技术领域Technical field
本申请涉及图像处理技术领域，特别是涉及一种高光谱遥感图像识别方法、装置、电子设备及可读存储介质。The present application relates to the field of image processing technology, and in particular to a hyperspectral remote sensing image recognition method, device, electronic equipment and readable storage medium.
背景技术Background technique
随着高光谱传感器的发展，遥感影像的光谱信息越来越丰富，凭借众多指纹性特征，遥感技术进入到一个崭新阶段。高光谱影像具有高达纳米级的光谱采样间隔，可获得的像元具有上百波段光谱信息，广泛应用于地物分类识别、定量反演、目标检测和生态环境监测等领域。其中，在地物分类识领域，高光谱图像分类是根据其影像所蕴含的光谱特征与空间特征将每个像素赋予唯一类别标签，是目前高光谱遥感技术中最活跃的研究领域，并引起了遥感领域的广泛关注。传统的高光谱图像识别方法基于光谱信息进行分类，输入的信息为一维信息。例如：支持向量机、随机森林的分类方法、K最近邻等方法。近年来，空间特征与光谱特征更多地应用到高光谱图像识别分类中。With the development of hyperspectral sensors, the spectral information of remote sensing images is becoming more and more abundant. With many fingerprint characteristics, remote sensing technology has entered a new stage. Hyperspectral images have spectral sampling intervals as high as nanometers, and the available pixels have hundreds of bands of spectral information. They are widely used in fields such as ground object classification and identification, quantitative inversion, target detection, and ecological environment monitoring. Among them, in the field of ground object classification, hyperspectral image classification assigns each pixel a unique category label based on the spectral characteristics and spatial characteristics contained in the image. It is currently the most active research field in hyperspectral remote sensing technology, and has caused widespread attention in the field of remote sensing. Traditional hyperspectral image recognition methods classify based on spectral information, and the input information is one-dimensional information. For example: support vector machine, random forest classification method, K nearest neighbor and other methods. In recent years, spatial features and spectral features have been increasingly used in hyperspectral image recognition and classification.
随着深度学习在各个领域的成功应用，所属领域技术人员将深度学习引入到高光谱图像识别中，并取得了良好的分类性能。如深度玻尔兹曼机、降噪自编码以及深度卷积神经网络。其中，深度卷积神经网络可自动提取出有效的空间以及光谱的高维特征，常见卷积神经网络有一维卷积、二维卷积、三维卷积的深度神经网络。卷积网络一般基于块(patch)进行运算，在高光谱遥感影像分类中，块的大小设定会导致分类结果失真。由于存在着空间感知细节的能力差、网络深度过度增加将会给网络模型带来梯度消失等问题，同时由于网络结构复杂以及模型训练样本有限导致模型训练不充分，识别精度低下。With the successful application of deep learning in various fields, technicians in the field have introduced deep learning into hyperspectral image recognition and achieved good classification performance. Such as deep Boltzmann machines, denoising autoencoders and deep convolutional neural networks. Among them, deep convolutional neural networks can automatically extract effective spatial and spectral high-dimensional features. Common convolutional neural networks include one-dimensional convolution, two-dimensional convolution, and three-dimensional convolution deep neural networks. Convolutional networks generally operate based on patches. In hyperspectral remote sensing image classification, the size setting of the patches will cause distortion of the classification results. Due to the poor ability to perceive details in space, excessive increase in network depth will bring gradient disappearance to the network model. At the same time, due to the complex network structure and limited model training samples, the model training is insufficient and the recognition accuracy is low.
发明内容Contents of the invention
本申请提供了一种高光谱遥感图像识别方法、装置、电子设备及可读存储介质，有效提升高光谱遥感图像的识别精准度。This application provides a hyperspectral remote sensing image recognition method, device, electronic equipment and readable storage medium, which effectively improves the recognition accuracy of hyperspectral remote sensing images.
为解决上述技术问题，本发明实施例提供以下技术方案：In order to solve the above technical problems, embodiments of the present invention provide the following technical solutions:
本发明实施例一方面提供了一种高光谱遥感图像识别方法，包括：On the one hand, embodiments of the present invention provide a hyperspectral remote sensing image recognition method, including:
预先训练图像识别模型；所述图像识别模型包括网络输入组合模块、多尺度双通卷积模块、卷积特征混合模块、压缩扩张空谱注意力机制模块；Pre-train the image recognition model; the image recognition model includes a network input combination module, a multi-scale two-pass convolution module, a convolution feature mixing module, and a compressed and expanded spatial spectrum attention mechanism module;
将待识别高光谱遥感图像输入至所述图像识别模型中，得到所述待识别高光谱遥感图像的图像识别结果；Input the hyperspectral remote sensing image to be identified into the image recognition model, and obtain the image recognition result of the hyperspectral remote sensing image to be identified;
其中，所述网络输入组合模块将所述待识别高光谱遥感图像的邻域特征和光谱维度相互合并，并从合并结果中选取网络输入图像块；所述多尺度双通卷积模块利用第一通道对所述网络输入图像块进行固定核卷积操作，利用第二通道对所述网络输入图像块进行多尺度空洞组卷积处理；所述卷积特征混合模块将不同卷积过程中的多通道特征进行混合；所述压缩扩张空谱注意力机制模块在光谱维与空间维上学习目标特征，并基于注意力权重自适应调整空谱特征图，并将最终得到的目标空谱特征图输送至判别卷积层和全连接层进行降维操作。Among them, the network input combination module merges the neighborhood features and spectral dimensions of the hyperspectral remote sensing image to be identified with each other, and selects the network input image block from the merged result; the multi-scale two-pass convolution module uses the first The channel performs a fixed kernel convolution operation on the network input image block, and uses the second channel to perform multi-scale hole group convolution processing on the network input image block; the convolution feature mixing module combines the multi-scale hole group convolution processing in different convolution processes. The channel features are mixed; the compressed and expanded spatial spectrum attention mechanism module learns the target features in the spectral dimension and the spatial dimension, and adaptively adjusts the spatial spectrum feature map based on the attention weight, and transmits the final target spatial spectrum feature map. Go to the discriminative convolutional layer and the fully connected layer to perform dimensionality reduction operations.
可选的，所述多尺度双通卷积模块为由第一卷积层和第二卷积层以双通道形式构成；Optionally, the multi-scale two-pass convolution module is composed of a first convolution layer and a second convolution layer in a dual-pass form;
所述第一卷积层的卷积核尺寸固定；所述第二卷积层由至少三层卷积通过串联方式构成。The size of the convolution kernel of the first convolution layer is fixed; the second convolution layer is composed of at least three layers of convolution in series.
可选的，所述第二卷积层包括第一子卷积层、第二子卷积层和第三子卷积层；所述第一子卷积层、所述第二子卷积层和所述第三子卷积层的卷积核尺寸相同，且采用递增空洞卷积率。Optionally, the second convolution layer includes a first sub-convolution layer, a second sub-convolution layer and a third sub-convolution layer; the first sub-convolution layer, the second sub-convolution layer The size of the convolution kernel is the same as that of the third sub-convolution layer, and the increasing atrous convolution rate is used.
可选的，所述将不同卷积过程中的多通道特征进行混合，包括：Optionally, mixing multi-channel features in different convolution processes includes:
获取所述第一通道输出的第一通道特征信息、所述第二通道输出的第二通道特征信息；Obtain the first channel feature information output by the first channel and the second channel feature information output by the second channel;
按照通道维度合并所述第一通道特征信息和所述第二通道特征信息；Merge the first channel feature information and the second channel feature information according to channel dimensions;
将合并结果的通道均分为两组，按照等位顺序对各组进行排列；提取相同位置的通道并重新进行排序。Divide the channels of the merged result into two groups and arrange each group in equal order; extract the channels at the same position and re-sort them.
可选的，所述分别在光谱维与空间维上学习目标特征，并自适应调整空谱特征图，并将最终得到的目标空谱特征图输送至判别卷积层和全连接层进行降维操作的过程，包括：Optionally, the target features are learned in the spectral dimension and the spatial dimension respectively, and the spatial spectrum feature map is adaptively adjusted, and the final target spatial spectrum feature map is sent to the discriminative convolution layer and the fully connected layer for dimensionality reduction. The operation process includes:
利用经全局池化层与多层感知器构成的神经网络对空谱特征进行通道维的操作，并经过RELU激活函数层输入至所述全连接层，再经过sigmoid激活函数层输出；The neural network composed of a global pooling layer and a multi-layer perceptron is used to perform channel-dimensional operations on the spatial spectral features, and is input to the fully connected layer through the RELU activation function layer, and then output through the sigmoid activation function layer;
经过通道维池化处理所述空谱特征得到单通道，经单层卷积操作和sigmoid激活激活函数层处理，生成空间注意力；The spatial spectral features are processed by channel dimension pooling to obtain a single channel, and then processed by a single layer convolution operation and a sigmoid activation function layer to generate spatial attention;
根据所述空间注意力与通道注意力确定空谱注意力；Determine spatial spectrum attention according to the spatial attention and channel attention;
根据所述空谱注意力和所述空谱特征，得到通道维缩放特征。According to the spatial spectrum attention and the spatial spectrum feature, channel dimension scaling features are obtained.
本发明实施例另一方面提供了一种高光谱遥感图像识别装置，包括：Another aspect of the embodiment of the present invention provides a hyperspectral remote sensing image recognition device, including:
模型预训练模块，用于预先训练图像识别模型；所述图像识别模型包括网络输入组合模块、多尺度双通卷积模块、卷积特征混合模块、压缩扩张空谱注意力机制模块；The model pre-training module is used to pre-train the image recognition model; the image recognition model includes a network input combination module, a multi-scale two-pass convolution module, a convolution feature mixing module, and a compressed and expanded spatial spectrum attention mechanism module;
图像识别模块，用于将待识别高光谱遥感图像输入至所述图像识别模型中，得到所述待识别高光谱遥感图像的图像识别结果；所述网络输入组合模块将所述待识别高光谱遥感图像的邻域特征和光谱维度相互合并，并从合并结果中选取网络输入图像块；所述多尺度双通卷积模块利用第一通道对所述网络输入图像块进行固定核卷积操作，利用第二通道对所述网络输入图像块进行多尺度空洞组卷积处理；所述卷积特征混合模块将不同卷积过程中的多通道特征进行混合；所述压缩扩张空谱注意力机制模块在光谱维与空间维上学习目标特征，并基于注意力权重自适应调整空谱特征图，并将最终得到的目标空谱特征图输送至判别卷积层和全连接层进行降维操作。The image recognition module is used to input the hyperspectral remote sensing image to be identified into the image recognition model to obtain the image recognition result of the hyperspectral remote sensing image to be identified; the network input combination module converts the hyperspectral remote sensing image to be identified to The neighborhood features and spectral dimensions of the image are merged with each other, and the network input image block is selected from the merged result; the multi-scale two-pass convolution module uses the first channel to perform a fixed kernel convolution operation on the network input image block, using The second channel performs multi-scale hole group convolution processing on the network input image block; the convolution feature mixing module mixes multi-channel features in different convolution processes; the compression and expansion spatial spectrum attention mechanism module The target features are learned in the spectral dimension and the spatial dimension, and the spatial spectrum feature map is adaptively adjusted based on the attention weight, and the final target spatial spectrum feature map is sent to the discriminative convolution layer and the fully connected layer for dimensionality reduction.
可选的，所述多尺度双通卷积模块为由第一卷积层和第二卷积层以双通道形式构成；Optionally, the multi-scale two-pass convolution module is composed of a first convolution layer and a second convolution layer in a dual-pass form;
所述第一卷积层的卷积核尺寸固定；所述第二卷积层由至少三层卷积通过串联方式构成。The size of the convolution kernel of the first convolution layer is fixed; the second convolution layer is composed of at least three layers of convolution in series.
可选的，所述第二卷积层包括第一子卷积层、第二子卷积层和第三子卷积层；所述第一子卷积层、所述第二子卷积层和所述第三子卷积层的卷积核尺寸相同，且采用递增空洞卷积率。Optionally, the second convolution layer includes a first sub-convolution layer, a second sub-convolution layer and a third sub-convolution layer; the first sub-convolution layer, the second sub-convolution layer The size of the convolution kernel is the same as that of the third sub-convolution layer, and the increasing atrous convolution rate is used.
本发明实施例还提供了一种电子设备，包括处理器，所述处理器用于执行存储器中存储的计算机程序时实现如前任一项所述高光谱遥感图像识别方法的步骤。An embodiment of the present invention also provides an electronic device, including a processor, which is configured to implement the steps of the hyperspectral remote sensing image recognition method as described in the previous item when executing a computer program stored in the memory.
本发明实施例最后还提供了一种可读存储介质，所述可读存储介质上存储有计算机程序，所述计算机程序被处理器执行时实现如前任一项所述高光谱遥感图像识别方法的步骤。Finally, the embodiment of the present invention also provides a readable storage medium, the readable storage medium stores a computer program, and when the computer program is executed by the processor, the hyperspectral remote sensing image recognition method as described in the previous item is implemented. step.
本申请提供的技术方案的优点在于，采用空谱邻域数据对网络输入进行组合，输入至由固定核卷积层与串联变式卷积层组成的双通道提取图像块进行多尺度特征分离式学习。不同尺度空洞卷积组合方式提取空谱特征时，可在不增加网络参数量和不消减数据特征的情况下同时提高卷积核的感受野，从而有效解决了空间感知细节的能力差的问题，提取表达能力更强的深层空谱特征，进而提升图像识别精准度。高效结合通道空间注意力机制，从通道和空间域两个方面增强特定图像特征的表现力，学习更多具有代表性的特征，进一步提升了高光谱遥感图像的识别精度。进一步地，整个图像识别模型的结构简单，数据处理速度快，有利于提高图像识别分类效率，有利于进一步推进了基于深度学习方法的高光谱图像分类的广泛应用。The advantage of the technical solution provided by this application is that the network input is combined using spatial spectrum neighborhood data, and the input is input to a dual-channel extraction image block composed of a fixed kernel convolution layer and a series variable convolution layer for multi-scale feature separation. study. When extracting spatial spectral features using a combination of dilated convolutions at different scales, the receptive field of the convolution kernel can be simultaneously increased without increasing the amount of network parameters or reducing data features, thus effectively solving the problem of poor spatial perception of details. Extract deep spatial spectrum features with stronger expressive capabilities to improve image recognition accuracy. Efficiently combines the channel spatial attention mechanism to enhance the expressiveness of specific image features from both channel and spatial domains, learn more representative features, and further improve the recognition accuracy of hyperspectral remote sensing images. Furthermore, the entire image recognition model has a simple structure and fast data processing speed, which is conducive to improving the efficiency of image recognition and classification, and is conducive to further promoting the widespread application of hyperspectral image classification based on deep learning methods.
此外，本发明实施例还针对高光谱遥感图像识别方法提供了相应的实现装置、电子设备及可读存储介质，进一步使得所述方法更具有实用性，所述装置、电子设备及可读存储介质具有相应的优点。In addition, embodiments of the present invention also provide corresponding implementation devices, electronic devices and readable storage media for the hyperspectral remote sensing image recognition method, further making the method more practical. The devices, electronic devices and readable storage media have corresponding advantages.
应当理解的是，以上的一般描述和后文的细节描述仅是示例性的，并不能限制本公开。It should be understood that the above general description and the following detailed description are only exemplary and do not limit the present disclosure.
附图说明Description of the drawings
为了更清楚的说明本发明实施例或相关技术的技术方案，下面将对实施例或相关技术描述中所需要使用的附图作简单的介绍，显而易见地，下面描述中的附图仅仅是本发明的一些实施例，对于本领域普通技术人员来讲，在不付出创造性劳动的前提下，还可以根据这些附图获得其他的附图。In order to more clearly illustrate the embodiments of the present invention or the technical solutions of related technologies, the drawings needed to be used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the drawings in the following description are only for the present invention. For some embodiments, those of ordinary skill in the art can also obtain other drawings based on these drawings without exerting creative efforts.
图1为本发明实施例提供的一种高光谱遥感图像识别方法的流程示意图；Figure 1 is a schematic flow chart of a hyperspectral remote sensing image recognition method provided by an embodiment of the present invention;
图2为本发明实施例提供的另一种高光谱遥感图像识别方法的流程示意图；Figure 2 is a schematic flow chart of another hyperspectral remote sensing image recognition method provided by an embodiment of the present invention;
图3为本发明实施例提供的网络输入组合模块在一个示例性应用场景的结构示意图；Figure 3 is a schematic structural diagram of the network input combination module provided by the embodiment of the present invention in an exemplary application scenario;
图4为本发明实施例提供的多尺度双通卷积模块在一个示例性应用场景的结构示意图；Figure 4 is a schematic structural diagram of the multi-scale two-pass convolution module provided by an embodiment of the present invention in an exemplary application scenario;
图5为本发明实施例提供的卷积特征混合模块在一个示例性应用场景的结构示意图；Figure 5 is a schematic structural diagram of the convolution feature mixing module provided by the embodiment of the present invention in an exemplary application scenario;
图6为本发明实施例提供的压缩扩张空谱注意力机制模块在一个示例性应用场景的结构示意图；Figure 6 is a schematic structural diagram of the compressed and expanded spatial spectrum attention mechanism module provided by the embodiment of the present invention in an exemplary application scenario;
图7为本发明实施例提供的图像识别模型在一个示例性应用场景的结构示意图；Figure 7 is a schematic structural diagram of the image recognition model provided by the embodiment of the present invention in an exemplary application scenario;
图8为本发明实施例提供的Indian Pines数据集中的一个示意性例子的示意图；Figure 8 is a schematic diagram of a schematic example in the Indian Pines data set provided by the embodiment of the present invention;
图9为本发明实施例提供的图8对应的真实地物图；Figure 9 is a real feature map corresponding to Figure 8 provided by an embodiment of the present invention;
图10为本发明实施例提供的高光谱图像识别装置的一种具体实施方式结构图；Figure 10 is a structural diagram of a specific implementation of the hyperspectral image recognition device provided by an embodiment of the present invention;
图11为本发明实施例提供的电子设备的一种具体实施方式结构图。Figure 11 is a structural diagram of a specific implementation of the electronic device provided by the embodiment of the present invention.
具体实施方式Detailed ways
为了使本技术领域的人员更好地理解本发明方案，下面结合附图和具体实施方式对本发明作进一步的详细说明。显然，所描述的实施例仅仅是本发明一部分实施例，而不是全部的实施例。基于本发明中的实施例，本领域普通技术人员在没有做出创造性劳动前提下所获得的所有其他实施例，都属于本发明保护的范围。In order to enable those skilled in the art to better understand the solution of the present invention, the present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments. Obviously, the described embodiments are only some of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the scope of protection of the present invention.
本申请的说明书和权利要求书及上述附图中的术语“第一”、“第二”、“第三”、“第四”等是用于区别不同的对象，而不是用于描述特定的顺序。此外术语“包括”和“具有”以及他们任何变形，意图在于覆盖不排他的包含。例如包含了一系列步骤或单元的过程、方法、系统、产品或设备没有限定于已列出的步骤或单元，而是可包括没有列出的步骤或单元。The terms "first", "second", "third", "fourth", etc. in the description and claims of this application and the above-mentioned drawings are used to distinguish different objects, rather than to describe specific objects. order. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device that includes a series of steps or units is not limited to the listed steps or units, but may include unlisted steps or units.
在介绍了本发明实施例的技术方案后，下面详细的说明本申请的各种非限制性实施方式。After introducing the technical solutions of the embodiments of the present invention, various non-limiting implementations of the present application are described in detail below.
首先参见图1，图1为本发明实施例提供的一种高光谱遥感图像识别方法的流程示意图，本发明实施例可包括以下内容：First referring to Figure 1, Figure 1 is a schematic flow chart of a hyperspectral remote sensing image recognition method provided by an embodiment of the present invention. The embodiment of the present invention may include the following:
S101：预先训练图像识别模型。S101: Pre-train the image recognition model.
本实施例的图像识别模型可包括网络输入组合模块、多尺度双通卷积模块、卷积特征混合模块、压缩扩张空谱注意力机制模块。网络输入组合模块为网络输入组织阶段，以输入的训练集数据块为基础，其首先进行全局降维，利用随机邻域嵌入方法进行光谱维度降维，降维凭借至光谱维度，得到初始数据，即将合并降维后像素邻域特征与光谱维度组成训练基块。具体的，可对输入影像的邻域特征与光谱维度相互合并之后，通过滑块窗口选取网络输入图像块。多尺度双通卷积模块，设计为双通道网络输入模式，一通道做固定核卷积，另一通道做多尺度空洞组卷积处理，以学习多尺度空谱特征表达。本实施例的空谱特征是指输入的遥感图像的空间特征和光谱特征。多尺度双通卷积模块将输入的训练集数据块作为训练单元，利用固定核卷积操作学习空谱特征基础表达，利用含有不同空洞率的串联式卷积操作实现多尺度空谱特征表达。卷积特征混合模块为特征组合方式，其设计重复混合过程，将不同卷积过程的多通道特征进行混合，也即将双通道学习的深层特征进行合并，利用张量转置方法，将两通道的输出进行对应位置对齐，将相同位置的特征进行提取并组合，然后将组合后的新空谱特征张量反转置回初始形状，增强各阶段卷积表达特征学习能力。举例来说，多尺度双通卷积模块可由2维固定核卷积层与2维3层串联变式卷积层以双通道方式构成，卷积特征混合模块可将合并两个分支的通道维度形成新组合特征，将双通道学习的深层特征进行混合合并，将分离式卷积与卷积合并进行迭代三次，最终输出分离式特征。压缩扩张空谱注意力机制模块完成通道的重要性权重学习和空间注意力，其在光谱维与空间维的两个维度上学习有意义的特征，依次推断出注意力权重，自适应调整空谱特征图，得到最终的空谱特征图。模型训练过程可包括：如图2所示，首先，对获取的原始的高光谱图像样本数据进行归一化且进行数据划分，将这些样本数据划分为训练集、验证集和测试集。其中，可从高光谱图像中随机划分出少量标签数据作为训练集与验证集，剩余标签数据作为测试集，训练集和测试集的比例可以设置为2：1：7。而后利用少量标签数据训练基于图像识别模型，最后结合已训练好的图像识别模型进行分类识别。在深度学习空谱联合网络中，输入数据首先会进行基于随机邻域嵌入(t-Distributed Stochastic NeighborEmbedding)方法进行降维，组成新组合特，传送给多尺度双通卷积模块中，得到空谱特征提取的特征图；随后经过卷积特征混合模块进行高阶特征混合，得到最终的空间特征图，最后经过压缩扩张空谱注意力机制模块校正特征图。当然，本实施例的图像识别模型不可避免的还可包括分类识别模块，分类识别模块由执行降维操作的全连接层和分类函数构成，分类函数例如可为softmax分类函数，分类识别模块将最终空间特征图和光谱特征图相拼接，并利用全连接融合所有特征信息，得到最后光谱空间特征，并传递给softmax函数进行目标分类。其中，softmax函数单元数量为相应类别总数。也就是说，将经上述网络模型的各网络模块提取的空谱特征深度输送至全连接层，全连接层融合所有特征至1维，再传递给softmax函数进行分类。The image recognition model in this embodiment may include a network input combination module, a multi-scale two-pass convolution module, a convolution feature mixing module, and a compressed and expanded spatial spectrum attention mechanism module. The network input combination module is the network input organization stage. Based on the input training set data block, it first performs global dimensionality reduction and uses the random neighborhood embedding method to reduce the spectral dimension. The dimensionality reduction relies on the spectral dimension to obtain the initial data. The dimensionally reduced pixel neighborhood features and spectral dimensions will be combined to form a training base block. Specifically, the neighborhood features and spectral dimensions of the input image can be combined with each other, and then the network input image block can be selected through the slider window. The multi-scale two-pass convolution module is designed as a dual-channel network input mode. One channel performs fixed kernel convolution, and the other channel performs multi-scale hole group convolution processing to learn multi-scale spatial spectrum feature expression. The spatial spectral features in this embodiment refer to the spatial features and spectral features of the input remote sensing image. The multi-scale two-pass convolution module uses the input training set data block as a training unit, uses fixed kernel convolution operations to learn the basic expression of spatial spectrum features, and uses tandem convolution operations with different hole rates to achieve multi-scale spatial spectrum feature expressions. The convolution feature mixing module is a feature combination method. It is designed to repeat the mixing process to mix multi-channel features of different convolution processes, that is, to merge the deep features of dual-channel learning, and use the tensor transposition method to combine the two-channel features. The output is aligned to the corresponding positions, features at the same position are extracted and combined, and then the combined new empty spectrum feature tensor is inverted and returned to its initial shape to enhance the convolutional expression feature learning ability at each stage. For example, the multi-scale two-pass convolution module can be composed of a 2-dimensional fixed kernel convolution layer and a 2-dimensional 3-layer serial variant convolution layer in a dual-channel manner. The convolution feature mixing module can merge the channel dimensions of the two branches. Form a new combined feature, mix and merge the deep features of dual-channel learning, iterate three times by combining separate convolution and convolution, and finally output the separate feature. The compressed and expanded spatial spectrum attention mechanism module completes the importance weight learning and spatial attention of the channel. It learns meaningful features in the two dimensions of spectral dimension and spatial dimension, infers the attention weight in turn, and adaptively adjusts the spatial spectrum. feature map to obtain the final spatial spectrum feature map. The model training process may include: As shown in Figure 2, first, the obtained original hyperspectral image sample data is normalized and divided into data, and the sample data is divided into a training set, a verification set and a test set. Among them, a small amount of label data can be randomly divided from the hyperspectral image as the training set and verification set, and the remaining label data can be used as the test set. The ratio of the training set and the test set can be set to 2:1:7. Then use a small amount of label data to train the image recognition model, and finally combine the trained image recognition model for classification and recognition. In the deep learning space-spectrum joint network, the input data will first be dimensionally reduced based on the stochastic neighbor embedding (t-Distributed Stochastic NeighborEmbedding) method to form a new combination feature, which is then sent to the multi-scale two-pass convolution module to obtain the space spectrum. The feature map of feature extraction is then mixed with high-order features through the convolutional feature mixing module to obtain the final spatial feature map. Finally, the feature map is corrected through the compressed and expanded spatial spectrum attention mechanism module. Of course, the image recognition model of this embodiment can also inevitably include a classification recognition module. The classification recognition module is composed of a fully connected layer that performs dimensionality reduction operations and a classification function. The classification function can be, for example, a softmax classification function. The classification recognition module will finally The spatial feature map and the spectral feature map are spliced, and all feature information is fused using full connection to obtain the final spectral space feature, which is passed to the softmax function for target classification. Among them, the number of softmax function units is the total number of corresponding categories. That is to say, the spatial spectrum feature depth extracted by each network module of the above network model is sent to the fully connected layer, and the fully connected layer fuses all features into 1 dimension, and then passes it to the softmax function for classification.
S102：将待识别高光谱遥感图像输入至图像识别模型中，得到待识别高光谱遥感图像的图像识别结果。S102: Input the hyperspectral remote sensing image to be identified into the image recognition model, and obtain the image recognition result of the hyperspectral remote sensing image to be identified.
为了提高图像识别精准度，可对获取的待识别高光谱遥感图像先预处理，如去噪、平滑等。然后将待识别高光谱遥感图像输入至上个步骤训练所得的图像识别模型中，图像识别模型的网络输入组合模块向将待识别高光谱遥感图像的邻域特征和光谱维度相互合并，并从合并结果中选取网络输入图像块，并将这些网络输入图像块输入至多尺度双通卷积模块中，多尺度双通卷积模块利用第一通道对网络输入图像块进行固定核卷积操作，利用第二通道对网络输入图像块进行多尺度空洞组卷积处理。卷积特征混合模块将多尺度双通卷积模块在不同卷积过程中的多通道特征进行混合。压缩扩张空谱注意力机制模块在光谱维与空间维上学习目标特征，并基于注意力权重自适应调整空谱特征图，并将最终得到的目标空谱特征图输送至判别卷积层和全连接层进行降维操作，在依赖图像识别模型的分类识别函数如softmax函数进行分类，并输出类别预测相应的识别结果，例如可得到分类结果图。In order to improve the accuracy of image recognition, the acquired hyperspectral remote sensing images to be identified can be pre-processed, such as denoising, smoothing, etc. Then the hyperspectral remote sensing image to be identified is input into the image recognition model trained in the previous step. The network input combination module of the image recognition model merges the neighborhood features and spectral dimensions of the hyperspectral remote sensing image to be identified with each other, and from the merged result Select the network input image blocks from , and input these network input image blocks into the multi-scale two-pass convolution module. The multi-scale two-pass convolution module uses the first channel to perform a fixed kernel convolution operation on the network input image blocks, and uses the second channel The channel performs multi-scale atrous group convolution processing on the network input image patches. The convolution feature mixing module mixes the multi-channel features of the multi-scale two-pass convolution module in different convolution processes. The compressed and expanded spatial spectrum attention mechanism module learns target features in spectral and spatial dimensions, adaptively adjusts the spatial spectrum feature map based on attention weights, and sends the final target spatial spectrum feature map to the discriminative convolution layer and the full The connection layer performs dimensionality reduction operations, performs classification on a classification recognition function such as softmax function that relies on the image recognition model, and outputs the recognition results corresponding to the category prediction. For example, the classification result map can be obtained.
在本发明实施例提供的技术方案中，采用空谱邻域数据对网络输入进行组合，输入至由固定核卷积层与串联变式卷积层组成的双通道提取图像块进行多尺度特征分离式学习。不同尺度空洞卷积组合方式提取空谱特征时，可在不增加网络参数量和不消减数据特征的情况下同时提高卷积核的感受野，从而有效解决了空间感知细节的能力差的问题，提取表达能力更强的深层空谱特征，进而提升图像识别精准度。高效结合通道空间注意力机制，从通道和空间域两个方面增强特定图像特征的表现力，学习更多具有代表性的特征，进一步提升了高光谱遥感图像的识别精度。进一步地，整个图像识别模型的结构简单，数据处理速度快，有利于提高图像识别分类效率，有利于进一步推进了基于深度学习方法的高光谱图像分类的广泛应用。In the technical solution provided by the embodiment of the present invention, the spatial spectrum neighborhood data is used to combine the network input, and the input is input to a dual-channel extraction image block composed of a fixed kernel convolution layer and a series variable convolution layer for multi-scale feature separation. style learning. When extracting spatial spectral features using a combination of dilated convolutions at different scales, the receptive field of the convolution kernel can be simultaneously increased without increasing the amount of network parameters or reducing data features, thus effectively solving the problem of poor spatial perception of details. Extract deep spatial spectrum features with stronger expressive capabilities to improve image recognition accuracy. Efficiently combines the channel spatial attention mechanism to enhance the expressiveness of specific image features from both channel and spatial domains, learn more representative features, and further improve the recognition accuracy of hyperspectral remote sensing images. Furthermore, the entire image recognition model has a simple structure and fast data processing speed, which is conducive to improving the efficiency of image recognition and classification, and is conducive to further promoting the widespread application of hyperspectral image classification based on deep learning methods.
需要说明的是，本申请中各步骤之间没有严格的先后执行顺序，只要符合逻辑上的顺序，则这些步骤可以同时执行，也可按照某种预设顺序执行，图1-图2只是一种示意方式，并不代表只能是这样的执行顺序。It should be noted that there is no strict order of execution between the steps in this application. As long as they comply with the logical order, these steps can be executed at the same time or in a certain preset order. Figures 1-2 are just a diagram. This way of indicating does not mean that this is the only execution sequence.
在上述实施例中，对于网络输入组合模块的结构并不做限定，本实施例中给出网络输入组合模块对输入数据的一种可选的处理方式，可包括如下步骤：In the above embodiment, the structure of the network input combination module is not limited. This embodiment provides an optional processing method for the input data by the network input combination module, which may include the following steps:
网络输入组合模块以网络输入组合的单个图像块为输入单元来构建，如图3所示，网络输入组合模块组合初始空谱特征网络输入基块的过程如下：The network input combination module is constructed with a single image block of the network input combination as the input unit. As shown in Figure 3, the process of the network input combination module combining the initial spatial spectrum feature network input base blocks is as follows:
以高光谱图像块为输入单元，网络输入组合模块首先进行全局降维，利用随机邻域嵌入方法进行光谱维度降维，对降维后的每个像素的前3个维度提取，在此基础上对每个像素给定领域的降维数据合并为一维向量，附加至其光谱维度后，组成新组合特征。同时，对每个像素选取以其为中心的固定大小邻域数据块作为网络输入块进行训练与预测。若将高光谱影像记作HSI(X，Y，S)，其中，X是影像长度，Y是影像宽度，S是影像光谱维度。利用随机邻域嵌入方法对HSI进行降维，得到HSI_down(X，Y，S_down)，其中S_down取降维后的前3维。对整个影像进行滑动窗口处理，对每个像素围绕中心邻域的所有降维后特征进行合并成1维向量，并拼接到对应像素光谱特征向量后，形成HSI_new(X，Y，S_new)，其中S_new的维度为合成后的特征维度。将HSI_new(X，Y，S_new)按步长1进行窗口裁取，输入至网络中。Taking the hyperspectral image block as the input unit, the network input combination module first performs global dimensionality reduction, uses the random neighborhood embedding method to reduce the spectral dimension, and extracts the first three dimensions of each pixel after dimensionality reduction. On this basis The dimensionality reduction data in a given area of each pixel are combined into a one-dimensional vector, which is appended to its spectral dimension to form a new combined feature. At the same time, a fixed-size neighborhood data block centered on each pixel is selected as the network input block for training and prediction. If the hyperspectral image is recorded as HSI (X, Y, S), where X is the image length, Y is the image width, and S is the image spectral dimension. The random neighborhood embedding method is used to reduce the dimensionality of HSI, and HSI_down(X, Y, S_down) is obtained, where S_down takes the first 3 dimensions after dimensionality reduction. Perform sliding window processing on the entire image, merge all reduced-dimensional features of each pixel around the central neighborhood into a 1-dimensional vector, and splice them into the corresponding pixel spectral feature vector to form HSI_new(X, Y, S_new), where The dimension of S_new is the synthesized feature dimension. Cut the window of HSI_new(X, Y, S_new) with a step size of 1 and input it into the network.
在上述实施例中，对于如何多尺度双通卷积模块的结构并不做限定，本实施例中给出多尺度双通卷积模块的一种可选的结构实施方式，请参阅图4，可包括：In the above embodiment, there is no limitation on the structure of the multi-scale two-pass convolution module. This embodiment provides an optional structural implementation of the multi-scale two-pass convolution module. Please refer to Figure 4. Can include:
多尺度双通卷积模块可为由第一卷积层和第二卷积层以双通道形式构成；第一卷积层的卷积核尺寸固定；第二卷积层由至少三层卷积通过串联方式构成。其中，第二卷积层包括第一子卷积层、第二子卷积层和第三子卷积层；第一子卷积层、第二子卷积层和第三子卷积层的卷积核尺寸相同，且采用递增空洞卷积率。The multi-scale two-pass convolution module can be composed of a first convolution layer and a second convolution layer in a dual-channel form; the convolution kernel size of the first convolution layer is fixed; the second convolution layer consists of at least three layers of convolution Constructed through series connection. Wherein, the second convolution layer includes a first sub-convolution layer, a second sub-convolution layer and a third sub-convolution layer; the first sub-convolution layer, the second sub-convolution layer and the third sub-convolution layer are The convolution kernel size is the same, and the increasing atrous convolution rate is used.
在本实施例中，以待识别高光谱图像的网络输入图像块为输入单元，多尺度双通卷积模块由2维固定核卷积层与2维串联变式卷积层以双通道方式构成，固定核卷积层由相同大小的卷积核组成，串联变式卷积层的由至少三层卷积通过串联方式组合构成，其中，三层卷积层的卷积核大小应一致，且采用递增空洞卷据率，如第一层空洞卷积率为1，第二层空洞卷积率为3，第三层空洞卷积率为5。基于图4所示的结构，多尺度双通卷积模块提取空谱特征的过程可如下：In this embodiment, the network input image block of the hyperspectral image to be recognized is used as the input unit, and the multi-scale two-pass convolution module is composed of a 2-dimensional fixed kernel convolution layer and a 2-dimensional series variant convolution layer in a dual-channel manner. , the fixed kernel convolution layer is composed of convolution kernels of the same size, and the series variant convolution layer is composed of at least three layers of convolutions combined in series. Among them, the convolution kernel sizes of the three layers of convolution layers should be consistent, and Increasing atrous convolution rates are used, such as the atrous convolution rate of the first layer is 1, the atrous convolution rate of the second layer is 3, and the atrous convolution rate of the third layer is 5. Based on the structure shown in Figure 4, the process of extracting spatial spectral features by the multi-scale two-pass convolution module can be as follows:
(1)并列层第一部分卷积层为2维固定核卷积层。(1) The first part of the convolution layer of the parallel layer is a 2-dimensional fixed kernel convolution layer.
(2)并列层第二部分卷积层为2维串联变式卷积层，设定至少三层带空洞的卷积提取层，所有层的卷积核大小保持一致，空洞率依次增加。(2) The second part of the convolution layer of the parallel layer is a 2-dimensional series variant convolution layer. At least three convolution extraction layers with holes are set. The convolution kernel size of all layers remains the same, and the hole rate increases sequentially.
(3)卷积模块中每一层网络都增加了BN层，同时采用ReLU激活函数，网络采用输入输出等大小的padding方式。(3) Each layer of the network in the convolution module adds a BN layer and uses the ReLU activation function. The network uses a padding method of equal size for input and output.
上述实施例对如何将不同卷积过程中的多通道特征进行混合的实施方式并不做限定，本实施例还给出一种可选的实施方式，可包括：The above embodiment does not limit the implementation of how to mix multi-channel features in different convolution processes. This embodiment also provides an optional implementation, which may include:
获取第一通道输出的第一通道特征信息、第二通道输出的第二通道特征信息；按照通道维度合并第一通道特征信息和第二通道特征信息；将合并结果的通道均分为两组，按照等位顺序对各组进行排列；提取相同位置的通道并重新进行排序。Obtain the first channel feature information output by the first channel and the second channel feature information output by the second channel; merge the first channel feature information and the second channel feature information according to the channel dimensions; divide the channels of the merged result into two groups, Arrange groups in equal order; channels at the same position are extracted and reordered.
在本实施例中，用于实现将不同卷积过程中的多通道特征进行混合的功能模块为卷积特征混合模块，结合图5所示，卷积特征混合模块的工作过程如下：通过合并两个分支的通道维度形成新组合特征，将双通道学习的深层特征进行合并，利用张量转置方法，将两通道的输出进行对应位置对齐，将相同位置的特征进行提取并组合，然后将组合后的新空谱特征张量反转置回初始形状，增强网络特征的学习能力，每次选择1/2进行混合卷积层，通过多个重复上述方式实现空谱特征充分混合。基于如图5所示卷积特征混合模块卷积层其结构，卷积特征混合模块提取空谱特征图的过程可如下：In this embodiment, the functional module used to mix multi-channel features in different convolution processes is a convolution feature mixing module. As shown in Figure 5, the working process of the convolution feature mixing module is as follows: by merging two The channel dimensions of each branch form a new combined feature, merge the deep features of dual-channel learning, use the tensor transposition method to align the output of the two channels at the corresponding positions, extract and combine the features at the same position, and then combine The new spatial spectrum feature tensor is inverted and reset to its initial shape to enhance the learning ability of network features. 1/2 is selected each time for the mixed convolution layer, and the spatial spectrum features are fully mixed by repeating the above method multiple times. Based on the structure of the convolution layer of the convolutional feature mixing module as shown in Figure 5, the process of extracting the spatial spectrum feature map by the convolutional feature mixing module can be as follows:
多尺度双通卷积模块输出的特征可称为：F1(c，x，y)和F2(c，x，y)，将这两个通道特征输出进行特征维度转置，形成新的张量F_t(2，c，x，y)，在第一个维度上进行合并，形成张量F_t(c，2，x，y)，最后合并张量的第2个维度，生成最终混合特征张量F(2c，x，y)，卷积特征混合模块将F(2c，x，y)输送至后续的压缩扩张空谱注意力机制模块进行处理。The features output by the multi-scale two-pass convolution module can be called: F1 (c, x, y) and F2 (c, x, y). The feature dimensions of these two channel feature outputs are transposed to form a new tensor. F_t(2,c,x,y), merge in the first dimension to form the tensor F_t(c,2,x,y), and finally merge the second dimension of the tensor to generate the final mixed feature tensor F(2c,x,y), the convolutional feature mixing module sends F(2c,x,y) to the subsequent compression and expansion spatial spectrum attention mechanism module for processing.
上述对如何执行分别在光谱维与空间维上学习目标特征，并自适应调整空谱特征图，并不做限定，本申请还给出了一种可选的实施方式，可包括：The above does not limit how to learn the target features in the spectral dimension and the spatial dimension respectively, and adaptively adjust the spatial spectrum feature map. This application also provides an optional implementation method, which may include:
利用经全局池化层与多层感知器构成的神经网络对空谱特征进行通道维的操作，并经过RELU激活函数层输入至全连接层，再经过sigmoid激活函数层输出；经过通道维池化处理空谱特征得到单通道，经单层卷积操作和sigmoid激活激活函数层处理，生成空间注意力；根据空间注意力与通道注意力确定空谱注意力；根据空谱注意力和空谱特征，得到通道维缩放特征。The neural network composed of a global pooling layer and a multi-layer perceptron is used to perform channel-dimensional operations on the spatial spectral features, and is input to the fully connected layer through the RELU activation function layer, and then output through the sigmoid activation function layer; after channel-dimensional pooling The spatial spectrum features are processed to obtain a single channel. After single-layer convolution operation and sigmoid activation function layer processing, spatial attention is generated; the spatial spectrum attention is determined according to the spatial attention and channel attention; the spatial spectrum attention and the spatial spectrum characteristics are determined according to the spatial attention and the spatial spectrum characteristics. , to obtain channel dimension scaling features.
在本实施例中，压缩扩张空谱注意力机制模块调整空谱特征图的过程如下：以学习的空谱特征图为单位，首先进行通道维的操作，经全局池化与多层感知器构成的神经网络，经过RELU激活函数输入至全连接网络，再经过sigmoid激活输出，完成通道的重要性权重学习。同时面向输入空谱特征，经过通道维池化获得单通道，经单层卷积操作和sigmoid激活，生成空间注意力，将此与通道注意力点乘形成空谱注意力，将结果乘以原始输入特征，得到新通道维的缩放特征。压缩扩张空谱注意力机制模块及其结构可如图6所示，基于提取空谱特征图的过程如下：In this embodiment, the process of adjusting the spatial spectrum feature map by the compression and expansion spatial spectrum attention mechanism module is as follows: taking the learned spatial spectrum feature map as a unit, first perform channel-dimensional operations, and then construct it through global pooling and multi-layer perceptrons. The neural network is input to the fully connected network through the RELU activation function, and then output through sigmoid activation to complete the importance weight learning of the channel. At the same time, facing the input spatial spectrum features, a single channel is obtained through channel dimension pooling. After a single layer convolution operation and sigmoid activation, spatial attention is generated. This is multiplied with the channel attention point to form the spatial spectrum attention. The result is multiplied by the original input. Features, get the scaled features of the new channel dimension. The compression expansion spatial spectrum attention mechanism module and its structure can be shown in Figure 6. The process based on extracting the spatial spectrum feature map is as follows:
(1)以混合后的空谱特征图为单位，首先进行通道维的操作，经过全局平均池经全局池化与多层感知器构成的神经网络，经过RELU激活函数输入至全连接网络，再经过sigmoid激活输出，完成通道的重要性权重学习。(2)同时以混合后的空谱特征图为单位，经过全局平均池化得到二维的空间特征描述图，经单层卷积操作和sigmoid激活，生成空间注意力，将此与通道注意力点乘形成空谱注意力，将结果乘以原始输入特征，得到新通道维的缩放特征。(1) Taking the mixed spatial spectrum feature map as a unit, the channel dimension is first operated, and the neural network composed of global pooling and multi-layer perceptron is passed through the global average pool, and then input to the fully connected network through the RELU activation function, and then After sigmoid activation output, the importance weight learning of the channel is completed. (2) At the same time, using the mixed spatial spectrum feature map as a unit, a two-dimensional spatial feature description map is obtained through global average pooling. After a single layer convolution operation and sigmoid activation, spatial attention is generated, which is combined with the channel attention point The multiplication forms the spatial spectral attention, and the result is multiplied by the original input features to obtain the scaled features of the new channel dimension.
为了验证本申请技术方案的有效性，本申请还提供了一个验证性的实施例，可包括：In order to verify the effectiveness of the technical solution of this application, this application also provides a verification example, which may include:
在本实施例中，将所有训练样本数据按照20％、10％、70％随机划分为训练数据、验证数据、测试数据。基于图7所示的网络结构作为图像识别模型，采用Adam优化方法优化图像识别模型。网络的训练次数为1000次，每批次训练数量设置为60。学习率设置为0.001。同时采用退化学习率的训练策略，其中每250次使学习率减少10％。按照如上参数，基于训练样本数据对网络结构进行训练，得到训练好的图像识别模型。In this embodiment, all training sample data are randomly divided into training data, verification data, and test data according to 20%, 10%, and 70%. Based on the network structure shown in Figure 7 as the image recognition model, the Adam optimization method is used to optimize the image recognition model. The number of training times of the network is 1000 times, and the number of trainings in each batch is set to 60. The learning rate is set to 0.001. At the same time, a training strategy of degenerated learning rate is adopted, in which the learning rate is reduced by 10% every 250 times. According to the above parameters, the network structure is trained based on the training sample data to obtain the trained image recognition model.
为了验证本申请的图像识别模型可有效提升高光谱遥感图像的识别精准度，利用图像识别模型对真实的IndianPines的高光谱影像进行分类识别，如图8及图9所示，该影像大小为145行×145列×224波段，去除相应的水汽波段、低信噪比波段，剩余200个可用波进行分类。按如下步骤采用上述实施例所提供的技术方案进行高光谱遥感图像的识别分类：In order to verify that the image recognition model of this application can effectively improve the recognition accuracy of hyperspectral remote sensing images, the image recognition model was used to classify and identify the real hyperspectral images of Indian Pines. As shown in Figure 8 and Figure 9, the image size is 145 Rows × 145 columns × 224 bands, removing the corresponding water vapor bands and low signal-to-noise ratio bands, leaving 200 available waves for classification. Follow the following steps to use the technical solutions provided in the above embodiments to identify and classify hyperspectral remote sensing images:
首先，对于IndianPines高光谱数据集，为了验证算法的有效性，选择了4个基准算法SVM、KNN、1D-CNN、2D-CNN进行比较。First, for the IndianPines hyperspectral data set, in order to verify the effectiveness of the algorithm, four benchmark algorithms SVM, KNN, 1D-CNN, and 2D-CNN were selected for comparison.
其次，将实验中每种算法的分类结果与地面真实值进行比较，采用总体识别精度(OverallAccuracy，OA)、平均分类精度(AverageAccuracy，AA)、Kappa系数作为性能评价指标。其计算结果见表1所示：Secondly, the classification results of each algorithm in the experiment are compared with the ground truth value, and the overall recognition accuracy (OverallAccuracy, OA), average classification accuracy (AverageAccuracy, AA), and Kappa coefficient are used as performance evaluation indicators. The calculation results are shown in Table 1:
表1高光谱遥感图像识别方法识别性能对比信息Table 1 Recognition performance comparison information of hyperspectral remote sensing image recognition methods
从表1可见，本申请技术方案所提的图像识别模型的各性能评价指标显著高于另外4种基准算法。It can be seen from Table 1 that the performance evaluation indicators of the image recognition model proposed in the technical solution of this application are significantly higher than the other four benchmark algorithms.
由上可知，本实施例采用双通道分离式多尺度卷积与卷积混合操作对高光谱影像数据库连续学习空谱特征，构造空谱注意机制，增强特定区域特征的表示能力，在面对大量冗余信息时，学习更多具有代表性的特征。相比现有技术，具有更高的分类精度与能力，有效解决了卷积神经网络的分类方法存在的空间感知细节的能力差与网络模型性能退化问题，分类精度进一步提升。It can be seen from the above that this embodiment uses dual-channel separated multi-scale convolution and convolution hybrid operations to continuously learn the spatial spectrum features of the hyperspectral image database, constructs a spatial spectrum attention mechanism, and enhances the representation ability of specific area features. In the face of a large number of When there is redundant information, learn more representative features. Compared with the existing technology, it has higher classification accuracy and capability, effectively solves the problems of poor spatial perception details and network model performance degradation problems of the convolutional neural network classification method, and further improves the classification accuracy.
本发明实施例还针对高光谱遥感图像识别方法提供了相应的装置，进一步使得方法更具有实用性。其中，装置可从功能模块的角度和硬件的角度分别说明。下面对本发明实施例提供的高光谱遥感图像识别装置进行介绍，下文描述的高光谱遥感图像识别装置与上文描述的高光谱遥感图像识别方法可相互对应参照。Embodiments of the present invention also provide corresponding devices for the hyperspectral remote sensing image recognition method, further making the method more practical. Among them, the device can be described separately from the perspective of functional modules and the perspective of hardware. The hyperspectral remote sensing image recognition device provided by the embodiment of the present invention is introduced below. The hyperspectral remote sensing image recognition device described below and the hyperspectral remote sensing image recognition method described above can correspond to each other.
基于功能模块的角度，参见图10，图10为本发明实施例提供的高光谱遥感图像识别装置在一种具体实施方式下的结构图，该装置可包括：From the perspective of functional modules, see Figure 10. Figure 10 is a structural diagram of a hyperspectral remote sensing image recognition device provided by an embodiment of the present invention in a specific implementation. The device may include:
模型预训练模块101，用于预先训练图像识别模型；图像识别模型包括网络输入组合模块、多尺度双通卷积模块、卷积特征混合模块、压缩扩张空谱注意力机制模块。The model pre-training module 101 is used to pre-train the image recognition model; the image recognition model includes a network input combination module, a multi-scale two-pass convolution module, a convolution feature mixing module, and a compressed and expanded spatial spectrum attention mechanism module.
图像识别模块1002，用于将待识别高光谱遥感图像输入至图像识别模型中，得到待识别高光谱遥感图像的图像识别结果。网络输入组合模块将待识别高光谱遥感图像的邻域特征和光谱维度相互合并，并从合并结果中选取网络输入图像块；多尺度双通卷积模块利用第一通道对网络输入图像块进行固定核卷积操作，利用第二通道对网络输入图像块进行多尺度空洞组卷积处理；卷积特征混合模块将不同卷积过程中的多通道特征进行混合；压缩扩张空谱注意力机制模块在光谱维与空间维上学习目标特征，并基于注意力权重自适应调整空谱特征图，并将最终得到的目标空谱特征图输送至判别卷积层和全连接层进行降维操作。The image recognition module 1002 is used to input the hyperspectral remote sensing image to be recognized into the image recognition model, and obtain the image recognition result of the hyperspectral remote sensing image to be recognized. The network input combination module combines the neighborhood features and spectral dimensions of the hyperspectral remote sensing image to be identified, and selects the network input image blocks from the merged results; the multi-scale two-pass convolution module uses the first channel to fix the network input image blocks Kernel convolution operation uses the second channel to perform multi-scale hole group convolution processing on the network input image block; the convolution feature mixing module mixes multi-channel features in different convolution processes; the compression expansion spatial spectrum attention mechanism module The target features are learned in the spectral dimension and the spatial dimension, and the spatial spectrum feature map is adaptively adjusted based on the attention weight, and the final target spatial spectrum feature map is sent to the discriminative convolution layer and the fully connected layer for dimensionality reduction.
可选的，在本实施例的一些实施方式中，上述多尺度双通卷积模块为由第一卷积层和第二卷积层以双通道形式构成；第一卷积层的卷积核尺寸固定；第二卷积层由至少三层卷积通过串联方式构成。Optionally, in some implementations of this embodiment, the above-mentioned multi-scale two-pass convolution module is composed of a first convolution layer and a second convolution layer in a dual-pass form; the convolution kernel of the first convolution layer The size is fixed; the second convolutional layer is composed of at least three layers of convolution in series.
作为本实施例的一种可选的实施方式，上述第二卷积层包括第一子卷积层、第二子卷积层和第三子卷积层；第一子卷积层、第二子卷积层和第三子卷积层的卷积核尺寸相同，且采用递增空洞卷积率。As an optional implementation of this embodiment, the above-mentioned second convolution layer includes a first sub-convolution layer, a second sub-convolution layer and a third sub-convolution layer; the first sub-convolution layer, the second sub-convolution layer and the third sub-convolution layer. The convolution kernel sizes of the sub-convolution layer and the third sub-convolution layer are the same, and the increasing atrous convolution rate is used.
可选的，在本实施例的另一些实施方式中，上述图像识别模块1002可用于：获取第一通道输出的第一通道特征信息、第二通道输出的第二通道特征信息；按照通道维度合并第一通道特征信息和第二通道特征信息；将合并结果的通道均分为两组，按照等位顺序对各组进行排列；提取相同位置的通道并重新进行排序。Optionally, in other implementations of this embodiment, the above-mentioned image recognition module 1002 can be used to: obtain the first channel feature information output by the first channel and the second channel feature information output by the second channel; merge according to channel dimensions The first channel feature information and the second channel feature information; divide the channels of the merged result into two groups, arrange each group in equal order; extract the channels at the same position and reorder them.
可选的，在本实施例的其他一些实施方式中，上述图像识别模块1002还可进一步用于：利用经全局池化层与多层感知器构成的神经网络对空谱特征进行通道维的操作，并经过RELU激活函数层输入至全连接层，再经过sigmoid激活函数层输出；经过通道维池化处理空谱特征得到单通道，经单层卷积操作和sigmoid激活激活函数层处理，生成空间注意力；根据空间注意力与通道注意力确定空谱注意力；根据空谱注意力和空谱特征，得到通道维缩放特征。Optionally, in other implementations of this embodiment, the above-mentioned image recognition module 1002 can be further used to perform channel-dimensional operations on spatial spectral features using a neural network composed of a global pooling layer and a multi-layer perceptron. , and is input to the fully connected layer through the RELU activation function layer, and then output through the sigmoid activation function layer; the spatial spectrum feature is obtained by channel dimension pooling to obtain a single channel, and is processed by a single layer convolution operation and a sigmoid activation function layer to generate a space Attention; determine spatial spectrum attention based on spatial attention and channel attention; obtain channel dimension scaling features based on spatial spectrum attention and spatial spectrum features.
本发明实施例高光谱遥感图像识别装置的各功能模块的功能可根据上述方法实施例中的方法具体实现，其具体实现过程可以参照上述方法实施例的相关描述，此处不再赘述。The functions of each functional module of the hyperspectral remote sensing image recognition device in the embodiment of the present invention can be specifically implemented according to the method in the above method embodiment. The specific implementation process can be referred to the relevant description of the above method embodiment, and will not be described again here.
由上可知，本发明实施例可有效提升高光谱遥感图像的识别精准度。It can be seen from the above that embodiments of the present invention can effectively improve the recognition accuracy of hyperspectral remote sensing images.
上文中提到的高光谱遥感图像识别装置是从功能模块的角度描述，进一步的，本申请还提供一种电子设备，是从硬件角度描述。图11为本申请实施例提供的电子设备在一种实施方式下的结构示意图。如图11所示，该电子设备包括存储器110，用于存储计算机程序；处理器111，用于执行计算机程序时实现如上述任一实施例提到的高光谱遥感图像识别方法的步骤。The hyperspectral remote sensing image recognition device mentioned above is described from the perspective of functional modules. Furthermore, this application also provides an electronic device, which is described from the perspective of hardware. FIG. 11 is a schematic structural diagram of an electronic device provided by an embodiment of the present application in an implementation manner. As shown in Figure 11, the electronic device includes a memory 110 for storing a computer program; a processor 111 for executing the computer program to implement the steps of the hyperspectral remote sensing image recognition method mentioned in any of the above embodiments.
其中，处理器111可以包括一个或多个处理核心，比如4核心处理器、8核心处理器，处理器111还可为控制器、微控制器、微处理器或其他数据处理芯片等。处理器111可以采用DSP(Digital Signal Processing，数字信号处理)、FPGA(Field－Programmable GateArray，现场可编程门阵列)、PLA(Programmable Logic Array，可编程逻辑阵列)中的至少一种硬件形式来实现。处理器111也可以包括主处理器和协处理器，主处理器是用于对在唤醒状态下的数据进行处理的处理器，也称CPU(Central Processing Unit，中央处理器)；协处理器是用于对在待机状态下的数据进行处理的低功耗处理器。在一些实施例中，处理器111可以集成有GPU(Graphics Processing Unit，图像处理器)，GPU用于负责显示屏所需要显示的内容的渲染和绘制。一些实施例中，处理器111还可以包括AI(ArtificialIntelligence，人工智能)处理器，该AI处理器用于处理有关机器学习的计算操作。The processor 111 may include one or more processing cores, such as a 4-core processor or an 8-core processor. The processor 111 may also be a controller, a microcontroller, a microprocessor or other data processing chips. The processor 111 can be implemented in at least one hardware form among DSP (Digital Signal Processing, digital signal processing), FPGA (Field-Programmable GateArray, field programmable gate array), and PLA (Programmable Logic Array, programmable logic array). . The processor 111 may also include a main processor and a co-processor. The main processor is a processor used to process data in the wake-up state, also called CPU (Central Processing Unit, central processing unit); the co-processor is A low-power processor used to process data in standby mode. In some embodiments, the processor 111 may be integrated with a GPU (Graphics Processing Unit, image processor), and the GPU is responsible for rendering and drawing content to be displayed on the display screen. In some embodiments, the processor 111 may also include an AI (Artificial Intelligence, artificial intelligence) processor, which is used to process computing operations related to machine learning.
存储器110可以包括一个或多个计算机可读存储介质，该计算机可读存储介质可以是非暂态的。存储器110还可包括高速随机存取存储器以及非易失性存储器，比如一个或多个磁盘存储设备、闪存存储设备。存储器110在一些实施例中可以是电子设备的内部存储单元，例如服务器的硬盘。存储器110在另一些实施例中也可以是电子设备的外部存储设备，例如服务器上配备的插接式硬盘，智能存储卡(Smart Media Card，SMC)，安全数字(Secure Digital，SD)卡，闪存卡(Flash Card)等。进一步地，存储器110还可以既包括电子设备的内部存储单元也包括外部存储设备。存储器110不仅可以用于存储安装于电子设备的应用软件及各类数据，例如：执行漏洞处理方法的程序的代码等，还可以用于暂时地存储已经输出或者将要输出的数据。本实施例中，存储器110至少用于存储以下计算机程序1101，其中，该计算机程序被处理器111加载并执行之后，能够实现前述任一实施例公开的高光谱遥感图像识别方法的相关步骤。另外，存储器110所存储的资源还可以包括操作系统1102和数据1103等，存储方式可以是短暂存储或者永久存储。其中，操作系统1102可以包括Windows、Unix、Linux等。数据1103可以包括但不限于高光谱遥感图像识别结果对应的数据等。Memory 110 may include one or more computer-readable storage media, which may be non-transitory. The memory 110 may also include high-speed random access memory and non-volatile memory, such as one or more magnetic disk storage devices, flash memory storage devices. The memory 110 in some embodiments may be an internal storage unit of the electronic device, such as a hard drive of a server. In other embodiments, the memory 110 may also be an external storage device of an electronic device, such as a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (SD) card, or a flash memory equipped on a server. Flash Card, etc. Further, the memory 110 may also include both an internal storage unit of the electronic device and an external storage device. The memory 110 can not only be used to store application software and various types of data installed on the electronic device, such as codes for executing programs for vulnerability processing methods, etc., but can also be used to temporarily store data that has been output or is to be output. In this embodiment, the memory 110 is at least used to store the following computer program 1101. After the computer program is loaded and executed by the processor 111, the relevant steps of the hyperspectral remote sensing image recognition method disclosed in any of the foregoing embodiments can be implemented. In addition, the resources stored in the memory 110 may also include the operating system 1102, data 1103, etc., and the storage method may be temporary storage or permanent storage. Among them, the operating system 1102 may include Windows, Unix, Linux, etc. Data 1103 may include but is not limited to data corresponding to hyperspectral remote sensing image recognition results, etc.
在一些实施例中，上述电子设备还可包括有显示屏112、输入输出接口113、通信接口114或者称为网络接口、电源115以及通信总线116。其中，显示屏112、输入输出接口113比如键盘(Keyboard)属于用户接口，可选的用户接口还可以包括标准的有线接口、无线接口等。可选地，在一些实施例中，显示器可以是LED显示器、液晶显示器、触控式液晶显示器以及OLED(Organic Light-Emitting Diode，有机发光二极管)触摸器等。显示器也可以适当的称为显示屏或显示单元，用于显示在电子设备中处理的信息以及用于显示可视化的用户界面。通信接口114可选的可以包括有线接口和/或无线接口，如WI-FI接口、蓝牙接口等，通常用于在电子设备与其他电子设备之间建立通信连接。通信总线116可以是外设部件互连标准(peripheral component interconnect，简称PCI)总线或扩展工业标准结构(extended industry standard architecture，简称EISA)总线等。该总线可以分为地址总线、数据总线、控制总线等。为便于表示，图11中仅用一条粗线表示，但并不表示仅有一根总线或一种类型的总线。In some embodiments, the above-mentioned electronic device may also include a display screen 112, an input and output interface 113, a communication interface 114 or also called a network interface, a power supply 115 and a communication bus 116. Among them, the display screen 112 and the input and output interface 113 such as a keyboard belong to the user interface, and optional user interfaces may also include standard wired interfaces, wireless interfaces, etc. Alternatively, in some embodiments, the display may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, an OLED (Organic Light-Emitting Diode, organic light-emitting diode) touch device, or the like. A display, which may also appropriately be called a display screen or display unit, is used for displaying information processed in the electronic device and for displaying a visual user interface. The communication interface 114 may optionally include a wired interface and/or a wireless interface, such as a WI-FI interface, a Bluetooth interface, etc., and is generally used to establish communication connections between electronic devices and other electronic devices. The communication bus 116 may be a peripheral component interconnect (PCI) bus or an extended industry standard architecture (EISA) bus, or the like. The bus can be divided into address bus, data bus, control bus, etc. For ease of presentation, only one thick line is used in Figure 11, but it does not mean that there is only one bus or one type of bus.
本领域技术人员可以理解，图11中示出的结构并不构成对该电子设备的限定，可以包括比图示更多或更少的组件，例如还可包括实现各类功能的传感器117。Those skilled in the art can understand that the structure shown in FIG. 11 does not constitute a limitation of the electronic device, and may include more or fewer components than shown in the figure, for example, it may also include sensors 117 that implement various functions.
本发明实施例所述电子设备的各功能模块的功能可根据上述方法实施例中的方法具体实现，其具体实现过程可以参照上述方法实施例的相关描述，此处不再赘述。The functions of each functional module of the electronic device described in the embodiment of the present invention can be specifically implemented according to the method in the above method embodiment. For the specific implementation process, reference can be made to the relevant description of the above method embodiment, which will not be described again here.
由上可知，本发明实施例可有效提升高光谱遥感图像的识别精准度。It can be seen from the above that embodiments of the present invention can effectively improve the recognition accuracy of hyperspectral remote sensing images.
可以理解的是，如果上述实施例中的高光谱遥感图像识别方法以软件功能单元的形式实现并作为独立的产品销售或使用时，可以存储在一个计算机可读取存储介质中。基于这样的理解，本申请的技术方案本质上或者说对现有技术做出贡献的部分或者该技术方案的全部或部分可以以软件产品的形式体现出来，该计算机软件产品存储在一个存储介质中，执行本申请各个实施例方法的全部或部分步骤。而前述的存储介质包括：U盘、移动硬盘、只读存储器(Read-Only Memory，ROM)、随机存取存储器(Random Access Memory，RAM)、电可擦除可编程ROM、寄存器、硬盘、多媒体卡、卡型存储器(例如SD或DX存储器等)、磁性存储器、可移动磁盘、CD-ROM、磁碟或者光盘等各种可以存储程序代码的介质。It can be understood that if the hyperspectral remote sensing image recognition method in the above embodiment is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application is essentially or contributes to the existing technology, or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium , execute all or part of the steps of the methods of various embodiments of this application. The aforementioned storage media include: U disk, mobile hard disk, read-only memory (ROM), random access memory (Random Access Memory, RAM), electrically erasable programmable ROM, register, hard disk, multimedia Cards, card-type memories (such as SD or DX memories, etc.), magnetic memories, removable disks, CD-ROMs, magnetic disks or optical disks and other media that can store program codes.
基于此，本发明实施例还提供了一种可读存储介质，存储有计算机程序，所述计算机程序被处理器执行时如上任意一实施例所述高光谱遥感图像识别方法的步骤。Based on this, embodiments of the present invention also provide a readable storage medium that stores a computer program. When the computer program is executed by a processor, the steps of the hyperspectral remote sensing image recognition method described in any of the above embodiments are performed.
本说明书中各个实施例采用递进的方式描述，每个实施例重点说明的都是与其它实施例的不同之处，各个实施例之间相同或相似部分互相参见即可。对于实施例公开的硬件包括装置及电子设备而言，由于其与实施例公开的方法相对应，所以描述的比较简单，相关之处参见方法部分说明即可。Each embodiment in this specification is described in a progressive manner. Each embodiment focuses on its differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other. As for the hardware including devices and electronic equipment disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple. For relevant details, please refer to the description in the method section.
专业人员还可以进一步意识到，结合本文中所公开的实施例描述的各示例的单元及算法步骤，能够以电子硬件、计算机软件或者二者的结合来实现，为了清楚地说明硬件和软件的可互换性，在上述说明中已经按照功能一般性地描述了各示例的组成及步骤。这些功能究竟以硬件还是软件方式来执行，取决于技术方案的特定应用和设计约束条件。专业技术人员可以对每个特定的应用来使用不同方法来实现所描述的功能，但是这种实现不应认为超出本发明的范围。Those skilled in the art may further realize that the units and algorithm steps of each example described in connection with the embodiments disclosed herein can be implemented by electronic hardware, computer software, or a combination of both. In order to clearly illustrate the possible functions of hardware and software, Interchangeability, in the above description, the composition and steps of each example have been generally described according to functions. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Skilled artisans may implement the described functionality using different methods for each specific application, but such implementations should not be considered to be beyond the scope of the present invention.
以上对本申请所提供的一种高光谱遥感图像识别方法、装置、电子设备及可读存储介质进行了详细介绍。本文中应用了具体个例对本发明的原理及实施方式进行了阐述，以上实施例的说明只是用于帮助理解本发明的方法及其核心思想。应当指出，对于本技术领域的普通技术人员来说，在不脱离本发明原理的前提下，还可以对本申请进行若干改进和修饰，这些改进和修饰也落入本申请权利要求的保护范围内。The above describes in detail the hyperspectral remote sensing image recognition method, device, electronic equipment and readable storage medium provided by this application. This article uses specific examples to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method and the core idea of the present invention. It should be noted that for those of ordinary skill in the art, several improvements and modifications can be made to the present application without departing from the principles of the present invention, and these improvements and modifications also fall within the protection scope of the claims of the present application.
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111137984.4ACN113822207B (en) | 2021-09-27 | 2021-09-27 | Hyperspectral remote sensing image identification method and device, electronic equipment and storage medium |
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111137984.4ACN113822207B (en) | 2021-09-27 | 2021-09-27 | Hyperspectral remote sensing image identification method and device, electronic equipment and storage medium |
| Publication Number | Publication Date |
|---|---|
| CN113822207A CN113822207A (en) | 2021-12-21 |
| CN113822207Btrue CN113822207B (en) | 2024-02-02 |
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202111137984.4AActiveCN113822207B (en) | 2021-09-27 | 2021-09-27 | Hyperspectral remote sensing image identification method and device, electronic equipment and storage medium |
| Country | Link |
|---|---|
| CN (1) | CN113822207B (en) |
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114387225B (en)* | 2021-12-23 | 2024-12-10 | 沈阳东软智能医疗科技研究院有限公司 | Bone joint image recognition method, device, electronic device and readable medium |
| CN114022748B (en)* | 2022-01-06 | 2022-04-08 | 深圳市海清视讯科技有限公司 | Target identification method, device, equipment and storage medium |
| CN114399696A (en)* | 2022-01-20 | 2022-04-26 | 中国农业银行股份有限公司 | Target detection method and device, storage medium and electronic equipment |
| CN114581789B (en)* | 2022-02-21 | 2025-03-14 | 山东锋士信息技术有限公司 | A hyperspectral image classification method and system |
| CN115424089A (en)* | 2022-08-26 | 2022-12-02 | 中国电力工程顾问集团华北电力设计院有限公司 | Image ground object identification method using UNET model |
| CN115908829B (en)* | 2022-09-09 | 2025-08-26 | 广州大学 | 3D point cloud object detection method based on second-order multi-attention mechanism of point column |
| CN115731391B (en)* | 2022-10-08 | 2025-10-03 | 北京瀚景锦河科技有限公司 | Target area extraction device and extraction method |
| CN116051896B (en)* | 2023-01-28 | 2023-06-20 | 西南交通大学 | A Lightweight Hybrid Tensor Neural Network for Hyperspectral Image Classification |
| CN116630700A (en)* | 2023-05-22 | 2023-08-22 | 齐鲁工业大学(山东省科学院) | Remote sensing image classification method based on introduction channel-space attention mechanism |
| CN116958696B (en)* | 2023-07-31 | 2025-10-03 | 长安大学 | A method and system for classifying objects using hyperspectral remote sensing technology from unmanned aerial vehicles |
| CN116773534B (en)* | 2023-08-15 | 2024-03-05 | 宁德思客琦智能装备有限公司 | Detection method and device, electronic equipment and computer readable medium |
| CN118823486A (en)* | 2024-09-14 | 2024-10-22 | 北京观微科技有限公司 | Hyperspectral image classification method, device and electronic equipment |
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110148103A (en)* | 2019-04-29 | 2019-08-20 | 中国科学院西安光学精密机械研究所 | EO-1 hyperion and Multispectral Image Fusion Methods, computer readable storage medium, electronic equipment based on combined optimization |
| CN110705457A (en)* | 2019-09-29 | 2020-01-17 | 核工业北京地质研究院 | Remote sensing image building change detection method |
| CN111626267A (en)* | 2019-09-17 | 2020-09-04 | 山东科技大学 | Hyperspectral remote sensing image classification method using void convolution |
| CN111914907A (en)* | 2020-07-13 | 2020-11-10 | 河海大学 | Hyperspectral image classification method based on deep learning space-spectrum combined network |
| KR102197297B1 (en)* | 2019-09-27 | 2020-12-31 | 서울대학교산학협력단 | Change detection method using recurrent 3-dimensional fully convolutional network for hyperspectral image |
| CN112906662A (en)* | 2021-04-02 | 2021-06-04 | 海南长光卫星信息技术有限公司 | Method, device and equipment for detecting change of remote sensing image and storage medium |
| CN113011499A (en)* | 2021-03-22 | 2021-06-22 | 安徽大学 | Hyperspectral remote sensing image classification method based on double-attention machine system |
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110148103A (en)* | 2019-04-29 | 2019-08-20 | 中国科学院西安光学精密机械研究所 | EO-1 hyperion and Multispectral Image Fusion Methods, computer readable storage medium, electronic equipment based on combined optimization |
| CN111626267A (en)* | 2019-09-17 | 2020-09-04 | 山东科技大学 | Hyperspectral remote sensing image classification method using void convolution |
| KR102197297B1 (en)* | 2019-09-27 | 2020-12-31 | 서울대학교산학협력단 | Change detection method using recurrent 3-dimensional fully convolutional network for hyperspectral image |
| CN110705457A (en)* | 2019-09-29 | 2020-01-17 | 核工业北京地质研究院 | Remote sensing image building change detection method |
| CN111914907A (en)* | 2020-07-13 | 2020-11-10 | 河海大学 | Hyperspectral image classification method based on deep learning space-spectrum combined network |
| CN113011499A (en)* | 2021-03-22 | 2021-06-22 | 安徽大学 | Hyperspectral remote sensing image classification method based on double-attention machine system |
| CN112906662A (en)* | 2021-04-02 | 2021-06-04 | 海南长光卫星信息技术有限公司 | Method, device and equipment for detecting change of remote sensing image and storage medium |
| Title |
|---|
| Tri_training for remote sensing classification based on multi-scale homogeneity;J. Zhu, K. Tan and Q. Du;《2016 IEEE International Geoscience and Remote Sensing Symposium》;全文* |
| 基于卷积神经网络的遥感图像降噪;潘凯;侯亮;;现代信息科技(第12期);全文* |
| 多尺度特征融合空洞卷积 ResNet遥感图像建筑物分割;徐胜军;欧阳朴衍;郭学源;Taha Muthar Khan;段中兴;;光学精密工程(第07期);全文* |
| Publication number | Publication date |
|---|---|
| CN113822207A (en) | 2021-12-21 |
| Publication | Publication Date | Title |
|---|---|---|
| CN113822207B (en) | Hyperspectral remote sensing image identification method and device, electronic equipment and storage medium | |
| CN113822209B (en) | Hyperspectral image recognition method and device, electronic equipment and readable storage medium | |
| CN113159147B (en) | Image recognition method and device based on neural network and electronic equipment | |
| CN110837836B (en) | Semi-supervised semantic segmentation method based on maximized confidence | |
| RU2743931C1 (en) | Image processing system and method using deep neural networks | |
| US20220277572A1 (en) | Artificial intelligence-based image processing method, apparatus, device, and storage medium | |
| JP6961011B2 (en) | Systems and methods for data management | |
| CN109685819B (en) | A 3D Medical Image Segmentation Method Based on Feature Enhancement | |
| CN107767408B (en) | Image processing method, processing device and processing equipment | |
| CN109300530B (en) | Pathological picture recognition method and device | |
| US12333853B2 (en) | Face parsing method and related devices | |
| CN111553406A (en) | Target detection system, method and terminal based on improved YOLO-V3 | |
| CN110298396A (en) | Hyperspectral image classification method based on deep learning multiple features fusion | |
| US11144291B1 (en) | Loop-oriented neural network compilation | |
| CN108229341A (en) | Sorting technique and device, electronic equipment, computer storage media, program | |
| CN114419372B (en) | Multi-scale point cloud classification method and system | |
| CN105144239A (en) | Image processing device, program, and image processing method | |
| CN112560967B (en) | Multi-source remote sensing image classification method, storage medium and computing device | |
| US12248782B2 (en) | Accelerating processor based artificial neural network computation | |
| CN113705749B (en) | Two-dimensional code recognition method, device, equipment and storage medium based on deep learning | |
| CN111950702A (en) | A kind of neural network structure determination method and device | |
| Zhou et al. | Fast minutiae extractor using neural network | |
| CN120340023A (en) | Cell nucleus segmentation method, device and computer equipment based on multimodal information | |
| CN116168420A (en) | Lightweight gesture image recognition method based on improved GhostNet | |
| CN108363962A (en) | A kind of method for detecting human face and system based on multi-level features deep learning |
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant | ||
| PE01 | Entry into force of the registration of the contract for pledge of patent right | ||
| PE01 | Entry into force of the registration of the contract for pledge of patent right | Denomination of invention:Hyperspectral remote sensing image recognition method, device, electronic equipment, and storage medium Granted publication date:20240202 Pledgee:Hainan Rural Commercial Bank Co.,Ltd. Wenchang Nanhai Branch Pledgor:Hainan Changguang Satellite Information Technology Co.,Ltd. Registration number:Y2025980017038 |