技术领域Technical Field
本发明涉及电子领域，尤其涉及一种基于提示引导的稀疏遥感卫星图像去雾方法。The invention relates to the field of electronics, and in particular to a sparse remote sensing satellite image defogging method based on prompt guidance.
背景技术Background Art
随着遥感图像可以提供广泛、全方位的地球表面信息，为环境研究、资源管理、灾害监测等领域提供重要的数据支持。然而，远距离拍摄的遥感图像经常受到雾霾的影响，导致明显的模糊区域。这些退化的图像严重降低了视觉清晰度，增加了实现上述任务的难度。因此，遥感去雾任务是遥感图像复原领域中一项关键且具有挑战性的研究。As remote sensing images can provide extensive and comprehensive information about the earth's surface, they provide important data support for environmental research, resource management, disaster monitoring and other fields. However, remote sensing images taken at a long distance are often affected by haze, resulting in obvious blurred areas. These degraded images severely reduce the visual clarity and increase the difficulty of achieving the above tasks. Therefore, the remote sensing dehazing task is a key and challenging research in the field of remote sensing image restoration.
之前的去雾方法主要利用基于先验的方法来实现遥感图像的雾霾去除，例如暗通道先验(dark channel prior,DCP)、雾优化变换(haze optimized transformation,HOT)和Retinex先验(Retinex prior)。虽然这些方法可以提高图像质量，但手工方法无法有效模拟雾霾的分布，导致复杂的遥感场景中出现不良干扰。Previous dehazing methods mainly use prior-based methods to remove haze from remote sensing images, such as dark channel prior (DCP), haze optimized transformation (HOT) and Retinex prior. Although these methods can improve image quality, manual methods cannot effectively simulate the distribution of haze, resulting in undesirable interference in complex remote sensing scenes.
近年来，数据驱动的方法在去雾任务中表现出了出色的性能。巨大的卷积神经网络(Convolutional Neural Networks,CNN)由于其强大的建模能力而被广泛应用于遥感图像的去雾。DehazeNet采用非线性回归方法从雾霾图像估计大气光和透射图。FFA-Net将注意力与卷积相结合，进一步从模糊图像中提取关键特征。H2RL-Net对图像进行多尺度建模以恢复无雾图像。FCFT使用双重注意力网络从粗到细细化捕获的特征，提高除雾效果。然而，遥感图像中的雾霾分布往往是大尺度且连续的，不同区域之间存在着密切的关系。由于其固有的局限性，卷积算子无法探索遥感图像中全局信息的潜在相关性。因此，全局雾霾消除对CNN来说是一个巨大的挑战。为了解决这些问题，大多数现有模型都采用基于Transformer的方法来弥补CNN在全局特征提取方面的缺陷。Dehazeformer率先使用Transformer进行图像去雾，实现了显着的性能改进。Uformer采用U-Net架构逐步捕获雾霾图像的关键特征，进一步提高捕获空间上下文信息的能力。RSDformer将自注意力与卷积块相结合，以增强全局和局部信息之间的交互。In recent years, data-driven methods have shown excellent performance in dehazing tasks. Huge convolutional neural networks (CNNs) have been widely used in dehazing remote sensing images due to their powerful modeling capabilities. DehazeNet uses nonlinear regression to estimate atmospheric light and transmittance maps from hazy images. FFA-Net combines attention with convolution to further extract key features from hazy images. H2RL-Net performs multi-scale modeling on images to restore haze-free images. FCFT uses a dual attention network to refine the captured features from coarse to fine to improve the dehazing effect. However, the distribution of haze in remote sensing images is often large-scale and continuous, and there is a close relationship between different regions. Due to its inherent limitations, the convolution operator cannot explore the potential correlation of global information in remote sensing images. Therefore, global haze removal is a huge challenge for CNN. To address these problems, most existing models use Transformer-based methods to make up for the shortcomings of CNN in global feature extraction. Dehazeformer is the first to use Transformer for image dehazing, achieving significant performance improvements. Uformer uses the U-Net architecture to gradually capture the key features of haze images and further improves the ability to capture spatial context information. RSDformer combines self-attention with convolutional blocks to enhance the interaction between global and local information.
然而，由于自注意力的密集计算模式，图像中的所有元素都会无差别地参与计算。这会导致大量不必要的资源占用和额外的噪声干扰。因此，有必要探索最有用的自注意力值，以便可以充分利用这些特征来更好地恢复图像。However, due to the intensive computational mode of self-attention, all elements in the image will participate in the computation indiscriminately. This will lead to a lot of unnecessary resource occupation and additional noise interference. Therefore, it is necessary to explore the most useful self-attention value so that these features can be fully utilized to better restore the image.
发明内容Summary of the invention
本发明的目的在于针对上述现有技术的不足，提供了一种基于提示引导的稀疏遥感卫星图像去雾方法，自适应处理最相关的全局特征，并有效地恢复遥感卫星图像中被雾霾遮挡的区域。The purpose of the present invention is to address the deficiencies of the above-mentioned prior art and provide a sparse remote sensing satellite image defogging method based on prompt guidance, adaptively process the most relevant global features, and effectively restore the areas obscured by haze in remote sensing satellite images.
为实现上述目的，本发明采用了如下技术方案：To achieve the above object, the present invention adopts the following technical solutions:
本发明提供了一种基于提示引导的稀疏遥感卫星图像去雾方法，包括，建立用于遥感卫星图像去雾的基提示引导的稀疏Transformer算法；The present invention provides a sparse remote sensing satellite image defogging method based on prompt guidance, including establishing a sparse Transformer algorithm based on prompt guidance for remote sensing satellite image defogging;
假设给定输入图像和输出图像分别为Iin和Iout，基于提示引导的稀疏Transformer遥感卫星图像去雾方法表示为：Assuming that the input image and output image are Iin and Iout respectively, the prompt-guided sparse Transformer remote sensing satellite image dehazing method is expressed as:
Fi+1＝ATGB(Fi),i＝0,1,4,5；Fi+1 =ATGB(Fi ),i=0,1,4,5;
Fi+1＝Concat(ATGB(Fi),F4-i),i＝2,3；Fi+1 =Concat(ATGB(Fi ),F4-i ),i=2,3;
其中，为卷积核大小为3×3的卷积算子；ATGB(·)为自适应Top-k引导块；F0为通过3×3卷积获取的浅层特征；Concat(·)为通道合并操作；in, is a convolution operator with a convolution kernel size of 3×3; ATGB(·) is an adaptive Top-k guide block; F0 is a shallow feature obtained by 3×3 convolution; Concat(·) is a channel merging operation;
自适应Top-k引导块由自适应Top-k知道注意力和频率选择前馈网络组成，表示为：The adaptive Top-k guidance block consists of an adaptive Top-k aware attention and frequency selective feed-forward network, expressed as:
ATGB(X)＝FSFN(LN(ATGA(LN(X))+X))+(ATGA(LN(X))+X)；ATGB(X)=FSFN(LN(ATGA(LN(X))+X))+(ATGA(LN(X))+X);
其中，LN(·)为层归一化操作；ATGA(·)为自适应Top-k指导注意力；FSFN(·)为频率选择前馈网络；X为输入特征。Among them, LN(·) is the layer normalization operation; ATGA(·) is the adaptive Top-k guided attention; FSFN(·) is the frequency selective feedforward network; X is the input feature.
进一步，还包括自适应Top-k指导注意力计算过程，Furthermore, it also includes adaptive Top-k guided attention calculation process,
假设给定输入特征为X∈RH×W×C，使用投影矩阵生成执行注意力计算过程的所需的查询Q、键K和值V，为：Assuming that the input feature is X∈RH×W×C , the projection matrix is used to generate the required query Q, key K and value V to perform the attention calculation process, as follows:
其中，为生成查询Q、键K和值V所需要的投影矩阵；Chunk为通道拆分操作；in, To generate the projection matrix required for query Q, key K and value V; Chunk is the channel splitting operation;
为了实现特征形状对齐，对特征执行重塑操作，为：To achieve feature shape alignment, a reshape operation is performed on the features as follows:
其中，为特征重塑操作。in, Reshape operation for features.
进一步，对查询Q、键K采用矩阵乘法操作实现注意力计算：Furthermore, the attention calculation is realized by using matrix multiplication operation on query Q and key K:
其中，sk为Top-k选择算子，用于实现稀疏注意力，提高关键特征的交互和聚合能力；γ为学习的参数，用于控制点积数值的大小；SK为注意力矩阵；j为sk中的列数；Among them,sk is the Top-k selection operator, which is used to realize sparse attention and improve the interaction and aggregation ability of key features; γ is the learning parameter, which is used to control the size of the dot product value;SK is the attention matrix; j is the number of columns insk ;
假设密集度为键K的注意力矩阵，表示为SK＝[W1,W2…,WC]T；其中，第i个稀疏表示为Assume that the density is the attention matrix of key K, represented by SK = [W1 ,W2 …,WC ]T ; where the i-th sparse representation is
其中，wi为第i个注意力值。Among them,wi is the i-th attention value.
进一步，第i个注意力头的平均值表示为：Furthermore, the average value of the i-th attention head is expressed as:
将每个头中的所有输出连接起来，并用1×1卷积聚合：Concatenate all outputs from each head and aggregate them with a 1×1 convolution:
其中，为in, for
最终输出。Final output.
进一步，LPB的提示过程为：Furthermore, the prompting process of LPB is:
其中，Xi为所有自注意力特征；ε(·)为全局平均池化；为代表逐元素乘法；σ为sigmoid算子；k1,k2,k3为原始权重分布映射出三个动态参数；FC为全连接层；XL为一组学习参数；为3x3深度卷积核全局平均池化；Among them,Xi is all self-attention features; ε(·) is the global average pooling; represents element-by-element multiplication; σ is the sigmoid operator; k1 , k2 , k3 are three dynamic parameters mapped from the original weight distribution; FC is the fully connected layer;XL is a set of learning parameters; It is a 3x3 depth convolution kernel global average pooling;
TOS通过掩码操作选择元素的细化过程为：The refinement process of TOS selecting elements through mask operation is:
其中，M为掩码操作；Xo为稀疏化后的输出。Among them, M is the mask operation;Xo is the output after sparseness.
进一步，FSFN的计算过程表示为：Furthermore, the calculation process of FSFN is expressed as:
其中，XF和X'F分别为FSFN的输入和出；ε(·)为全局平均池化操作；为最大池化操作。Among them, XF andX'F are the input and output of FSFN respectively; ε(·) is the global average pooling operation; It is the maximum pooling operation.
本发明的有益效果为：有效地学习遥感雾霾图像中的大尺度和非均匀雾霾特征，从而在获得良好的去雾效果同时更好的保留图像原本的纹理细节。The beneficial effect of the present invention is to effectively learn the large-scale and non-uniform haze characteristics in remote sensing haze images, thereby obtaining a good defogging effect while better retaining the original texture details of the image.
开发了一个自适应Top-k指导注意力，包括Top-k选择算子和可学习提示块。ATGA利用TSO自适应增强自注意力计算的稀疏性。经过精心设计的LPB，以确保选择过程中的精确指导；An adaptive Top-k guided attention is developed, including a Top-k selection operator and a learnable hint block. ATGA uses TSO to adaptively enhance the sparsity of self-attention computation. A carefully designed LPB is used to ensure precise guidance during the selection process;
提出了频率选择前馈网络将ATGA的注意特征映射到频域空间，以进一步促进特征变换中雾霾和干净图像的分离；在不同的遥感图像去雾基准数据集上证明了本发明的合理性，并且我们的方法相对于其他竞争方法取得了良好的性能。A frequency selective feedforward network is proposed to map the attention features of ATGA to the frequency domain space to further promote the separation of haze and clean images in feature transformation; the rationality of the invention is demonstrated on different remote sensing image dehazing benchmark datasets, and our method achieves good performance compared with other competing methods.
附图说明BRIEF DESCRIPTION OF THE DRAWINGS
图1为基于提示引导的稀疏遥感卫星图像去雾方法的结构图；FIG1 is a structural diagram of a sparse remote sensing satellite image defogging method based on prompt guidance;
图2是自适应top-k指导注意力的结构图；Figure 2 is a diagram of the structure of adaptive top-k guided attention;
图3是频率选择前馈网络结构图；Fig. 3 is a diagram of the frequency selection feedforward network structure;
图4是本发明与其它方法在SateHaze1k数据集上可视化定性对比结果；FIG4 is a visualized qualitative comparison result of the present invention and other methods on the SateHaze1k dataset;
图5是本发明与其他方法在RCIE数据集上可视化定性对比结果；FIG5 is a visualization qualitative comparison result of the present invention and other methods on the RCIE dataset;
图6是本发明与其他方法在RRSD300数据集上可视化定性对比结果。FIG6 is a visualized qualitative comparison result of the present invention and other methods on the RRSD300 dataset.
具体实施方式DETAILED DESCRIPTION
为了使本发明的目的、技术方案及优点更加清楚明白，下面结合附图，对本发明进行进一步详细说明。应当理解，此处所描述的具体实施例仅用以解释本发明，并不用于限定本发明。In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.
本发明提出的基于提示引导的稀疏遥感卫星图像去雾方法，如图1所示，整体网络架构采用自适应top-k引导块作为基本特征提取单元，自适应top-k引导块自适应top-k知道注意力和频率选择前馈网络组成。The sparse remote sensing satellite image defogging method based on prompt guidance proposed in the present invention is shown in Figure 1. The overall network architecture adopts an adaptive top-k guided block as a basic feature extraction unit, and the adaptive top-k guided block is composed of an adaptive top-k aware attention and frequency selection feedforward network.
一种基于提示引导的稀疏遥感卫星图像去雾方法，包括，建立用于遥感卫星图像去雾的基提示引导的稀疏Transformer算法；A cue-guided sparse remote sensing satellite image defogging method, comprising: establishing a cue-guided sparse Transformer algorithm for remote sensing satellite image defogging;
假设给定输入图像和输出图像分别为Iin和Iout，基于提示引导的稀疏Transformer遥感卫星图像去雾方法表示为：Assuming that the input image and output image are Iin and Iout respectively, the prompt-guided sparse Transformer remote sensing satellite image dehazing method is expressed as:
Fi+1＝ATGB(Fi),i＝0,1,4,5；Fi+1 =ATGB(Fi ),i=0,1,4,5;
Fi+1＝Concat(ATGB(Fi),F4-i),i＝2,3；Fi+1 =Concat(ATGB(Fi ),F4-i ),i=2,3;
其中，为卷积核大小为3×3的卷积算子；ATGB(·)为自适应Top-k引导块；F0为通过3×3卷积获取的浅层特征；Concat(·)为通道合并操作；in, is a convolution operator with a convolution kernel size of 3×3; ATGB(·) is an adaptive Top-k guide block; F0 is a shallow feature obtained by 3×3 convolution; Concat(·) is a channel merging operation;
自适应Top-k引导块由自适应Top-k知道注意力和频率选择前馈网络组成，表示为：The adaptive Top-k guidance block consists of an adaptive Top-k aware attention and frequency selective feed-forward network, expressed as:
ATGB(X)＝FSFN(LN(ATGA(LN(X))+X))+(ATGA(LN(X))+X)；ATGB(X)=FSFN(LN(ATGA(LN(X))+X))+(ATGA(LN(X))+X);
其中，LN(·)为层归一化操作；ATGA(·)为自适应Top-k指导注意力；FSFN(·)为频率选择前馈网络；X为输入特征。Among them, LN(·) is the layer normalization operation; ATGA(·) is the adaptive Top-k guided attention; FSFN(·) is the frequency selective feedforward network; X is the input feature.
还包括自适应Top-k指导注意力计算过程，It also includes adaptive Top-k guided attention calculation process,
如图2所示，假设给定输入特征为X∈RH×W×C，使用投影矩阵生成执行注意力计算过程的所需的查询Q、键K和值V，为：As shown in Figure 2, assuming that the given input feature is X∈RH×W×C , the projection matrix is used to generate the required query Q, key K, and value V to perform the attention calculation process, as follows:
其中，为生成查询Q、键K和值V所需要的投影矩阵；Chunk为通道拆分操作；in, To generate the projection matrix required for query Q, key K and value V; Chunk is the channel splitting operation;
为了实现特征形状对齐，对特征执行重塑操作，为：To achieve feature shape alignment, a reshape operation is performed on the features as follows:
其中，为特征重塑操作。in, Reshape operation for features.
对查询Q、键K采用矩阵乘法操作实现注意力计算：The attention calculation is implemented by using matrix multiplication operation for query Q and key K:
其中，sk为Top-k选择算子，用于实现稀疏注意力，提高关键特征的交互和聚合能力；γ为学习的参数，用于控制点积数值的大小；SK为注意力矩阵；j为sk中的列数；Among them,sk is the Top-k selection operator, which is used to realize sparse attention and improve the interaction and aggregation ability of key features; γ is the learning parameter, which is used to control the size of the dot product value;SK is the attention matrix; j is the number of columns insk ;
假设密集度为键K的注意力矩阵，表示为SK＝[W1,W2…,WC]T；其中，第i个稀疏表示为Assume that the density is the attention matrix of key K, represented by SK = [W1 ,W2 …,WC ]T ; where the i-th sparse representation is
其中，wi为第i个注意力值。Among them,wi is the i-th attention value.
第i个注意力头的平均值表示为：The average value of the i-th attention head is expressed as:
将每个头中的所有输出连接起来，并用1×1卷积聚合：Concatenate all outputs from each head and aggregate them with a 1×1 convolution:
其中，为in, for
最终输出。Final output.
由于TSO中可调参数k值的选择存在争议，因为完全依赖于手动设置，这可能会在识别焦点注意力特征时产生偏差，最终导致无关细节的不准确分配，为了更准确地设置可调参数k，通过可学习提示块来提供更准确的参数提示指导；Since the choice of the adjustable parameter k value in TSO is controversial, it relies entirely on manual settings, which may cause bias in identifying focal attention features and ultimately lead to inaccurate assignment of irrelevant details. In order to set the adjustable parameter k more accurately, a learnable hint block is used to provide more accurate parameter hint guidance;
LPB首先引入3×3深度卷积和全局平均池化，以生产紧凑的提示向量w∈R1×C；然后，引入一组可学习参数XL；这些参数按元素与w相乘；LPB first introduces 3×3 depthwise convolution and global average pooling to produce a compact hint vector w∈R1×C ; then, a set of learnable parametersXL are introduced; these parameters are multiplied by w element-wise;
接下来，执行全连接层(Fully Connection Layer，FC)和sigmoid计算，根据原始权重分布映射出三个动态参数k1,k2,k3。然后使用它们作为TSO内稀疏度控制的参数；Next, perform the fully connected layer (FC) and sigmoid calculation to map out three dynamic parameters k1 , k2 , k3 according to the original weight distribution. Then use them as parameters for sparsity control in TSO;
LPB的提示过程为：The prompting process of LPB is:
其中，Xi为所有自注意力特征；ε(·)为全局平均池化；为代表逐元素乘法；σ为sigmoid算子；k1,k2,k3为原始权重分布映射出三个动态参数；FC为全连接层；XL为一组学习参数；为3x3深度卷积核全局平均池化；Among them,Xi is all self-attention features; ε(·) is the global average pooling; represents element-by-element multiplication; σ is the sigmoid operator; k1 , k2 , k3 are three dynamic parameters mapped from the original weight distribution; FC is the fully connected layer;XL is a set of learning parameters; It is a 3x3 depth convolution kernel global average pooling;
当矩阵中的权重分数通常较小时TSO，仍然保留了一些低相关度的特征。因此，在到LPB中利用掩码操作以进一步细化TSO筛选的元素。掩码阈值m由LPB从原始特征Xi自适应学习。当ATGA中引入掩码操作时，可以更有效地提取重要特征。When the weight scores in the matrix are usually small, TSO still retains some features with low correlation. Therefore, the mask operation is used in LPB to further refine the elements screened by TSO. The mask threshold m is adaptively learned from the original featuresXi by LPB. When the mask operation is introduced in ATGA, important features can be extracted more effectively.
TOS通过掩码操作选择元素的细化过程为：The refinement process of TOS selecting elements through mask operation is:
其中，M为掩码操作；Xo为稀疏化后的输出。Among them, M is the mask operation;Xo is the output after sparseness.
请参阅图3，假设FSFN的输入特征为XF，首先使用1×1卷积来融合来自不同频率的信息并扩展通道维度；随后，设计了双分支并行结构来合并全局平均池化和最大池化。通过上述结构，FSFN可以实现高通道和低通滤波器的功能，实现跨越各种频率的自适应特征交互；此外，引入由3×3深度卷积组成的残差连接起来进一步有效地对特征的频率信息进行建模，最后，使用1×1卷积来恢复通道维度。Please refer to Figure 3. Assuming that the input feature of FSFN is XF , 1×1 convolution is first used to fuse information from different frequencies and expand the channel dimension; then, a dual-branch parallel structure is designed to merge global average pooling and maximum pooling. Through the above structure, FSFN can realize the functions of high-pass and low-pass filters, and realize adaptive feature interaction across various frequencies; in addition, the residual connection composed of 3×3 deep convolution is introduced to further effectively model the frequency information of the feature. Finally, 1×1 convolution is used to restore the channel dimension.
FSFN的计算过程表示为：The calculation process of FSFN is expressed as:
其中，XF和X′F分别为FSFN的输入和出；ε(·)为全局平均池化操作；为最大池化操作。Where XF and X′F are the input and output of FSFN respectively; ε(·) is the global average pooling operation; It is the maximum pooling operation.
参见图3、图4和图5，为了证明本发明的有效性，将本发明与基于先验的方法(DCP)、基于CNN的去雾基线(DehazeNet、AODNet、PFFNet、FFA-Net、FCTF、MSBDN、LD-Net、SCANet)以及最近的基于Transformer的方法(Dehazeformer、UFormer、AIDNet和RSDformer)进行对比。具体说明如下：Referring to Figures 3, 4 and 5, in order to demonstrate the effectiveness of the present invention, the present invention is compared with the prior-based method (DCP), the CNN-based dehazing baselines (DehazeNet, AODNet, PFFNet, FFA-Net, FCTF, MSBDN, LD-Net, SCANet) and the recent Transformer-based methods (Dehazeformer, UFormer, AIDNet and RSDformer). The specific description is as follows:
图4显示了本发明和其他方法在SateHaze1k数据集上的一些遥感去雾结果。可以看出，PFFNet和SCANet无法有效恢复被雾霾遮挡的区域，导致大量信息丢失。UFormer和RSDformer对建筑色彩的还原不够，导致局部色彩失真。相比之下，PGSformer可以更清晰地保留图像颜色并保留准确的纹理结构。图4显示本发明可以更清晰地还原地面和山脉的纹理细节，并且更接近地面真实情况。Figure 4 shows some remote sensing dehazing results of the present invention and other methods on the SateHaze1k dataset. It can be seen that PFFNet and SCANet cannot effectively restore the areas obscured by haze, resulting in a large amount of information loss. UFormer and RSDformer do not restore the color of the building enough, resulting in local color distortion. In contrast, PGSformer can preserve the image color more clearly and retain the accurate texture structure. Figure 4 shows that the present invention can restore the texture details of the ground and mountains more clearly and is closer to the ground truth.
为了进一步评估定性性能，我们使用RRSD300真实基准数据集进行了额外的实验。视觉结果如图5和图6所示。结果表明，大多数模型难以有效处理大范围且不均匀分布的现实世界雾霾，导致其输出中出现明显的雾霾效应。相比之下，与其他比较模型相比，我们的模型取得了令人印象深刻的遥感去雾结果。所提出的模型可以有效消除大部分雾霾干扰，从而产生视觉上令人愉悦的恢复效果。这表明在真实的遥感除雾场景中，我们的网络表现出卓越的输出质量，内容更清晰，感知质量增强。To further evaluate the qualitative performance, we conduct additional experiments using the RRSD300 real benchmark dataset. The visual results are shown in Figures 5 and 6. The results show that most models have difficulty in effectively handling large-scale and unevenly distributed real-world haze, resulting in obvious haze effects in their output. In contrast, our model achieves impressive remote sensing dehazing results compared to other comparison models. The proposed model can effectively remove most of the haze interference, resulting in visually pleasing restoration results. This shows that in real remote sensing dehazing scenes, our network exhibits superior output quality, with clearer content and enhanced perceptual quality.
综上所述，本发明提出了一种基于提示引导的稀疏Transformer遥感卫星图像去雾方法，其采用自适应Top-k引导注意力利用Top-k选择算子来保留每个查询的键中最重要的注意力分数，防止自注意力计算中低相关性查询键对的干扰。In summary, the present invention proposes a sparse Transformer remote sensing satellite image dehazing method based on prompt guidance, which adopts adaptive Top-k guided attention to utilize the Top-k selection operator to retain the most important attention scores in the key of each query, and prevent the interference of low-correlation query key pairs in the self-attention calculation.
同时，我们在ATGA中设计了可学习提示块，以进一步提高稀疏选择以增强注意力的准确性。其中LPB指导TSO动态优化稀疏率并自适应学习掩码阈值以进一步提取所选特征。此外，还设计了频率选择前馈网络来自适应获取频率信息，使整体管道能够提高双频特征的学习能力。大量的实验表明，与其他竞争方法相比，所提出的方法获得了更有吸引力的性能。Meanwhile, we design a learnable hint block in ATGA to further improve the accuracy of sparse selection for enhanced attention. Among them, LPB guides TSO to dynamically optimize the sparse rate and adaptively learn the mask threshold to further extract the selected features. In addition, a frequency-selective feedforward network is designed to adaptively obtain frequency information, enabling the overall pipeline to improve the learning ability of dual-frequency features. Extensive experiments show that the proposed method obtains more attractive performance compared with other competing methods.
以上所述实施例仅表达了本发明的实施方式，其描述较为具体和详细，但并不能因此而理解为对本发明专利范围的限制。应当指出的是，对于本领域的普通技术人员来说，在不脱离本发明构思的前提下，还可以做出若干变形和改进，这些都属于本发明的保护范围。因此，本发明专利的保护范围应以所附权利要求。The above-mentioned embodiments only express the implementation methods of the present invention, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the patent of the present invention. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention should be based on the attached claims.
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202410879305.8ACN118887132A (en) | 2024-07-02 | 2024-07-02 | A cue-guided dehazing method for sparse remote sensing satellite images |
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202410879305.8ACN118887132A (en) | 2024-07-02 | 2024-07-02 | A cue-guided dehazing method for sparse remote sensing satellite images |
| Publication Number | Publication Date |
|---|---|
| CN118887132Atrue CN118887132A (en) | 2024-11-01 |
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202410879305.8APendingCN118887132A (en) | 2024-07-02 | 2024-07-02 | A cue-guided dehazing method for sparse remote sensing satellite images |
| Country | Link |
|---|---|
| CN (1) | CN118887132A (en) |
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US6714925B1 (en)* | 1999-05-01 | 2004-03-30 | Barnhill Technologies, Llc | System for identifying patterns in biological data using a distributed network |
| US20230260279A1 (en)* | 2020-10-07 | 2023-08-17 | Wuhan University | Hyperspectral remote sensing image classification method based on self-attention context network |
| CN117809294A (en)* | 2023-12-29 | 2024-04-02 | 天津大学 | Text detection method based on feature correction and difference guiding attention |
| CN117953366A (en)* | 2023-11-30 | 2024-04-30 | 长沙理工大学 | Method and system for extracting water body from SAR image based on water body extraction network |
| CN118195962A (en)* | 2024-03-22 | 2024-06-14 | 南京理工大学 | Image flare removal system and method based on cross-channel feature information interaction |
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US6714925B1 (en)* | 1999-05-01 | 2004-03-30 | Barnhill Technologies, Llc | System for identifying patterns in biological data using a distributed network |
| US20230260279A1 (en)* | 2020-10-07 | 2023-08-17 | Wuhan University | Hyperspectral remote sensing image classification method based on self-attention context network |
| CN117953366A (en)* | 2023-11-30 | 2024-04-30 | 长沙理工大学 | Method and system for extracting water body from SAR image based on water body extraction network |
| CN117809294A (en)* | 2023-12-29 | 2024-04-02 | 天津大学 | Text detection method based on feature correction and difference guiding attention |
| CN118195962A (en)* | 2024-03-22 | 2024-06-14 | 南京理工大学 | Image flare removal system and method based on cross-channel feature information interaction |
| Title |
|---|
| S. KIESEL; UNIVERSITY OF STUTTGART; M. STIEMERLING; H-DA;: "Application-Layer Traffic Optimization (ALTO) Cross-Domain Server Discovery", IETF, 28 February 2020 (2020-02-28)* |
| 吴嘉炜;余兆钗;李佐勇;刘维娜;张祖昌;: "一种基于深度学习的两阶段图像去雾网络", 计算机应用与软件, no. 04, 12 April 2020 (2020-04-12)* |
| Publication | Publication Date | Title |
|---|---|---|
| CN109493303B (en) | An Image Dehazing Method Based on Generative Adversarial Networks | |
| Wang et al. | Multi-scale dilated convolution of convolutional neural network for image denoising | |
| CN111062880B (en) | Underwater image real-time enhancement method based on condition generation countermeasure network | |
| CN114742719B (en) | An end-to-end image dehazing method based on multi-feature fusion | |
| CN111160533A (en) | Neural network acceleration method based on cross-resolution knowledge distillation | |
| CN109272455A (en) | Image defogging method based on weak supervision to generate confrontation network | |
| CN111681188B (en) | Image Deblurring Method Based on Combining Image Pixel Prior and Image Gradient Prior | |
| CN108427921A (en) | A kind of face identification method based on convolutional neural networks | |
| CN110060286A (en) | A kind of monocular depth estimation method | |
| CN110189260B (en) | An Image Noise Reduction Method Based on Multi-scale Parallel Gated Neural Network | |
| CN115063318A (en) | Low-light image enhancement method and related equipment based on adaptive frequency decomposition | |
| CN114626042B (en) | A face verification attack method and device | |
| CN115899598B (en) | A heating network status monitoring method and system integrating auditory and visual features | |
| CN114897884A (en) | No-reference screen content image quality evaluation method based on multi-scale edge feature fusion | |
| WO2021042857A1 (en) | Processing method and processing apparatus for image segmentation model | |
| CN115619677A (en) | An Image Dehazing Method Based on Improved CycleGAN | |
| CN110807744A (en) | An Image Dehazing Method Based on Convolutional Neural Network | |
| CN110135501A (en) | High Dynamic Range Image Forensics Method Based on Neural Network Framework | |
| CN117455757A (en) | Image processing method, device, equipment and storage medium | |
| CN117611456A (en) | Atmospheric turbulence image restoration method and system based on multiscale generation countermeasure network | |
| CN118333900A (en) | An unpaired dehazing method based on cyclic generative adversarial network | |
| CN118115378A (en) | Low-light image enhancement method of image hierarchical structure network based on stream learning | |
| CN116309178A (en) | A Visible Light Image Denoising Method Based on Adaptive Attention Mechanism Network | |
| Bai et al. | CEPDNet: a fast CNN-based image denoising network using edge computing platform | |
| CN109523478B (en) | Image descreening method and storage medium |
| 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 |