CN115909086A - SAR target detection and recognition method based on multi-level enhanced network - Google Patents
SAR target detection and recognition method based on multi-level enhanced networkDownload PDFInfo
- Publication number
- CN115909086A CN115909086ACN202211449058.5ACN202211449058ACN115909086ACN 115909086 ACN115909086 ACN 115909086ACN 202211449058 ACN202211449058 ACN 202211449058ACN 115909086 ACN115909086 ACN 115909086A
- Authority
- CN
- China
- Prior art keywords
- module
- network
- convolution
- layer
- loss
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Images
Landscapes
- Image Analysis (AREA)
Abstract
Description
Translated fromChinese技术领域technical field
本发明属于雷达遥感技术领域，更进一步涉及一种SAR目标检测识别方法，可用于战场侦察和态势感知。The invention belongs to the technical field of radar remote sensing, and further relates to a SAR target detection and recognition method, which can be used for battlefield reconnaissance and situation awareness.
背景技术Background technique
合成孔径雷达SAR是一种主动式微波成像传感器，通过发射大时宽-带宽积信号并利用孔径合成获得二维高分辨图像。与光学和红外传感器相比，SAR具有全天时、全天候、作用距离远、穿透力强等独特优势，成为对地观测的重要手段，并广泛应用于军事和民用领域。随着SAR系统的不断完善和SAR成像水平的不断提高，SAR图像解译技术逐渐受到相关领域学者及研究人员的关注。作为其中的难点和关键步骤，对重点目标的精确检测和识别具有重要意义和研究价值。Synthetic Aperture Radar (SAR) is an active microwave imaging sensor that emits a large time-width-bandwidth product signal and uses aperture synthesis to obtain a two-dimensional high-resolution image. Compared with optical and infrared sensors, SAR has unique advantages such as all-day, all-weather, long range, and strong penetrating power. It has become an important means of earth observation and is widely used in military and civilian fields. With the continuous improvement of SAR system and the continuous improvement of SAR imaging level, SAR image interpretation technology has gradually attracted the attention of scholars and researchers in related fields. As one of the difficulties and key steps, the precise detection and recognition of key targets is of great significance and research value.
传统的SAR目标检测识别方法主要采用三级处理流程，包括目标检测、目标鉴别和目标识别。其中，目标检测主要基于恒虚警率CFAR算法，在假设背景杂波满足某种概率分布模型的前提下，该算法通过在SAR图像上进行滑窗并将选定值与自适应阈值进行比较来实现目标检测，然而，由于难以对非均匀强杂波背景进行有效建模，该算法在复杂场景中适应性较差，检测精度较低。目标鉴别和目标识别主要依据图像的统计信息和物理特性进行手工特征设计和分类器构建，然而，这需要较强的专业知识和专家经验，且算法的精度和灵活性较差，难以在实际应用中达到理想的效果。此外，传统的三级处理流程中各环节之间的低效连接也极大地降低了算法的运算效率，亟需开发新的架构体系。Traditional SAR target detection and recognition methods mainly adopt three-level processing flow, including target detection, target discrimination and target recognition. Among them, the target detection is mainly based on the constant false alarm rate CFAR algorithm. Under the premise that the background clutter satisfies a certain probability distribution model, the algorithm performs sliding windows on the SAR image and compares the selected value with the adaptive threshold. However, due to the difficulty in effectively modeling the non-uniform strong clutter background, the algorithm has poor adaptability in complex scenes and low detection accuracy. Target identification and target recognition are mainly based on the statistical information and physical characteristics of the image for manual feature design and classifier construction. However, this requires strong professional knowledge and expert experience, and the accuracy and flexibility of the algorithm are poor, which is difficult to apply in practice. to achieve the desired effect. In addition, the inefficient connection between the various links in the traditional three-level processing flow also greatly reduces the computational efficiency of the algorithm, and it is urgent to develop a new architecture system.
近年来，随着深度学习技术的不断发展，以深度神经网络为主体的目标检测识别方法在计算机视觉领域取得了重大突破。得益于深度网络的特殊结构，这些算法能够同时预测目标的位置和类别信息，而无需进行多级处理，显著提高了检测识别的性能和效率。目前，主流的检测识别算法可分为单阶段和双阶段两种模式，前者直接将网络所提特征进行解码来实现目标的检测识别，具有更快的推理速度，代表算法有YOLO、SSD和RetinaNet等；后者则增加了候选区域提取阶段，即首先通过深度网络从图像中提取可能包含重点目标的候选区域，然后再对候选区域的位置进一步修正，并得到识别结果，代表算法有R-CNN、Faster R-CNN和Cascade R-CNN等。相比于单阶段算法，双阶段算法具有更高的检测识别精度。In recent years, with the continuous development of deep learning technology, object detection and recognition methods based on deep neural networks have made major breakthroughs in the field of computer vision. Thanks to the special structure of the deep network, these algorithms can simultaneously predict the location and category information of the target without multi-level processing, which significantly improves the performance and efficiency of detection and recognition. At present, the mainstream detection and recognition algorithms can be divided into two modes: single-stage and two-stage. The former directly decodes the features extracted by the network to achieve target detection and recognition. It has faster reasoning speed. Representative algorithms include YOLO, SSD and RetinaNet. etc.; the latter adds a candidate area extraction stage, that is, first extracts candidate areas that may contain key targets from the image through a deep network, and then further corrects the position of the candidate area, and obtains the recognition result. The representative algorithm is R-CNN , Faster R-CNN and Cascade R-CNN, etc. Compared with the single-stage algorithm, the two-stage algorithm has higher detection and recognition accuracy.
虽然上述基于深度学习的方法为SAR目标检测识别提供了可行的途径，然而，与光学图像相比，SAR图像场景更为复杂、不同类别目标的相似性较高，且受相干斑噪声的影响，目标的边缘不清晰，因此依然存在复杂环境不稳健、相似类别难以区分的问题。Although the above deep learning-based methods provide a feasible way for SAR target detection and recognition, compared with optical images, SAR image scenes are more complex, the similarity of different types of targets is higher, and they are affected by coherent speckle noise. The edge of the target is not clear, so there are still problems that the complex environment is not robust and similar categories are difficult to distinguish.
申请号为201710461303.7的专利文献公开了“一种SAR图像目标检测识别一体化方法”，其首先通过卷积神经网络提取SAR图像特征，然后基于该特征生成可能包含目标的候选区域，最后利用全连接网络预测感兴趣区域的类别和位置信息来实现SAR目标的检测识别。由于该方法未针对SAR图像的特性进行相关优化，因此在复杂环境中不稳健，预测边界框不准确，虚警率和漏检率高，且难以对目标的细粒度特征进行有效挖掘，导致检测识别准确率较低。The patent document with the application number 201710461303.7 discloses "an integrated method for SAR image target detection and recognition", which firstly extracts SAR image features through a convolutional neural network, then generates candidate regions that may contain targets based on the features, and finally uses fully connected The network predicts the category and location information of the region of interest to realize the detection and recognition of SAR targets. Since this method is not optimized for the characteristics of SAR images, it is not robust in complex environments, the predicted bounding box is inaccurate, the false alarm rate and missed detection rate are high, and it is difficult to effectively mine the fine-grained features of the target, resulting in detection The recognition accuracy is low.
发明内容Contents of the invention
本发明的目的在于针对上述现有技术的不足，提出一种基于多级增强网络的SAR目标检测识别方法，以提高复杂环境中检测识别的稳健性，降低虚警率和漏检率，增强特征的可分性，显著提升SAR目标检测识别精度。The purpose of the present invention is to address the deficiencies of the above-mentioned prior art, and propose a SAR target detection and recognition method based on a multi-level enhanced network to improve the robustness of detection and recognition in complex environments, reduce false alarm rates and missed detection rates, and enhance feature The separability can significantly improve the accuracy of SAR target detection and recognition.
本发明的技术思路是，通过设计多级增强网络，提升复杂环境下SAR目标检测识别性能，其实现步骤包括如下：The technical idea of the present invention is to improve the performance of SAR target detection and recognition in complex environments by designing a multi-level enhanced network, and the implementation steps include the following:
(1)获取具有多类目标的SAR图像，对每幅SAR图像中的目标位置和目标类别进行标记，并将标记好的SAR图像进行随机划分，得到训练集和测试集；(1) Obtain SAR images with multiple types of targets, mark the target position and target category in each SAR image, and randomly divide the marked SAR images to obtain a training set and a test set;
(2)构建多级增强网络：(2) Build a multi-level enhanced network:
(2a)建立依次进行多尺度变换、随机翻转、随机旋转、幂次变换、随机噪声操作的数据级增强模块；(2a) Establish a data-level enhancement module that sequentially performs multi-scale transformation, random flip, random rotation, power transformation, and random noise operations;
(2b)建立由主干网络A、特征优选金字塔网络F、递归主干网络Q和递归特征优选金字塔网络E级联组成的特征级增强模块；(2b) Establish a feature-level enhancement module consisting of backbone network A, feature-optimized pyramid network F, recursive backbone network Q and recursive feature-optimized pyramid network E cascade;
(2c)选用现有区域建议网络组成区域建议模块G，并选用交叉熵损失和CIOU损失作为其分类和回归损失；(2c) Select the existing region proposal network to form the region proposal module G, and select the cross-entropy loss and CIOU loss as its classification and regression losses;
(2d)建立由三个决策器d1,d2,d3级联组成的决策级增强模块D，并选用交叉熵损失和CIOU损失作为其分类和回归损失；(2d) Establish a decision-level enhancement module D consisting of cascades of three decision makers d1 , d2 , and d3 , and select cross-entropy loss and CIOU loss as its classification and regression losses;
(2e)将数据级增强模块、特征级增强模块、区域建议模块、决策级增强模块依次级联，构成多级增强网络；(2e) cascading the data-level enhancement module, feature-level enhancement module, region suggestion module, and decision-level enhancement module sequentially to form a multi-level enhancement network;
(3)对多级增强网络进行训练：(3) Train the multi-level enhanced network:
(3a)从训练集中随机采样一组SAR图像输入到多级增强网络中，计算损失，基于该损失，通过随机梯度下降算法更新网络参数；(3a) Randomly sample a group of SAR images from the training set and input them into the multi-level enhanced network, calculate the loss, and update the network parameters through the stochastic gradient descent algorithm based on the loss;
(3b)重复(3a)过程，直至网络收敛，得到训练好的多级增强网络；(3b) Repeat the process of (3a) until the network converges to obtain a trained multi-level enhanced network;
(4)将测试集中的SAR图像输入到训练好的多级增强网络中，得到检测识别结果。(4) Input the SAR images in the test set into the trained multi-level enhanced network to obtain the detection and recognition results.
本发明与现有技术相比，具有以下优点：Compared with the prior art, the present invention has the following advantages:
第一，本发明通过设计数据级增强模块，模拟目标的尺度和方位变化以及杂波噪声干扰，提高了复杂环境中算法的稳健性，降低虚警率和漏检率。First, the present invention improves the robustness of the algorithm in a complex environment and reduces the false alarm rate and missed detection rate by designing a data-level enhancement module to simulate the scale and orientation changes of the target and clutter noise interference.
第二，本发明通过设计特征级增强模块，充分挖掘SAR图像中目标的细粒度特征，增强了相似类别目标的可分性；Second, the present invention fully excavates the fine-grained features of targets in SAR images by designing a feature-level enhancement module, and enhances the separability of targets of similar categories;
第三，本发明通过设计决策级增强模块，对预测结果进行多次微调，以逐步缩小目标预测位置与真实值之间的偏差，有效抑制了SAR目标边缘模糊对检测精度的影响。Thirdly, the present invention fine-tunes the prediction results multiple times by designing a decision-making level enhancement module to gradually reduce the deviation between the predicted position of the target and the real value, effectively suppressing the influence of blurred SAR target edges on the detection accuracy.
附图说明Description of drawings
图1为本发明的实现流程图；Fig. 1 is the realization flowchart of the present invention;
图2为本发明中构建的多级增强网路模型图；Fig. 2 is the multi-level enhanced network model figure that builds among the present invention;
图3为本发明的仿真结果图。Fig. 3 is a simulation result diagram of the present invention.
具体实施方式Detailed ways
下面结合附图对本发明的实例和效果作进一步详细说明。The examples and effects of the present invention will be described in further detail below in conjunction with the accompanying drawings.
参考图1，本实例中基于多级增强网络的SAR目标检测识别方法依次包括数据标注和划分、构建多级增强网络、训练多级增强网络、获取SAR目标检测识别结果，具体实现如下：Referring to Figure 1, the SAR target detection and recognition method based on the multi-level enhanced network in this example includes data labeling and division, building a multi-level enhanced network, training the multi-level enhanced network, and obtaining the SAR target detection and recognition results. The specific implementation is as follows:
步骤一，数据标注和划分。Step 1, data labeling and division.
获取具有多类目标的SAR图像，对每幅SAR图像中的目标位置和目标类别进行标注，并将标注好的SAR图像按照7:3进行随机划分，得到训练集和测试集。Obtain SAR images with multiple types of targets, mark the target position and target category in each SAR image, and divide the marked SAR images randomly according to 7:3 to obtain training set and test set.
在本发明的实施例中，SAR图像来自高分3号卫星上的星载雷达，SAR图像尺度包括600×600、1024×1024、2048×2048三种，共七类飞机目标，训练集图像数量为1400张，测试集图像数量为600张。In the embodiment of the present invention, the SAR image comes from the spaceborne radar on the Gaofen 3 satellite. The SAR image scale includes three types: 600×600, 1024×1024, and 2048×2048. There are seven types of aircraft targets in total. The number of images in the training set The number of images in the test set is 1400, and the number of images in the test set is 600.
步骤二，构建多级增强网络。Step two, build a multi-level enhanced network.
参考图2，本步骤构建的多级增强网络包括依次级联的数据级增强模块、特征级增强模块、区域建议模块和决策级增强模块，其构建步骤如下：Referring to Figure 2, the multi-level enhanced network constructed in this step includes sequentially cascaded data-level enhanced modules, feature-level enhanced modules, region proposal modules, and decision-level enhanced modules. The construction steps are as follows:
2.1)建立依次进行多尺度变换、随机翻转、随机旋转、幂次变换、随机噪声操作的数据级增强模块；在本发明的实施例中，训练时执行上述所有操作，测试时仅执行多尺度变换和随机翻转操作，其中多尺度变换的尺度包括1024×1024、1088×1088、1152×1152，随机翻转的方向包括水平、垂直、对角，随机旋转的角度包括90°、180°、270°，幂次变换的系数为[0.8,1.2]之间的随机值。2.1) Establish a data-level enhancement module that sequentially performs multi-scale transformation, random flip, random rotation, power transformation, and random noise operations; in the embodiment of the present invention, all the above-mentioned operations are performed during training, and only multi-scale transformation is performed during testing And random flip operation, where the scale of multi-scale transformation includes 1024×1024, 1088×1088, 1152×1152, the direction of random flip includes horizontal, vertical, diagonal, and the angle of random rotation includes 90°, 180°, 270°, The coefficient of the power transformation is a random value between [0.8,1.2].
2.2)建立由主干网络A、特征优选金字塔网络F、递归主干网络Q和递归特征优选金字塔网络E级联组成的特征级增强模块：2.2) Establish a feature-level enhancement module consisting of backbone network A, feature optimization pyramid network F, recursive backbone network Q and recursive feature optimization pyramid network E cascade:
2.2.1)建立包括5个级联卷积模块a1,a2,a3,a4,a5的主干网络A，其中：2.2.1) Establish a backbone network A including 5 cascaded convolution modules a1 , a2 , a3 , a4 , a5 , where:
第一个卷积模块a1由7×7标准卷积层、批归一化层、ReLU激活层以及最大池化下采样层级联组成；The first convolution module a1 consists of a 7×7 standard convolution layer, a batch normalization layer, a ReLU activation layer, and a cascade of max pooling downsampling layers;
第二个卷积模块a2由3个残差块级联组成；The second convolutional modulea2 consists of 3 residual block cascades;
第三个卷积模块a3由4个残差块级联组成；The third convolutional module a3 consists of 4 residual block cascades;
第四个卷积模块a4由6个残差块级联组成；The fourth convolution module a4 consists of 6 residual block cascades;
第五个卷积模块a5由3个残差块级联组成；The fifth convolution module a5 consists of 3 residual block cascades;
在本发明的实施例中，主干网络A用于提取多尺度特征图，对于宽和高分别为W和H的输入SAR图像
2.2.2)建立包括4个并联分支f1,f2,f3,f4的特征优选金字塔网络F，每个分支fi由特征优选模块fis和特征融合模块fiu级联组成，每个特征优选模块fis,由两个并联子分支
在本发明的实施例中，特征优选金字塔网络F用于对主干网络输出的多尺度特征图Yi(i＝2,3,4,5)进行特征优选和特征融合，Yi经过特征优选后表示为：In the embodiment of the present invention, the feature optimization pyramid network F is used to perform feature optimization and feature fusion on the multi-scale feature map Yi (i=2, 3, 4, 5) output by the backbone network. After the feature optimization, Yi Expressed as:
其中，Si为第i个特征优选模块fis的输出特征图，⊙表示逐通道相乘操作，GAP(·)表示全局平均池化，Conv1d(·)表示1维卷积，σ(·)表示Sigmoid函数，Conv1×1(·)表示1×1标准卷积；Si经过特征融合后表示为：Among them, Si is the output feature map of the i-th feature selection module fis , ⊙ represents channel-by-channel multiplication operation, GAP(·) represents global average pooling, Conv1d(·) represents 1-dimensional convolution, σ(· ) represents the Sigmoid function, Conv1×1(·) represents the 1×1 standard convolution; Si is expressed as:
其中，Ui为第i个特征融合模块fiu的输出特征图，Conv3×3(·)表示3×3标准卷积，Up(·)表示双线性插值上采样函数。Among them, Ui is the output feature map of the i-th feature fusion module fiu , Conv3×3(·) represents the 3×3 standard convolution, and Up(·) represents the bilinear interpolation upsampling function.
2.2.3)建立包括5个级联卷积模块q1,q2,q3,q4,q5的递归主干网络Q，其中：2.2.3) Establish a recursive backbone network Q including 5 cascaded convolution modules q1 , q2 , q3 , q4 , q5 , where:
第1个卷积模块q1由7×7标准卷积层、批归一化层、ReLU激活层以及最大池化下采样层级联组成；The first convolution module q1 consists of a 7×7 standard convolution layer, a batch normalization layer, a ReLU activation layer, and a cascade of maximum pooling downsampling layers;
第2个卷积模块q2由3个级联的残差块和一个1×1标准卷积层并联组成；The second convolutional module q2 consists of 3 cascaded residual blocks and a 1×1 standard convolutional layer connected in parallel;
第3个卷积模块q3由4个级联的残差块和一个1×1标准卷积层并联组成；The third convolutional module q3 consists of 4 cascaded residual blocks and a 1×1 standard convolutional layer connected in parallel;
第4个卷积模块q4由6个级联的残差块和一个1×1标准卷积层并联组成；The fourth convolutional module q4 consists of 6 cascaded residual blocks and a 1×1 standard convolutional layer connected in parallel;
第5个卷积模块q5由3个级联的残差块和一个1×1标准卷积层并联组成；The fifth convolutional module q5 consists of 3 cascaded residual blocks and a 1×1 standard convolutional layer connected in parallel;
在本发明的实施例中，递归主干网络Q用于提取多尺度特征图，对于宽和高分别为W和H的输入SAR图像
其中，Zi为第i个卷积模块qi的输出特征图，Ui-1为特征优选金字塔网络第i-1个特征融合模块
2.2.4)建立与特征优选金字塔网络F的结构、参数相同的递归特征优选金字塔网络E。2.2.4) Establish a recursive feature optimization pyramid network E with the same structure and parameters as the feature optimization pyramid network F.
在本发明的实施例中，递归特征优选金字塔E用于对递归主干网络Q输出的多尺度特征图Zi(i＝2,3,4,5)进行特征优选和特征融合，Zi经过特征优选后表示为：In the embodiment of the present invention, the recursive feature optimization pyramid E is used to perform feature optimizationand feature fusion on the multi-scale feature map Zi (i=2, 3, 4, 5) output by the recurrent backbone network Q. After optimization, it is expressed as:
Wi＝fis(Zi+1)Wi ＝fis (Zi+1 )
＝Conv1×1(σ(Conv1d(GAP(Zi+1)))⊙Zi+1),i＝1,2,3,4=Conv1×1(σ(Conv1d(GAP(Zi+1 )))⊙Zi+1 ), i=1,2,3,4
其中，Wi为第i个特征优选模块fis的输出特征图，⊙表示逐通道相乘操作，GAP(·)表示全局平均池化，Conv1d(·)表示1维卷积，σ(·)表示Sigmoid函数，Conv1×1(·)表示1×1标准卷积；Wi经过特征融合后表示为：Among them, Wi is the output feature map of the i-th feature selection module fis , ⊙ represents channel-by-channel multiplication operation, GAP(·) represents global average pooling, Conv1d(·) represents 1-dimensional convolution, σ(· ) represents the Sigmoid function, Conv1×1( ) represents the 1×1 standard convolution; Wi is expressed as:
其中，Pi为第i个特征融合模块fiu的输出特征图，Conv3×3(·)表示3×3标准卷积，Up(·)表示双线性插值上采样函数。Among them, Pi is the output feature map of the i-th feature fusion module fiu , Conv3×3(·) represents the 3×3 standard convolution, and Up(·) represents the bilinear interpolation upsampling function.
2.3)建立由区域建议网络组成的区域建议模块G，区域建议网络依次包括候选区域生成模块、分类回归模块、后处理模块、正负样本分配模块，其中：2.3) Establish a region proposal module G consisting of a region proposal network. The region proposal network includes a candidate region generation module, a classification regression module, a post-processing module, and a positive and negative sample allocation module in turn, where:
候选区域生成模块，用于在输入特征图Pi(i＝1,2,3,4)每个点上生成面积为22i+2，宽、高比分别为1:2、1:1、2:1的三个矩形候选区域；The candidate area generation module is used to generate an area of 22i+2 on each point of the input feature map Pi (i=1,2,3,4), and the width and height ratios are 1:2, 1:1, 2:1 three rectangular candidate areas;
分类回归模块，包括两个并联的3×3标准卷积层g1和g2，第一卷积层g1用于调整候选区域的中心点位置和宽高，第二卷积层g2用于预测候选区域的目标置信度；Classification and regression module, including two parallel 3×3 standard convolutional layers g1 and g2 , the first convolutional layer g1 is used to adjust the center point position, width and height of the candidate area, and the second convolutional layer g2 is used to Target confidence in predicting candidate regions;
后处理模块，用于过滤冗余的候选区域并输出目标置信度最高的前N个候选区域{(pj,bj)|j＝1,2,...,N}，其中pj为第j个候选区域的目标置信度，bj＝(xj,yj,wj,hj)为第j个候选区域的边界框，(xj,yj)为边界框的中心点坐标，(wj,hj)为边界框的宽高；The post-processing module is used to filter redundant candidate regions and output the top N candidate regions {(pj ,bj )|j=1,2,...,N} with the highest target confidence, where pj is The target confidence of the j-th candidate area, bj = (xj ,yj ,wj ,hj ) is the bounding box of the j-th candidate area, (xj ,yj ) is the coordinates of the center point of the bounding box , (wj ,hj ) is the width and height of the bounding box;
正负样本分配模块，用于将候选区域分配为正负样本，即将与标注边界框交并比大于0.7的候选区域分配为正样本，将与标注边界框交并比小于0.3的候选区域分配为负样本；The positive and negative sample allocation module is used to assign the candidate area as a positive and negative sample, that is, the candidate area with an intersection ratio greater than 0.7 with the labeled bounding box is assigned as a positive sample, and the candidate area with an intersection ratio with the labeled bounding box that is less than 0.3 is assigned as negative sample;
在本发明的实施例中，区域建议模块G输出的边界框RG表示为：In an embodiment of the present invention, the bounding box RG output by the region proposal module G is expressed as:
RG＝{bj＝(xj,yj,wj,hj)|j＝1,2,...N}。RG ={bj =(xj ,yj ,wj ,hj )|j=1,2,...N}.
2.4)建立由三个级联子决策器d1,d2,d3组成的决策级增强模块D，每个子决策器结构相同，依次包括感兴趣区域提取模块、分类回归模块、正负样本分配模块，其中：2.4) Establish a decision-level enhancement module D consisting of three cascaded sub-decision makers d1 , d2 , and d3 . Each sub-decision maker has the same structure, including the region of interest extraction module, classification and regression module, and positive and negative sample distribution. module, where:
感兴趣区域提取模块，由自适应平均池化层和展平层级联组成，自适应平均池化层用于将候选区域特征池化为7×7的感兴趣特征，展平层用于展平感兴趣特征；The region of interest extraction module consists of an adaptive average pooling layer and a flattening layer cascade. The adaptive average pooling layer is used to pool the candidate region features into 7×7 features of interest, and the flattening layer is used to flatten features of interest;
分类回归模块，由两个线性层并联组成，第一个线性层调整边界框的位置，第二个线性层预测边界框的类别得分；The classification regression module consists of two linear layers in parallel, the first linear layer adjusts the position of the bounding box, and the second linear layer predicts the category score of the bounding box;
正负样本分配模块，用于将边界框分配为正负样本，即将与标注边界框的交并比大于阈值的边界框分配为正样本，将与标注边界框的交并比小于阈值的边界框分配为负样本，三个子决策器的阈值分别为0.5，0.6，0.7；The positive and negative sample allocation module is used to assign the bounding box as a positive and negative sample, that is, assign the bounding box whose intersection ratio with the labeled bounding box is greater than the threshold as a positive sample, and assign the bounding box whose intersection ratio with the labeled bounding box is smaller than the threshold Assigned as a negative sample, the thresholds of the three sub-decision makers are 0.5, 0.6, 0.7;
决策级增强模块输出的边界框RD表示为：The bounding box RD output by the decision-level augmentation module is expressed as:
其中N为边界框个数，
其中Nc为类别总数，
步骤三，训练多级增强网络。Step three, train the multi-level enhanced network.
3.1)从训练集中随机采样一组SAR图像输入到多级增强网络中计算其损失，基于该损失，通过随机梯度下降算法更新网络参数：3.1) Randomly sample a set of SAR images from the training set and input them into the multi-level enhanced network to calculate its loss. Based on the loss, update the network parameters through the stochastic gradient descent algorithm:
3.1.1)计算多级增强网络的损失
其中,NG为区域建议模块随机采样的候选区域个数，pm和
3.1.2)求解3.1.1)中多级增强网络损失
其中
3.1.3)根据3.1.2)中求解的梯度
其中θ′为更新后网络参数，θ为更新前网络参数；lr为学习率，其根据输入图像批次大小设置，在本发明的实施例中，设lr＝0.005。Wherein θ' is the network parameter after updating, θ is the network parameter before updating; lr is the learning rate, which is set according to the batch size of the input image, in the embodiment of the present invention, set lr=0.005.
3.2)重复步骤3.1)，直至网络收敛，得到训练好的多级增强网络。3.2) Repeat step 3.1) until the network converges to obtain a trained multi-level enhanced network.
步骤四，获取SAR目标检测识别结果。Step 4, obtaining the SAR target detection and recognition result.
将测试集中的SAR图像输入到训练好的多级增强网络中，得到检测识别结果。Input the SAR images in the test set into the trained multi-level enhanced network to obtain the detection and recognition results.
本发明的效果可以通过以下仿真实验进一步说明：Effect of the present invention can be further illustrated by following simulation experiments:
一、仿真实验条件：1. Simulation experiment conditions:
本发明仿真实验的软件平台为：Ubuntu 18.04操作系统和Pytorch 1.8.0，硬件配置为：Core i9-10980XE CPU和NVIDIA GeForce RTX 3090 GPU。The software platform of the simulation experiment of the present invention is: Ubuntu 18.04 operating system and Pytorch 1.8.0, and the hardware configuration is: Core i9-10980XE CPU and NVIDIA GeForce RTX 3090 GPU.
本发明的仿真实验使用高分3号SAR实测数据，场景类型为机场，图像分辨率为1m×1m，SAR图像数量为2000张，图像大小有600×600,1024×1024,2048×2048三种，目标类别数为7类，目标总数为6556，训练集图像数量为1400张，测试集图像数量为600张。The simulation experiment of the present invention uses the measured data of Gaofen No. 3 SAR, the scene type is an airport, the image resolution is 1m×1m, the number of SAR images is 2000, and the image sizes are 600×600, 1024×1024, and 2048×2048. , the number of target categories is 7, the total number of targets is 6556, the number of images in the training set is 1400, and the number of images in the test set is 600.
二、仿真内容与结果分析：2. Simulation content and result analysis:
在上述仿真条件下，用本发明和现有的“一种SAR图像目标检测识别一体化方法”分别在训练集上完成训练，然后随机挑选一张测试集图像输入到训练好的网络中，并将检测识别结果可视化到测试集图像上，结果如图3所示。其中，图3(a)为现有技术的检测识别结果，图3(b)为本发明的检测识别结果。图中绿色矩形框表示检测识别正确的目标，红色矩形框表示检测或识别错误的目标。Under the above simulation conditions, use the present invention and the existing "a kind of SAR image target detection and recognition integration method" to complete the training on the training set respectively, then randomly select a test set image and input it into the trained network, and The detection and recognition results are visualized on the test set images, and the results are shown in Figure 3. Among them, Fig. 3(a) is the detection and recognition result of the prior art, and Fig. 3(b) is the detection and recognition result of the present invention. The green rectangular box in the figure indicates the target detected and recognized correctly, and the red rectangular box indicates the target detected or recognized incorrectly.
对比图3(a)和3(b)可以看出，现有技术得到的检测识别结果中虚警和漏检较多，本发明所得检测识别结果虚警和漏检较少。Comparing Figures 3(a) and 3(b), it can be seen that there are more false alarms and missed detections in the detection and recognition results obtained in the prior art, and fewer false alarms and missed detections in the detection and recognition results obtained in the present invention.
对比本发明和现有技术在所有测试集图像上的检测识别指标，包括七类目标的平均精确率、平均召回率、平均F1分数、类平均精度，结果如表1所示：Comparing the detection and recognition indicators of the present invention and the prior art on all test set images, including the average precision rate, average recall rate, average F1 score, and class average precision of seven types of targets, the results are shown in Table 1:
表1Table 1
从表1中可以看出，本发明的平均精确率、平均召回率、平均F1分数以及类平均精度均高于现有技术，说明本发明检测识别性能显著优于现有技术。It can be seen from Table 1 that the average precision rate, average recall rate, average F1 score and class average precision of the present invention are all higher than those of the prior art, indicating that the detection and recognition performance of the present invention is significantly better than that of the prior art.
Claims (9)
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211449058.5ACN115909086B (en) | 2022-11-18 | 2022-11-18 | SAR target detection and recognition method based on multi-level enhancement network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211449058.5ACN115909086B (en) | 2022-11-18 | 2022-11-18 | SAR target detection and recognition method based on multi-level enhancement network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN115909086Atrue CN115909086A (en) | 2023-04-04 |
| CN115909086B CN115909086B (en) | 2025-08-05 |
Family
ID=86473982
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211449058.5AActiveCN115909086B (en) | 2022-11-18 | 2022-11-18 | SAR target detection and recognition method based on multi-level enhancement network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115909086B (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116664823A (en)* | 2023-05-31 | 2023-08-29 | 西安电子科技大学 | Small-sample SAR target detection and recognition method based on meta-learning and metric learning |
| CN116977870A (en)* | 2023-06-28 | 2023-10-31 | 西安电子科技大学 | SAR target increment identification method based on dynamic structure and multi-level distillation |
| CN117574161A (en)* | 2024-01-17 | 2024-02-20 | 航天宏图信息技术股份有限公司 | Surface parameter estimation method, device and equipment based on generation of countermeasure network |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20190138830A1 (en)* | 2015-01-09 | 2019-05-09 | Irvine Sensors Corp. | Methods and Devices for Cognitive-based Image Data Analytics in Real Time Comprising Convolutional Neural Network |
| CN112465880A (en)* | 2020-11-26 | 2021-03-09 | 西安电子科技大学 | Target detection method based on multi-source heterogeneous data cognitive fusion |
| CN113378763A (en)* | 2021-06-25 | 2021-09-10 | 长沙理工大学 | SAR image-oriented target automatic detection method and system |
| CN113469088A (en)* | 2021-07-08 | 2021-10-01 | 西安电子科技大学 | SAR image ship target detection method and system in passive interference scene |
- 2022
- 2022-11-18CNCN202211449058.5Apatent/CN115909086B/enactiveActive
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20190138830A1 (en)* | 2015-01-09 | 2019-05-09 | Irvine Sensors Corp. | Methods and Devices for Cognitive-based Image Data Analytics in Real Time Comprising Convolutional Neural Network |
| CN112465880A (en)* | 2020-11-26 | 2021-03-09 | 西安电子科技大学 | Target detection method based on multi-source heterogeneous data cognitive fusion |
| CN113378763A (en)* | 2021-06-25 | 2021-09-10 | 长沙理工大学 | SAR image-oriented target automatic detection method and system |
| CN113469088A (en)* | 2021-07-08 | 2021-10-01 | 西安电子科技大学 | SAR image ship target detection method and system in passive interference scene |
Non-Patent Citations (1)
| Title |
|---|
| 刘洁瑜;赵彤;刘敏;: "基于RetinaNet的SAR图像舰船目标检测", 湖南大学学报(自然科学版), no. 02, 25 February 2020 (2020-02-25)* |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116664823A (en)* | 2023-05-31 | 2023-08-29 | 西安电子科技大学 | Small-sample SAR target detection and recognition method based on meta-learning and metric learning |
| CN116664823B (en)* | 2023-05-31 | 2025-07-29 | 西安电子科技大学 | Small sample SAR target detection and recognition method based on meta learning and metric learning |
| CN116977870A (en)* | 2023-06-28 | 2023-10-31 | 西安电子科技大学 | SAR target increment identification method based on dynamic structure and multi-level distillation |
| CN116977870B (en)* | 2023-06-28 | 2025-07-29 | 西安电子科技大学 | SAR target increment identification method based on dynamic structure and multi-level distillation |
| CN117574161A (en)* | 2024-01-17 | 2024-02-20 | 航天宏图信息技术股份有限公司 | Surface parameter estimation method, device and equipment based on generation of countermeasure network |
| CN117574161B (en)* | 2024-01-17 | 2024-04-16 | 航天宏图信息技术股份有限公司 | Surface parameter estimation method, device and equipment based on generation of countermeasure network |
Also Published As
| Publication number | Publication date |
|---|---|
| CN115909086B (en) | 2025-08-05 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110796168B (en) | Vehicle detection method based on improved YOLOv3 | |
| CN104732243B (en) | SAR target identification methods based on CNN | |
| CN103955702B (en) | SAR image terrain classification method based on depth RBF network | |
| CN107341488B (en) | An integrated method for target detection and recognition in SAR images | |
| CN107563411B (en) | Online SAR target detection method based on deep learning | |
| CN112446388A (en) | Multi-category vegetable seedling identification method and system based on lightweight two-stage detection model | |
| CN112529005B (en) | Target detection method based on semantic feature consistency supervision pyramid network | |
| CN115909086A (en) | SAR target detection and recognition method based on multi-level enhanced network | |
| CN110147807A (en) | A kind of ship intelligent recognition tracking | |
| CN108009526A (en) | A kind of vehicle identification and detection method based on convolutional neural networks | |
| CN115223017B (en) | Multi-scale feature fusion bridge detection method based on depth separable convolution | |
| Cepni et al. | Vehicle detection using different deep learning algorithms from image sequence | |
| CN109034224A (en) | Hyperspectral classification method based on double branching networks | |
| CN110378191B (en) | Pedestrian and vehicle classification method based on millimeter wave sensor | |
| CN111738114B (en) | Vehicle target detection method based on accurate sampling of remote sensing images without anchor points | |
| CN106408030A (en) | SAR image classification method based on middle lamella semantic attribute and convolution neural network | |
| CN108898065A (en) | Candidate regions quickly screen and the depth network Ship Target Detection method of dimension self-adaption | |
| CN111950488A (en) | An Improved Faster-RCNN Remote Sensing Image Object Detection Method | |
| Yang et al. | HCNN-PSI: A hybrid CNN with partial semantic information for space target recognition | |
| CN115082672A (en) | Infrared image target detection method based on bounding box regression | |
| CN113808166A (en) | Single-target tracking method based on clustering difference and depth twin convolutional neural network | |
| CN109446894A (en) | The multispectral image change detecting method clustered based on probabilistic segmentation and Gaussian Mixture | |
| CN108447057A (en) | SAR image change detection based on conspicuousness and depth convolutional network | |
| CN110969121A (en) | High-resolution radar target recognition algorithm based on deep learning | |
| CN114973019B (en) | A method and system for detecting and classifying geospatial information changes based on deep learning |
Legal Events
| 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 |