A target monitoring method, medium and system for high-resolution satellite remote sensing images

Abstract

The present invention provides a target monitoring method, medium and system for high-resolution satellite remote sensing images, which belongs to the field of remote sensing data processing technology, and includes: first, preprocessing the image and dividing it into multiple sub-regions by using a region segmentation algorithm. Then, a global-region enhancement network is used to iteratively enhance the sub-regions, and a game optimization model is established to balance local features and global features to generate candidate refined images. Next, a deep target detection network is used to perform a preliminary scan on the preprocessed image to obtain a target candidate region. Subsequently, a multi-layer attention mechanism is used to extract target features, and an occlusion distribution matrix is ​​solved based on an occlusion perception equation group to adaptively compensate for the target region with occlusion. Finally, the extracted target features are input into a pre-trained recognition model to realize target category recognition and position positioning. The problem that the existing technology for enhancing remote sensing images is difficult to cope with different types of targets in complex scenes is solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of remote sensing data processing, and in particular, relates to a target monitoring method, medium and system for high-resolution satellite remote sensing images. Background Art

[0002] With the continuous improvement of satellite sensor performance, the resolution of satellite remote sensing images currently available has reached 0.5 meters, providing a rich source of information for various applications. However, in complex natural and artificial environments, high-resolution remote sensing images are easily affected by factors such as illumination changes, atmospheric interference, and ground obstruction, resulting in problems such as target blur and loss of details, which greatly limits the actual application value of the images. In the process of realizing target monitoring of high-resolution satellite remote sensing images, the existing technology first pre-processes the remote sensing images including image enhancement, then uses target detection and recognition technology on the pre-processed images, and uses deep learning methods such as convolutional neural networks to accurately detect and identify the target of interest from the pre-processed images. These methods have improved the application effect of high-resolution remote sensing images to a certain extent, but because the existing technology only considers the single aspect of local features or global features in the enhancement processing of remote sensing images, it fails to make full use of multi-scale feature information, resulting in unsatisfactory enhancement effects. The main reason is that most traditional image enhancement methods use classic algorithms based on histogram equalization, wavelet transform, etc. Although these methods can improve image quality to a certain extent, since these algorithms use fixed parameters and processing procedures, they cannot be adaptively adjusted for different areas and different target types, and it is difficult to cope with different types of targets in complex scenes, resulting in the enhancement effect being either excessive or insufficient, and difficult to balance. In other words, the enhancement effect is often not adaptive enough.

[0003] Existing deep learning target detection methods are mainly designed for natural scene images, without fully considering the characteristics of remote sensing images. For example, the large scale differences of targets in remote sensing images, special viewing angles, and complex backgrounds make it difficult for detection accuracy and robustness to meet actual needs.

[0004] In summary, the application of high-resolution remote sensing image target recognition in complex environments faces many challenges, and it is necessary to study a new solution to overcome the limitations of existing technologies and improve the application effect of remote sensing images. Summary of the invention

[0005] In view of this, the present invention provides a target monitoring method, medium and system for high-resolution satellite remote sensing images, which can solve the problem that the existing technology is difficult to cope with different types of targets in complex scenes for high-resolution satellite remote sensing image enhancement processing.

[0006] The present invention is achieved in that:

[0007] A first aspect of the present invention provides a method for monitoring a target using high-resolution satellite remote sensing images, comprising the following steps:

[0008] S10, preprocessing the high-resolution satellite remote sensing image to generate an initial preprocessed image, and partitioning the initial preprocessed image based on the region segmentation algorithm to obtain multiple sub-region images; specifically: first, performing format conversion, radiation correction, geometric correction and other preprocessing operations on the original high-resolution satellite remote sensing image to eliminate the interference introduced by the sensing equipment, atmosphere, terrain and other factors, and obtain a clear and stable initial preprocessed image. Then, the image gradient information is calculated based on the improved multi-scale morphological gradient operator ,in is the input image. Then, the adaptive threshold segmentation method is used to determine the candidate area. ,in is the regional grayscale mean, is the standard deviation, is the adjustment coefficient, and its value is generally between 1.0 and 1.5. Finally, the improved region growing algorithm is used to finely segment the candidate region, and the growth criterion is ,in and are the mean and standard deviation of the seed region respectively. Through this step, the original high-resolution satellite remote sensing image is divided into multiple sub-region images.

[0009] S20. Use the pre-trained global-region enhancement network to iteratively enhance each sub-region image to obtain a sub-region enhanced temporary image, and splice all the sub-region enhanced temporary images obtained in each iteration into a global temporary image; specifically: first, the region generation sub-network adopts a multi-branch residual network structure, the input is the original sub-region image, and the output is the enhanced sub-region feature map. The network structure includes multiple residual blocks, each residual block consists of two 3x3 convolution layers, a BN layer, and a ReLU activation layer, and has a short-circuit connection. Secondly, the global splicing sub-network adopts an attention-enhanced encoder-decoder structure, the input is the spliced ​​global temporary image, and the output is the enhanced global feature map. The network structure includes an encoder and a decoder. The encoder consists of multiple convolution + pooling layers, and the decoder consists of multiple deconvolution + upsampling layers. The attention mechanism is used to enhance the feature expression ability. Finally, through iterative optimization, the parameters of the region generation sub-network and the global splicing sub-network are updated alternately to generate enhanced sub-region temporary images and global temporary images. The construction process of the training dataset of the global-regional enhancement network first requires the collection of a large number of high-resolution satellite remote sensing images. These images should cover different scene types, climate conditions and lighting conditions to ensure that the network has good generalization ability. The collected data needs to be uniformly preprocessed, including the uniform processing of image size, radiation correction and geometric correction. In order to expand the scale of the dataset and improve the robustness of the model, the original image needs to be enhanced, including filtering, denoising, improving clarity, and deformity correction. Then, the superpixel segmentation algorithm is used to divide the preprocessed training image into regions, generate sub-region samples, perform image enhancement on the sub-region samples, and save the pairing relationship between the sub-region and the corresponding global image. Finally, the processed dataset is divided into a training set and a validation set in a ratio of 8:2.

[0010] The training process of the network is divided into two stages: pre-training and joint training. In the pre-training stage, the regional generation sub-network and the global splicing sub-network are first trained independently, using mean square error loss and perceptual loss as supervision signals, using Adam optimizer for parameter optimization, and the initial learning rate is set to 0.0001. In the joint training stage, an objective function based on game optimization is constructed, and the balance between local features and global features is achieved by alternately optimizing the parameters of the two sub-networks. The design of the loss function includes four parts: L1 loss for ensuring reconstruction quality, perceptual loss based on VGG features, SSIM loss for measuring global consistency, and regional feature loss for measuring local structural similarity. The training adopts a progressive strategy, the batch size is set to 16, a total of 200 iterations are performed, and learning rate decay and early stopping mechanisms are used to prevent overfitting.

[0011] S30. Using the network parameter complexity, feature space consistency and information loss degree of the regional generation subnetwork and the global stitching subnetwork as adjustment factors, and the enhancement degree of the subregion enhanced temporary image and the enhancement degree of the global temporary image as the two parties in the game, a game optimization model is established to achieve a balance between local features and global features.

[0012] S40, based on the optimization result of the game optimization model, update the global-regional enhancement network, use the updated global-regional enhancement network to iteratively optimize the sub-regional image and the overall image, and generate multiple candidate refined images; select the best refined image among the candidate image clarity, information entropy, and edge preservation as the final preprocessed image; specifically: first, according to the game optimization result obtained in step S30, update the parameters of the global-regional enhancement network, and the parameters of the regional generation sub-network are obtained by Update, the parameters of the global splicing subnetwork are given by Then, the updated global-regional enhancement network is used to iteratively optimize each sub-region image and the overall image to generate multiple candidate refined images. Finally, the generated multiple candidate refined images are evaluated, and the indicators include clarity (using the Laplace operator to calculate the image gradient energy), information entropy (quantifying the image information content), and edge preservation (using the Sobel operator to calculate the edge response intensity). The refined image with the best combination of the three is selected as the final preprocessing result.

[0013] S50: According to preset target features, a deep target detection network is used to perform a preliminary scan on the final preprocessed image to generate a plurality of target candidate regions.

[0014] S60. Use a multi-layer attention mechanism to extract features from each target candidate region and construct a spatial attention matrix and a channel attention matrix.

[0015] S70, solving the preset occlusion perception equations based on the spatial attention matrix and the channel attention matrix to obtain an occlusion distribution matrix; specifically: first, define the occlusion perception equations, including the visibility equation, the occlusion degree equation, and the occlusion position equation. Then, using the spatial attention matrix and the channel attention matrix calculated in step S60, solve the above occlusion perception equations to obtain an occlusion distribution matrix ,in ,Each element in the matrix represents the occlusion intensity of the corresponding position.

[0016] S80. If there are elements in the occlusion distribution matrix that are greater than a preset threshold, it means that the target is occluded. Then, based on the occlusion distribution matrix, adaptive occlusion compensation is performed on the target area, and a multi-scale feature fusion method is used to extract target features, a target feature matrix is ​​constructed, and the process proceeds to step S90; otherwise, a multi-scale feature fusion method is directly used to extract target features, a target feature matrix is ​​constructed, and the process proceeds to step S90.

[0017] S90, input the target feature matrix into a pre-trained target recognition model to achieve target category recognition and position positioning, and output target monitoring results, specifically:

[0018] On the basis of the above technical solution, the target monitoring method of a high-resolution satellite remote sensing image of the present invention can also be improved as follows: First, construct a target feature matrix ,in Indicates The feature vector of the target, including the position coordinates , Width and Height , category probability and confidence Then, an improved fast regional convolutional neural network is used as the target recognition model. The model consists of a feature extraction backbone network (using a residual network), a region proposal network (using deformable convolution), a region of interest alignment layer, and a multi-task prediction head. During the training process, the region proposal network is trained first, and then the entire network is trained. The loss function includes classification loss (cross entropy), positioning loss (smoothing ) and occlusion prediction loss. Finally, the target feature matrix constructed in step S80 is input into the trained target recognition model to realize the category recognition and position positioning of the target, and output the final target monitoring result.

[0019] Among them, the global-regional enhancement network includes a region generation subnetwork and a global stitching subnetwork, the region generation subnetwork is used to perform feature enhancement on each of the subregion images, and the global stitching subnetwork performs feature enhancement on the global temporary image obtained in each iteration step.

[0020] Specifically, the region generation subnetwork is used to enhance the features of a single subregion. The input is the original subregion image, and the output is the enhanced subregion feature map. The structure adopts a multi-branch residual network. The global splicing subnetwork is used to enhance the overall features. The input is the spliced ​​global temporary image, and the output is the enhanced global feature map. The structure adopts an attention-enhanced encoder-decoder.

[0021] Furthermore, the objective function of the game optimization model is expressed as follows:

[0022] ;

[0023] In the formula, represents the region generation subnetwork, represents the global splicing subnetwork, represents mathematical expectation; Represents the real data distribution; represents the noise distribution; Representation variables obey Distribution expectations, Representation variables obey Distribution expectations; is the weight coefficient, and its value range is [0,1]; is the network parameter complexity; is the feature space consistency; is the degree of information loss; Consider generating a network Try to minimize the objective function, Considering the discriminant network Trying to maximize the objective function, the game adversarial approach ensures the dynamic balance of the two networks.

[0024] The purpose of taking logarithms for the two sub-networks is to convert multiplication into addition, which is convenient for gradient calculation. The logarithmic function can enhance numerical stability and prevent the value from being too small. At the same time, it conforms to the form of cross entropy loss, which is beneficial to network training. The discriminator believes is the probability of the true sample, Indicates the probability that the discriminator thinks the generated sample is fake; network parameter complexity The calculation of adopts the general network parameter complexity calculation method of neural network, and the preferred formula is expressed as:

[0025] ;

[0026] In the formula, is the number of network layers; For the The number of neurons in the layer; For the The input dimension of the layer; For the The weight matrix of the layer; For the The bias vector of the layer; is the balance coefficient, the default value is 0.5.

[0027] Feature space consistency The calculation of adopts the common feature consistency calculation method of neural network, and the preferred formula is:

[0028] ;

[0029] In the formula, is the sample size; and Respectively and Feature extraction function; is the input sample; Representation sample The neighborhood set of ; is the weight coefficient, the default value is 0.1; is the Gaussian kernel parameter, the default value is 1.

[0030] Information loss degree The calculation of adopts the general information loss degree calculation method of neural network, and the preferred formula is:

[0031] ;

[0032] In the formula, is the total number of samples, is the true probability distribution; is the predicted probability distribution; is the balance coefficient, which is used to adjust the weights of the two loss terms; For input The gradient operator of For input Information entropy of Represents the square of the L2 norm. In the prior art, the degree of information loss is ensured by cross entropy loss to ensure the consistency between the model output and the true label. Entropy gradient regularization is used to prevent excessive information loss during feature extraction. The prediction accuracy and information retention are comprehensively considered to ensure that the model remains accurate without losing too much information during feature extraction.

[0033] Furthermore, the occlusion perception equation group includes a visibility equation, an occlusion degree equation and an occlusion position equation, wherein the visibility equation in the occlusion perception equation group is used to estimate the visibility of each part of the target, with the input being the pixel position and depth information, and the output being the visibility value; the occlusion degree equation is used to quantify the severity of occlusion, with the input being the visibility map and the target mask, and the output being the occlusion degree map; the occlusion position equation is used to determine the specific location where occlusion occurs, with the input being the occlusion degree map, and the output being the occlusion position coordinates.

[0034] The visibility equation is expressed as: ,in Represents pixel The depth value of Indicates the perspective deviation, is the weight coefficient, is the attenuation coefficient.

[0035] The occlusion degree equation is expressed as: ,in is the target mask, is the target distance map, is the distance influencing factor.

[0036] The occlusion position equation is expressed as: ,in is the directional gradient operator, represents the convolution operation, is the shape prior, is the balance coefficient.

[0037] Furthermore, the partition division based on the region segmentation algorithm specifically includes: firstly, the image gradient is calculated using the improved multi-scale morphological gradient operator: ,in is the input image; then the adaptive threshold segmentation method is used to determine the candidate area: ,in is the regional grayscale mean, is the standard deviation, is the adjustment coefficient; finally, the region growing algorithm is used for fine segmentation, and the growth criterion is: ,in and are the mean and standard deviation of the seed region, respectively.

[0038] Furthermore, the specific formula of the spatial attention matrix is: ,in and are the height and width of the feature map respectively, and the matrix elements are calculated as follows: ,in For location The eigenvector at is the weight matrix, is the bias vector.

[0039] The specific formula of the channel attention matrix is: ,in is the number of channels of the feature map, and the matrix elements are calculated as follows: ,in For the The global pooling feature of each channel, is the weight matrix, is the bias vector.

[0040] The steps of adaptively compensating the target area for occlusion based on the occlusion distribution matrix specifically include: firstly calculating the feature reconstruction weight of the occluded area: ,in is the regularization coefficient, the default value is 1.5; then a multi-scale feature compensation network is constructed, which contains 3 branches: represents small-scale feature reconstruction, represents the mesoscale feature reconstruction, represents large-scale feature reconstruction, where Represents convolutional layers with different receptive fields; then adaptive feature fusion is performed: ,in is the adaptive weight coefficient, calculated by the attention mechanism: ; Finally, the residual connection is used to retain the valid information of the unoccluded area: ,in is the original feature map. In order to improve the compensation effect, reconstruction loss is introduced during the training process: ,in It is the feature map in the ideal unoccluded case. The compensation network uses an iterative optimization strategy, and updates the weight matrix in each iteration: ,in is the learning rate, set to 0.001, and the number of iterations is 5.

[0041] The process of solving the occlusion perception equations based on the spatial attention matrix and the channel attention matrix includes: firstly calculating the attention-weighted feature map ; then update the visibility estimate ; Then solve the degree of occlusion ;Finally locate the occlusion position .

[0042] The formula for the occlusion distribution matrix is: ,in Each element in the matrix represents the occlusion intensity at the corresponding position.

[0043] The steps of obtaining the preset threshold include: calculating the statistical characteristics of the occluded samples in the validation set and obtaining the mean and standard deviation ; Set the initial threshold ; The final threshold is obtained by optimizing the validation set ,in and Respectively represent the threshold The default value of the preset threshold is 0.65.

[0044] The formula of the target feature matrix is: ,in Indicates The elements of the target feature matrix include: target center coordinates , Width and Height , category probability and confidence score .

[0045] Furthermore, the target recognition model adopts an improved fast regional convolutional neural network structure, which specifically includes: a feature extraction backbone network (using a residual network), a region proposal network (using deformable convolution), a region of interest alignment layer, and a multi-task prediction head.

[0046] The steps for establishing the training data set for the target recognition model include: collecting satellite remote sensing images of target samples in different scenes, different lighting, and different angles; performing data enhancement (including random cropping, flipping, rotation, adjusting brightness and contrast, etc.); marking the target bounding box and category information; and dividing the training set, validation set, and test set (in a ratio of 6:2:2).

[0047] The training steps of the target recognition model include: initializing the backbone network with pre-trained weights; adopting a phased training strategy, first training the region proposal network and then training the entire network; the loss function includes classification loss (cross entropy), positioning loss (smooth L1) and occlusion prediction loss; the initial value of the learning rate is set to 0.01, and decays by 10 times every 30 cycles; using a stochastic gradient descent optimizer with a momentum coefficient of 0.9; training for 90 cycles until convergence.

[0048] A second aspect of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores program instructions, and when the program instructions are run in a computer, they are used to execute the above-mentioned target monitoring method of high-resolution satellite remote sensing images.

[0049] A third aspect of the present invention provides a target monitoring system for high-resolution satellite remote sensing images, which includes the above-mentioned computer-readable storage medium.

[0050] Compared with the prior art, the target monitoring method, medium and system of high-resolution satellite remote sensing images provided by the present invention have the following beneficial effects:

[0051] First, the original high-resolution image is partitioned by a regional segmentation algorithm to obtain multiple clear and stable sub-region images, laying the foundation for subsequent enhancement processing. Then, a pre-trained global-region enhancement network is used to iteratively enhance the sub-region images and the overall image to fully explore local details and overall features. In order to balance local enhancement and global enhancement, the present invention establishes a game optimization model, using network parameter complexity, feature space consistency and information loss as adjustment factors, and the degree of sub-region enhancement and global enhancement as the two parties in the game. The optimal balance between the two is achieved through the minimum maximization optimization strategy, making it more suitable for different types of targets in complex scenes.

[0052] Secondly, based on the target features extracted by the attention mechanism and combined with the preset occlusion perception model, the occlusion distribution of the target area is accurately estimated. For targets with occlusion, the present invention proposes an adaptive feature compensation strategy, which uses multi-scale feature fusion and attention mechanism to effectively compensate for the detailed information of the occluded area and extract clearer and more complete target features.

[0053] Finally, the compensated target features are input into the improved deep learning target recognition model to achieve accurate classification and positioning of the target and output high-quality monitoring results.

[0054] In summary, the present invention solves the problem that the existing technology for high-resolution satellite remote sensing image enhancement processing is difficult to cope with different targets in complex scenes. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 A flow chart of the method provided by the present invention;

[0056] Figure 2 The remote sensing image gradient value distribution map in Example 1;

[0057] Figure 3 It is a trend diagram of the loss function of the generator and the discriminator in the game optimization process in Example 1;

[0058] Figure 4 This is a comparison chart of the detection performance of various types of waste in Example 2. DETAILED DESCRIPTION

[0059] In order to make the purpose, technical solution and advantages of the embodiments of the present invention more clear, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0060] like Figure 1 As shown, it is a flow chart of a target monitoring method of high-resolution satellite remote sensing images provided by the first aspect of the present invention. The specific implementation of each step in this method is described in detail as follows:

[0061] The specific implementation of step S10 is as follows:

[0062] First, the original high-resolution satellite remote sensing images are preprocessed, including data format conversion, radiation correction, and geometric correction, to eliminate interference introduced by sensor equipment, atmosphere, and terrain, and obtain clear and stable initial preprocessed images. .

[0063] Then, the image gradient information is calculated based on the improved multi-scale morphological gradient operator. The gradient calculation formula is: ,in represents the pixel position coordinates, and Respectively indicate along Axis and Partial derivatives about the axes. Gradient map It can highlight the edges, textures and other significant features in the image, providing a basis for subsequent region segmentation.

[0064] Next, the adaptive threshold segmentation method is used to determine the candidate area. The threshold calculation formula is: ,in and Represent the mean and standard deviation of the regional grayscale, is the adjustment coefficient, and its value is generally Through this adaptive threshold selection method, different segmentation strengths can be used for regions with different features to improve the robustness of segmentation.

[0065] Finally, the improved region growing algorithm is used to finely segment the candidate region. The growing criterion is: ,in represents the current pixel position, and are the grayscale mean and standard deviation of the seed region, respectively. This criterion ensures that the grayscale difference of adjacent pixels does not exceed twice the standard deviation of the seed region, thus ensuring the connectivity and accuracy of the segmentation results.

[0066] Through the above steps, the original high-resolution satellite remote sensing image is divided into multiple sub-region images. , providing a basis for subsequent image enhancement and target detection.

[0067] The specific implementation of step S20 is as follows:

[0068] This step uses a pre-trained global-region enhancement network to iteratively enhance each sub-region image to obtain a sub-region enhanced temporary image, and splices all sub-region enhanced temporary images obtained in each iteration to form a global temporary image. The global-region enhancement network includes a region generation sub-network and a global splicing sub-network.

[0069] First, the region generation sub-network is used to generate a single sub-region image This sub-network adopts a multi-branch residual network structure, which can effectively extract and enhance local features. The network structure includes multiple residual blocks, each of which consists of two It consists of a convolutional layer, a BN layer, and a ReLU activation layer.

[0070] Secondly, the global splicing sub-network is used to enhance the overall features. This sub-network adopts an attention-enhanced encoder-decoder structure, which can adaptively extract and fuse multi-scale global features. The network structure includes an encoder and a decoder. The encoder consists of multiple convolution + pooling layers to extract multi-scale features; the decoder consists of multiple deconvolution + upsampling layers, and uses an attention mechanism to enhance feature expression capabilities.

[0071] Finally, through iterative optimization, the parameters of the region generation subnetwork and the global stitching subnetwork are updated alternately to generate enhanced temporary subregion images. and global temporary images The purpose of sub-region enhancement is to improve local details, while global enhancement aims to enhance overall information. The two work together to improve image quality.

[0072] The specific implementation of step S30 is as follows:

[0073] This step established a game optimization model, with the network parameter complexity, feature space consistency and information loss degree of the regional generation sub-network and the global splicing sub-network as adjustment factors, and the enhancement degree of the sub-regional enhanced temporary image and the enhancement degree of the global temporary image as the two parties of the game, to achieve a balance between local and global features through minimum and maximum optimization.

[0074] First, define the objective function of the game optimization model:

[0075] ;

[0076] in, represents the region generation subnetwork, represents the global splicing subnetwork, is the real data distribution, is the noise distribution, is the weight coefficient, and its value range is . is the network parameter complexity, is the feature space consistency, The degree of information loss.

[0077] Network parameter complexity The calculation formula is:

[0078] ;

[0079] in, is the number of network layers, For the The number of neurons in the layer, For the The input dimension of the layer, For the The weight matrix of the layer, For the The bias vector of the layer, is the balance coefficient, with a value of 0.5. The lower the network complexity, the simpler the model, but it may not fit the data distribution well.

[0080] Feature space consistency The calculation formula is:

[0081] ;

[0082] in, is the sample size, and They represent the feature extraction functions of the region generation sub-network and the global splicing sub-network respectively. is the input sample, Representation sample The neighborhood set of is the weight coefficient, with a value of 0.1. is the Gaussian kernel parameter, and its value is 1. This item encourages the sub-region features to be semantically consistent with the global features.

[0083] Information loss degree The calculation formula is:

[0084] ;

[0085] in, and represent the true probability distribution and the predicted probability distribution respectively, is the weight coefficient. This term hopes to minimize the difference in probability distribution and the norm of the gradient to reduce information loss.

[0086] Finally, through the minimum-maximum optimization strategy, the network parameter complexity, feature space consistency and information loss degree are adjusted to balance the enhancement degree of the sub-region enhanced temporary image and the global temporary image, and the optimal game result is obtained. This game optimization process can achieve the best balance between local features and global features, laying the foundation for subsequent refined preprocessing and target monitoring.

[0087] The specific implementation of step S40 is as follows:

[0088] This step updates the parameters of the global-region enhancement network based on the game optimization result obtained in step S30, and uses the updated network to iteratively optimize the sub-region image and the overall image to generate multiple candidate refined images. Finally, the generated candidate images are comprehensively evaluated in terms of clarity, information entropy, edge preservation, and other indicators, and the optimal refined image is selected as the final preprocessing result.

[0089] First, according to the game optimization result of step S30, the parameters of the global-regional enhancement network are updated. Among them, the parameters of the regional generation subnetwork are given by Update, the parameters of the global splicing subnetwork are given by This process needs to consider multiple factors such as network complexity, feature consistency and information loss to achieve the best balance between local details and overall information.

[0090] Then, the updated global-regional enhancement network is used to enhance each sub-region image. and overall image Perform iterative optimization to generate multiple candidate refined images This process aims to further improve the clarity of the image, maintain edge details and enrich the information content.

[0091] Finally, the generated multiple candidate refined images are evaluated. The specific indicators include:

[0092] Clarity ,Use the Laplacian operator to calculate the image gradient energy, the formula is:

[0093] ;

[0094] in, Represents the Laplace operator. The higher the definition, the clearer the image.

[0095] Information Entropy , used to quantify the information content of the image, the formula is:

[0096] ;

[0097] in, For the The higher the information entropy, the richer the image information.

[0098] Edge retention , use the Sobel operator to calculate the edge response intensity, the formula is:

[0099] ;

[0100] in, and Respectively indicate along Axis and The higher the edge preservation, the more complete the detail information.

[0101] Finally, the best refined image is selected as the final preprocessing result. The purpose of this step is to further optimize the global-regional enhancement network based on the aforementioned game optimization strategy, generate the best refined image, and provide high-quality input data for subsequent target monitoring.

[0102] The specific implementation of step S50 is as follows:

[0103] This step uses deep learning target detection technology to perform a preliminary scan of the preprocessed high-resolution image based on preset target features to generate multiple target candidate areas.

[0104] First, a training data set for the target detection model is established, including: collecting target sample images in different scenes, different lighting and different angles, performing data enhancement (such as random cropping, flipping, rotation, adjusting brightness and contrast, etc.), annotating the target bounding box and category information, and dividing them into training set, validation set and test set (ratio is 6:2:2).

[0105] Secondly, the improved Faster Regional Convolutional Neural Network (FasterR-CNN) is used as the target detection model. The model consists of a feature extraction backbone network, a region proposal network, a region of interest alignment layer, and a multi-task prediction head. The backbone network uses a residual network structure, which can effectively extract the semantic features of the image; the region proposal network uses deformable convolution, which can adaptively sense the geometric shape of the target; the region of interest alignment layer is used to extract the features of the region of interest; the multi-task prediction head performs target classification and bounding box regression simultaneously.

[0106] During the training process, the region proposal network is trained first, and then the entire network is trained. The loss function includes classification loss (cross entropy), positioning loss (smoothing ) and occlusion prediction loss.

[0107] Finally, the high-resolution image that has been finely preprocessed in step S40 is input into the trained target detection model for preliminary scanning to generate multiple target candidate regions. Each candidate region contains feature information such as the target's location coordinates, size, category probability, and confidence.

[0108] The specific implementation of step S60 is as follows:

[0109] This step uses a multi-layer attention mechanism to extract features from each target candidate region and construct a spatial attention matrix and a channel attention matrix to highlight significant spatial regions and channel features.

[0110] First, calculate the spatial attention matrix ,in and are the height and width of the feature map respectively. The matrix elements The calculation formula is:

[0111] ;

[0112] in, For location The eigenvector at is the weight matrix, is the bias vector. The spatial attention matrix can highlight the salient spatial regions.

[0113] Secondly, calculate the channel attention matrix ,in is the number of channels of the feature map. The matrix elements The calculation formula is:

[0114] ;

[0115] in, For the The global pooling feature of each channel, is the weight matrix, is the bias vector. The channel attention matrix can highlight significant channel features.

[0116] Finally, the spatial attention matrix and channel attention matrix are applied to the feature map of the target candidate region to obtain the attention-weighted feature representation ,in is the original feature map. This multi-layer attention mechanism can effectively extract the key features of the target area and provide better input for subsequent occlusion perception and target recognition.

[0117] The specific implementation of step S70 is as follows:

[0118] This step solves the preset occlusion perception equations based on the spatial attention matrix and channel attention matrix calculated in step S60 to obtain the occlusion distribution matrix.

[0119] First, define the occlusion perception equations, including the visibility equation, the occlusion degree equation and the occlusion position equation.

[0120] Then, the spatial attention matrix calculated in step S60 is used and the channel attention matrix , solve the above occlusion perception equations. The specific process includes:

[0121] Compute attention-weighted feature maps ;

[0122] Update viewability estimates ;

[0123] Solving the degree of occlusion ;

[0124] Locate the occlusion position .

[0125] Finally, the occlusion distribution matrix is ​​constructed based on the above results ,in ,Each element in the matrix represents the occlusion intensity of the corresponding position.,The purpose of this step is to use the feature information extracted by the,attention mechanism, combined with the preset occlusion perception model, to accurately estimate the,occlusion distribution of the target area and provide a basis for the subsequent,occlusion compensation.

[0126] The specific implementation of step S80 is as follows:

[0127] This step adaptively compensates the occluded target area based on the occlusion distribution matrix obtained in step S70 to extract clearer and more complete target features and construct a target feature matrix.

[0128] First, calculate the feature reconstruction weights of the occluded area ,in is the regularization coefficient, and its value is 1.5. This weight reflects the degree of occlusion at different positions and provides a basis for subsequent feature compensation.

[0129] Then, a multi-scale feature compensation network is constructed, which contains 3 branches:

[0130] represents small-scale feature reconstruction,

[0131] represents the mesoscale feature reconstruction,

[0132] represents large-scale feature reconstruction,

[0133] in Convolutional layers representing different receptive fields. This multi-scale feature reconstruction can better compensate for the details of occluded areas.

[0134] Next, perform feature adaptive fusion:

[0135] ;

[0136] in, is the adaptive weight coefficient, calculated by the attention mechanism:

[0137] ;

[0138] The attention mechanism can automatically adjust the fusion weights according to the significance of features at different scales.

[0139] Finally, the residual connection is used to retain the valid information of the unoccluded area:

[0140] ;

[0141] In order to improve the compensation effect, reconstruction loss is introduced during training:

[0142] ;

[0143] in, It is the feature map of the ideal unobstructed case. The compensation network uses an iterative optimization strategy, and updates the weight matrix in each iteration. ,in is the learning rate, the value is 0.001, and the number of iterations is 5.

[0144] Finally, according to the compensated feature map Constructing the target feature matrix ,in Indicates The feature vector of the target, including the position coordinates , Width and Height , category probability and confidence The purpose of this step is to effectively compensate for occluded targets, extract clearer and more complete target features, and provide high-quality input for subsequent target recognition.

[0145] The specific implementation of step S90 is as follows:

[0146] This step uses the improved Faster Regional Convolutional Neural Network (FasterR-CNN) as the target recognition model, and transforms the target feature matrix constructed in step S80 into As input, it realizes the category recognition and location positioning of the target and outputs the final target monitoring result.

[0147] First, the improved FasterR-CNN model consists of a feature extraction backbone network, a region proposal network, an interest region alignment layer, and a multi-task prediction head. The backbone network uses a residual network structure, which can effectively extract the semantic features of the image; the region proposal network uses deformable convolution, which can adaptively sense the geometric shape of the target; the interest region alignment layer is used to extract the features of the interest region; the multi-task prediction head performs target classification and bounding box regression simultaneously.

[0148] During the training process, the region proposal network is trained first, and then the entire network is trained. The loss function includes classification loss (cross entropy), positioning loss (smoothing ) and occlusion prediction loss. Classification loss The formula is:

[0149] ;

[0150] in, and Represent the true category label and predicted probability respectively. Positioning loss The formula is:

[0151] ;

[0152] in, and Denote the predicted bounding box parameters and the ground-truth bounding box parameters respectively. The occlusion prediction loss encourages the model to learn to recognize occluded objects.

[0153] Finally, the target feature matrix constructed in step S80 is Input the trained target recognition model to realize target category recognition and location positioning, and output the final target monitoring results.

[0154] A second aspect of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores program instructions, and when the program instructions are run in a computer, they are used to execute the above-mentioned target monitoring method of high-resolution satellite remote sensing images.

[0155] A third aspect of the present invention provides a target monitoring system for high-resolution satellite remote sensing images, which includes the above-mentioned computer-readable storage medium.

[0156] Specifically, the principle of the present invention is:

[0157] 1. Region segmentation and global-region enhancement

[0158] First, the present invention uses the regional segmentation algorithm to partition the original high-resolution image to obtain multiple clear and stable sub-region images. Through this segmentation strategy, the local detail information in the image can be better processed, laying the foundation for subsequent enhancement processing.

[0159] Then, the pre-trained global-region enhancement network is used to iteratively enhance the sub-region images and the overall image. The region generation sub-network uses a multi-branch residual structure to effectively extract and enhance the local features of the sub-region; the global splicing sub-network adopts an attention-enhanced encoder-decoder structure to adaptively extract and fuse multi-scale global features. The two coordinate with each other to jointly improve the overall quality of the image.

[0160] 2. Game optimization strategy

[0161] The existing technology for enhancing remote sensing images either pays too much attention to local details, resulting in the destruction of global consistency, or pursues global effects too much, resulting in the loss of local details. There is no effective mechanism to balance the relationship between the two. In order to balance local enhancement and global enhancement, the present invention establishes a game optimization model. The model uses network parameter complexity, feature space consistency and information loss as adjustment factors, and the degree of sub-region enhancement and global enhancement as the two parties in the game, and achieves the best balance between the two through the minimum maximization optimization strategy.

[0162] This game optimization process can ensure that local and global features reach the optimal state, fully explore the detailed information of the sub-region, and maintain the semantic consistency of the overall image, laying a good foundation for subsequent refined preprocessing and target monitoring.

[0163] 3. Occlusion perception and adaptive compensation

[0164] The present invention uses the attention mechanism to extract target features and combines it with a preset occlusion perception model to accurately estimate the occlusion distribution of the target area. The occlusion perception model includes the visibility equation, the occlusion degree equation and the occlusion position equation, which can comprehensively describe the visibility, occlusion severity and occlusion position of the target.

[0165] For targets with occlusion, the present invention proposes an adaptive feature compensation strategy. Through multi-scale feature fusion and attention mechanism, the detailed information of the occluded area is effectively compensated, thereby extracting clearer and more complete target features. This compensation strategy can make adaptive adjustments to different degrees of occlusion, greatly improving the detection ability of occluded targets.

[0166] 4. Deep Learning Object Recognition

[0167] Finally, the present invention inputs the compensated target features into the improved deep learning target recognition model to achieve accurate classification and positioning of the target. The model uses deformable convolution and attention mechanism to enhance the feature extraction capability and shows strong robustness in complex scenes.

[0168] The following is an example 1 of a specific application scenario of the present invention: a certain unit needs to use high-resolution optical satellite data in commercial operations to dynamically track and observe the planting conditions of field crops in coastal counties.

[0169] First, the unit purchased high-resolution satellite image data of the pilot area from March to September 2022 from commercial remote sensing data, totaling 7 scenes. The ground resolution of these image data is 0.5 meters, and the imaging bands include four bands: blue, green, red, and near-infrared. Taking into account factors such as cloud cover and atmospheric interference that may exist during satellite observation, the quality of these original image data is uneven, with a certain degree of blur and distortion.

[0170] In order to improve the subsequent monitoring and analysis results, the unit decided to use the high-resolution satellite remote sensing image target monitoring method proposed in this invention to pre-process the above image data and detect targets. The specific implementation steps are as follows:

[0171] Step S10: Preprocessing high-resolution satellite remote sensing images

[0172] First, the original 7-view satellite image data were subjected to radiation correction and geometric correction to eliminate the interference introduced by factors such as atmosphere and terrain. After this preprocessing step, the clarity and geometric accuracy of the image were improved to a certain extent.

[0173] Then, the improved multi-scale morphological gradient operator is used to calculate the gradient of the preprocessed image to obtain the gradient map. Taking the image of March 25, 2022 as an example, the gradient calculation results are shown in Table 1:

[0174] Table 1 Image gradient calculation results on March 25, 2022

[0175]

[0176] Figure 2 The distribution of gradient values ​​of the image on March 25, 2022 is shown. The frequency distribution of gradient values ​​is displayed in the form of a histogram, and the red dotted line indicates the average gradient value of 0.13.

[0177] Next, the adaptive threshold segmentation method is used to binarize the gradient image and determine the candidate area. The threshold calculation formula is: ,in is the mean value of the gradient, This threshold setting can better extract the salient areas in the image without missing out on valuable information due to being too strict.

[0178] Finally, the improved region growing algorithm is used to finely segment the candidate region, and the growth criterion is ,in and are the mean and standard deviation of the seed region respectively.Through this step, the original image is divided into 270 sub-regions.

[0179] Step S20: Using global-regional enhancement network for image enhancement

[0180] For the above 270 sub-region images, the unit uses a pre-trained global-region enhancement network for iterative enhancement. The network consists of two parts: a region generation sub-network and a global splicing sub-network.

[0181] The region generation sub-network uses a multi-branch residual structure, with the original sub-region image as input and the enhanced sub-region feature map as output. The specific parameters of the network are shown in Table 2:

[0182] Table 2 Region generation sub-network parameters

[0183]

[0184] The global splicing sub-network adopts an attention-enhanced encoder-decoder structure, with the input being the spliced ​​global temporary image and the output being the enhanced global feature map. The specific parameters of the network are shown in Table 3:

[0185] Table 3 Global splicing sub-network parameters

[0186]

[0187] Through 5 rounds of iterative optimization, the parameters of the region generation subnetwork and the global stitching subnetwork were updated alternately, and finally 270 enhanced sub-regional temporary images and 1 enhanced global temporary image were generated.

[0188] Step S30: Establishing a game optimization model

[0189] In order to balance local enhancement and global enhancement, the unit established the following game optimization model:

[0190] Objective function: ;

[0191] Network parameter complexity Calculation:

[0192] ;

[0193] Feature space consistency Calculation:

[0194] ;

[0195] Information loss degree Calculation:

[0196] ;

[0197] After 150 rounds of iterative optimization, the optimal game result was obtained. Figure 3 The changing trend of the loss functions of the generator and the discriminator during the game optimization process is shown. It can be seen that as the number of iterations increases, the two networks gradually reach a balance.

[0198] Step S40: Pre-processing

[0199] According to the game optimization results of step S30, the unit updated the parameters of the global-regional enhancement network. Then, the updated network was used to iteratively optimize the 270 sub-region images and the global temporary image, generating 10 candidate refined images.

[0200] The 10 candidate images were evaluated using the following indicators:

[0201] 1) Clarity : Use the Laplace operator to calculate the gradient energy, and the results are shown in Table 4.

[0202] Table 4 Alternative image clarity evaluation

[0203]

[0204] 2) Information Entropy :Quantify the image information content, the results are shown in Table 5.

[0205] Table 5 Evaluation of information entropy of candidate images

[0206]

[0207] 3) Edge retention :The edge response intensity is calculated using the Sobel operator. The results are shown in Table 6.

[0208] Table 6 Evaluation of edge preservation of candidate images

[0209]

[0210] Based on the three indicators, the unit selected image 3 as the final preprocessing result. The clarity, information entropy and edge preservation of this image are all at a high level, which can provide high-quality input data for subsequent target monitoring.

[0211] Step S50: Target detection using deep learning

[0212] For the satellite images that have been finely preprocessed, the unit uses the improved Faster Regional Convolutional Neural Network (FasterR-CNN) to perform preliminary target detection. The specific parameters of the model are shown in Table 7:

[0213] Table 7 Target detection model parameters

[0214]

[0215] After training, the model achieved 85% target detection accuracy on the validation set. It was applied to the satellite imagery of March 25, 2022, and a total of 156 food crop target areas were detected. The detection results are shown in Table 8:

[0216] Table 8 Target detection results

[0217]

[0218] Step S60: Extract target features using the attention mechanism

[0219] For the above 156 detection targets, the unit uses the attention mechanism to extract their feature information. The specific calculation is as follows:

[0220] First, calculate the spatial attention matrix ,in and are the height and width of the target feature map respectively. The matrix elements The calculation formula is:

[0221] ;

[0222] Secondly, calculate the channel attention matrix ,in is the number of channels of the target feature map. Matrix elements The calculation formula is:

[0223] ;

[0224] Finally, the spatial attention matrix and channel attention matrix are applied to the target feature map to obtain the attention-weighted feature representation .

[0225] Step S70: Estimating target occlusion distribution

[0226] Based on the attention weighted features extracted in step S60, the unit uses a preset occlusion perception equation set to estimate the target occlusion situation.

[0227] First, the visibility equation is calculated as follows:

[0228] ;

[0229] in, Represents pixel The depth value of Indicates perspective deviation.

[0230] Then, the occlusion degree equation is calculated as follows:

[0231] ;

[0232] in, is the target mask, The target distance map.

[0233] Finally, the occlusion position equation is calculated as follows:

[0234] ;

[0235] in, is the directional gradient operator, is the shape prior.

[0236] Combining the above calculation results, we get the occlusion distribution matrix ,in .

[0237] After analysis, among the 156 detected targets, 52 targets have obvious occlusion, accounting for 33.3%. In the occlusion distribution matrix of these targets, 16 elements have values ​​greater than 0.7, indicating serious occlusion.

[0238] Step S80: Adaptively compensate for the occluded target

[0239] According to the occlusion distribution matrix obtained in step S70, the unit performs feature compensation on 52 target areas with occlusion.

[0240] First, calculate the feature reconstruction weights of the occluded area .

[0241] Then, a multi-scale feature compensation network is constructed, which contains 3 branches:

[0242] represents small-scale feature reconstruction,

[0243] represents the mesoscale feature reconstruction,

[0244] represents large-scale feature reconstruction,

[0245] in Convolutional layers representing different receptive fields.

[0246] Next, perform feature adaptive fusion:

[0247] ;

[0248] Among them, the fusion weight Automatically adjusted via attention mechanism.

[0249] Finally, the residual connection is used to retain the valid information of the unoccluded area:

[0250] ;

[0251] During the training process, reconstruction loss is introduced ,After 5 rounds of optimization, the final compensation result was obtained.

[0252] Step S90: Using an improved target recognition model for monitoring

[0253] The unit inputs the target feature matrix constructed in step S80 into the improved FasterR-CNN model for final target recognition. The specific parameters of the model are shown in Table 9:

[0254] Table 9 Target recognition model parameters

[0255]

[0256] After training, the model achieved a target recognition accuracy of 92% on the test set. It was applied to the satellite imagery of March 25, 2022, and finally 143 food crop targets were identified, including 73 wheat targets, 44 corn targets, and 26 rice targets. The monitoring results are summarized in Table 10:

[0257] Table 10 Target monitoring results

[0258]

[0259] Through the high-resolution satellite remote sensing image target monitoring method proposed in this invention, the unit successfully realized the dynamic monitoring of grain crop planting in the pilot area. Compared with the existing preprocessing and target detection technology, this method has significant improvements in image quality improvement, occlusion information compensation and final monitoring accuracy.

[0260] The following is an example 2 of a specific application scenario of the present invention: This example 2 takes the waste monitoring in a certain coastline area as an example to explain in detail the specific implementation process of the present invention. The coastline is about 15 kilometers long and contains a variety of landforms such as beaches, reefs, and docks. The main goal of the monitoring task is to identify and locate various types of garbage and waste in the coastline area, including plastic waste, fishing gear, construction waste, etc.

[0261] Data acquisition and preprocessing stage: This embodiment 2 uses image data acquired by a certain model of high-resolution optical remote sensing satellite, with a spatial resolution of 0.5 meters and an image size of 12000×8000 pixels, including panchromatic bands and multispectral bands. Specific image parameters are shown in Table 11 below:

[0262] Table 11 Image parameter table

[0263]

[0264] First, the original image is subjected to radiation correction and geometric correction. The radiation correction adopts the absolute radiation correction method, and uses 4 ground reflectance control points obtained on site for parameter calculation. The geometric correction adopts the orthorectification method, and uses 10 ground control points to achieve sub-pixel accuracy geometric correction. The positioning accuracy of the corrected image is better than 1 meter, which meets the requirements of subsequent processing. Then the image is segmented, and the improved multi-scale morphological gradient operator is used for boundary extraction. The gradient threshold is set to 25, and the minimum area threshold is set to 100 square meters. Finally, the image is divided into 2876 sub-areas.

[0265] Image enhancement stage: The pre-trained global region enhancement network is used to iteratively enhance the segmented sub-region images. The network structure includes a region generation sub-network and a global stitching sub-network. The specific parameters are shown in Table 12 below:

[0266] Table 12 Network parameter table

[0267]

[0268] During the enhancement process, the number of iterations was set to 10, and the enhancement degree of the sub-region image and the global image was evaluated after each iteration. The average clarity of the sub-region enhanced temporary image was improved by 32.6%, and the information entropy increased by 18.4%. The edge preservation of the global temporary image was improved by 25.8%, and the overall contrast was improved by 21.3%.

[0269] Game optimization stage: In order to balance local features and global features, a game optimization model is constructed. The network parameter complexity weight λ1 is set to 0.3, the feature space consistency weight λ2 is set to 0.4, and the information loss degree weight λ3 is set to 0.3. After 50 rounds of iterative optimization, the model converges, and the final optimization indicators are shown in Table 13 below:

[0270] Table 13 Game adjustment factors

[0271]

[0272] Image refinement processing stage: Based on the game optimization results, the global region enhancement network is updated, and the sub-region images and the overall image are refined. Five groups of candidate refined images are generated and evaluated from three dimensions: clarity, information entropy, and edge preservation. The results are shown in Table 14 below:

[0273] Table 14 Image evaluation table

[0274]

[0275] According to the evaluation results, Scheme 3 was selected as the final refinement result, and its comprehensive score of clarity, information entropy, and edge preservation was the highest, which was 0.901.

[0276] Target detection stage: Use a deep target detection network to perform a preliminary scan on the refined image. The network uses an improved fast RCNN structure and is trained using 3,000 labeled samples, which include six types of targets such as plastic waste, fishing gear, and construction waste. The training parameters and detection results are shown in Tables 15 and 16 below:

[0277] Table 15 Target detection training parameters

[0278]

[0279] Table 16 Test results

[0280]

[0281] Figure 4 The bar chart shows the detection accuracy and recall rate of the system for six types of waste. The detection accuracy of abandoned ships is the highest, reaching 0.923, while the detection performance of other types of garbage is relatively low, which is related to the distribution of the number of samples.

[0282] Feature extraction stage:

[0283] A multi-layer attention mechanism is applied to the detected target candidate regions for feature extraction. The size of the spatial attention matrix is ​​28×28, and the size of the channel attention matrix is ​​512×512. The calculation results of the attention weights are shown in Table 17 below:

[0284] Table 17 Weight calculation table

[0285]

[0286] Occlusion processing stage: Calculate the occlusion distribution matrix based on the attention features, and set the occlusion judgment threshold to 0.65. In this monitoring, 42 targets were found to have varying degrees of occlusion, mainly concentrated in the dock area and reef-dense areas. The occlusion degree statistics are shown in Table 18 below:

[0287] Table 18 Occlusion statistics

[0288]

[0289] Adaptive compensation is performed on the occluded target. The compensation effect evaluation results are shown in Table 19 below:

[0290] Table 19 Compensation effect table

[0291]

[0292] Final recognition result

[0293] After a complete processing process, the system finally identified 368 pieces of garbage and waste in the coastline area, with a total area of ​​about 2,860 square meters. The specific statistical results are shown in Table 20 below:

[0294] Table 20 Final recognition results

[0295]

[0296] This Example 2 successfully achieved automatic monitoring of waste in the coastline area through a systematic processing flow. The overall performance of the system reached the expected goal, with a detection accuracy of 89.2%, a missed detection rate of less than 8.5%, and a false detection rate of less than 6.8%. The deployment of this system significantly improves the efficiency and accuracy of coastline environmental monitoring and provides strong technical support for marine environmental protection. The experimental results show that the method proposed in the present invention has good adaptability and stability when dealing with target monitoring problems in complex scenarios.

[0297] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A target monitoring method for high-resolution satellite remote sensing images, characterized in that: The following steps are involved: S10, preprocessing the high-resolution satellite remote sensing image to generate an initial preprocessed image, and partitioning the initial preprocessed image based on a region segmentation algorithm to obtain a plurality of sub-region images; S20, using a pre-trained global-region enhancement network to iteratively enhance each sub-region image to obtain a sub-region enhanced temporary image, and splicing all sub-region enhanced temporary images obtained in each iteration into a global temporary image; S30, using the network parameter complexity, feature space consistency and information loss degree of the regional generation sub-network and the global splicing sub-network as adjustment factors, and using the enhancement degree of the sub-region enhanced temporary image and the enhancement degree of the global temporary image as the two parties of the game, to establish a game optimization model for achieving a balance between local features and global features; S40, based on the optimization result of the game optimization model, updating the global-regional enhancement network, using the updated global-regional enhancement network to iteratively optimize the sub-region image and the overall image to generate multiple candidate refined images; selecting the best refined image among the candidate images in terms of clarity, information entropy, and edge preservation as the final preprocessed image; S50, performing a preliminary scan on the final preprocessed image using a deep target detection network according to preset target features to generate multiple target candidate regions; S60, using a multi-layer attention mechanism to extract features from each target candidate region, and construct a spatial attention matrix and a channel attention matrix; S70, solving a preset occlusion perception equation group based on the spatial attention matrix and the channel attention matrix to obtain an occlusion distribution matrix; S80, if there are elements in the occlusion distribution matrix that are greater than a preset threshold, it means that the target is occluded, then adaptive occlusion compensation is performed on the target area based on the occlusion distribution matrix, a multi-scale feature fusion method is used to extract target features, a target feature matrix is ​​constructed, and the process proceeds to step S90; Otherwise, directly use the multi-scale feature fusion method to extract the target features, construct the target feature matrix and enter step S90; S90, inputting the target feature matrix into a pre-trained target recognition model to achieve target category recognition and position positioning, and outputting target monitoring results; The global-region enhancement network includes a region generation subnetwork and a global splicing subnetwork. The region generation subnetwork is used to perform feature enhancement on each subregion image, and the global splicing subnetwork performs feature enhancement on the global temporary image obtained by each iteration. The region generation subnetwork adopts a multi-branch residual network structure, with the original subregion image as input and the enhanced subregion feature map as output. The global splicing subnetwork adopts an attention-enhanced encoder-decoder structure, with the spliced ​​global temporary image as input and the enhanced global feature map as output. The process of solving the occlusion perception equations based on the spatial attention matrix and the channel attention matrix includes: firstly, calculating the attention-weighted feature map ; then update the visibility estimate ; Then solve the degree of occlusion ;Finally locate the occlusion position ;in, is the spatial attention matrix, is the attention matrix, is the original feature map, is the visibility estimate, is the target mask, The formula for the occlusion distribution matrix is: ,in , each element in the matrix represents the occlusion intensity of the corresponding position.

2. The target monitoring method of high-resolution satellite remote sensing images according to claim 1 is characterized in that: The objective function of the game optimization model is expressed as follows: The objective function of the game optimization model is expressed as follows: ; In the formula, represents the region generation subnetwork, represents the global splicing subnetwork, represents mathematical expectation; Represents the real data distribution; represents the noise distribution; Representation variables obey distributed expectations; Representation variables obey Distribution expect; is the weight coefficient, and its value range is [0,1]; is the network parameter complexity; is the feature space consistency; is the degree of information loss; Consider generating a network Try to minimize the objective function, Considering the discriminant network Trying to maximize the objective function, the game adversarial approach ensures the dynamic balance of the two networks.

3. The target monitoring method of high-resolution satellite remote sensing images according to claim 2 is characterized in that: The occlusion perception equation group includes a visibility function, an occlusion degree function and an occlusion position function, wherein the visibility function in the occlusion perception equation group is used to estimate the visibility of each part of the target, with the input being the pixel position and depth information, and the output being the visibility value; the occlusion degree function is used to quantify the severity of occlusion, with the input being the visibility map and the target mask, and the output being the occlusion degree map; the occlusion position function is used to determine the specific location where occlusion occurs, with the input being the occlusion degree map, and the output being the occlusion position coordinates.

4. The target monitoring method of high-resolution satellite remote sensing images according to claim 3 is characterized in that: The visibility function is expressed as: ， in Represents pixel The depth value of Indicates the perspective deviation, is the weight coefficient, is the attenuation coefficient.

5. The target monitoring method of high-resolution satellite remote sensing images according to claim 4 is characterized in that: The occlusion degree function is expressed as: ， in is the target mask, is the target distance map, is the distance influencing factor.

6. The target monitoring method of high-resolution satellite remote sensing images according to claim 5, characterized in that: The occlusion position function is expressed as: ， in is the directional gradient operator, represents the convolution operation, is the shape prior, is the balance coefficient.

7. The target monitoring method of high-resolution satellite remote sensing images according to claim 6, characterized in that: The target recognition model adopts an improved fast regional convolutional neural network structure.

8. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores program instructions, and when the program instructions are executed in a computer, they are used to execute the target monitoring method of high-resolution satellite remote sensing images according to any one of claims 1 to 7.

9. A target monitoring system for high-resolution satellite remote sensing images, characterized in that: Contains the computer-readable storage medium of claim 8.

Images