Remote sensing image target detection method and system

Abstract

The invention relates to the technical field of remote sensing image processing, in particular to a remote sensing image target detection method and system, and the method comprises the steps: obtaining a to-be-detected remote sensing image, and inputting the to-be-detected remote sensing image into a trained target detection model; a data processing module in the model performs feature processing on a remote sensing image to obtain a multi-scale feature map. And the adaptive occlusion sensing module performs occlusion processing on the multi-scale feature graph to obtain an image occlusion mask graph, and then performs feature optimization according to the image occlusion mask graph. And the mapping module performs feature mapping to obtain mapping features, and finally obtains a shielding region detection result according to the mapping features. According to the model, target detection can be effectively carried out on remote sensing images through multi-module cooperation, the detection accuracy and reliability are improved, and powerful support is provided for remote sensing image analysis and application. The method solves the problems that an existing target detection method is high in detection speed, and target shielding is well handled.

Description

Technical Field

[0001] The invention relates to the technical field of remote sensing image processing, and in particular to a remote sensing image target detection method and system. Background Art

[0002] Current target detection methods can be roughly divided into two categories: Anchor-based detection methods and Anchor-free detection methods. Anchor-based detection methods mainly include FasterRCNN. The main implementation steps of this method are: extract feature maps from the input image through a convolutional neural network (CNN). Then, the feature map is sent to the RegionProposalNetwork (RPN) to generate candidate regions. RPN generates multiple anchor boxes through a sliding window, and uses classification and regression branches to predict the category and position of each anchor box. The generated candidate regions are processed by non-maximum suppression (NMS) to filter out the optimal region. Then, these candidate regions are passed to the FastR-CNN network for accurate classification and detection box regression to obtain the final target category and position. This method uses RPN to generate candidate regions. RPN generates multi-scale and multi-ratio anchor boxes at each pixel of the feature map through a sliding window, but most of these anchor boxes have little effect on the detection results, but greatly increase the computational complexity of the entire network, affecting the detection speed. FCOS (Fully Convolutional One-Stage Object Detection) is an anchor-free target detection algorithm. It makes predictions directly on the feature map and does not rely on predefined anchor boxes. The algorithm first extracts the feature map through a convolutional neural network, and then predicts the target category, center, and regression parameters of the detection box at each pixel position. This information is used to generate the final target detection box. Although this method does not rely on predefined anchor boxes, it greatly reduces the computational complexity and speeds up the training and reasoning process. However, in some cases, especially in scenes with complex backgrounds or dense targets, the target positioning accuracy may not be as good as the anchor box-based method.

[0003] In summary, anchor-based methods need to generate and process a large number of anchor boxes, which may lead to high training and inference costs on high-resolution images and large-scale datasets. Anchor-free methods cannot handle target occlusion problems well. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide a remote sensing image target detection method and system in view of the deficiencies in the above-mentioned prior art, so as to solve the technical problems that the existing target detection method has slow detection speed and cannot handle target occlusion well.

[0005] The objective of the present invention is achieved by the following technical solutions: In a first aspect, the present invention provides a remote sensing image target detection method, comprising: Acquire a remote sensing image to be detected; Inputting the remote sensing image to be detected into the trained target detection model, performing target detection on the remote sensing image, and obtaining a detection result; The target detection model includes a backbone network module, an adaptive occlusion perception module and a mapping module. The backbone network module is used to perform feature processing on the input remote sensing image to obtain a multi-scale feature map; the adaptive occlusion perception module is used to perform occlusion processing on the multi-scale feature map to obtain an image occlusion mask map; the features are optimized according to the image occlusion mask map; the mapping module is used to perform feature mapping to obtain mapping features, obtain the detection result of the occluded area according to the mapping features, obtain the detection result of the non-occluded area according to the features in the multi-scale feature map, and finally obtain the detection result according to the occluded area detection result and the non-occluded area detection result.

[0006] As a further improvement of the present invention, the backbone network module adopts a ResNet network; the backbone network module includes a backbone sub-module and a feature pyramid sub-module, the backbone sub-module performs feature extraction on the input remote sensing image to obtain a first feature map; the feature pyramid sub-module is used to perform convolution processing on the features in the first feature map to obtain a multi-scale feature map.

[0007] As a further improvement of the present invention, the adaptive occlusion perception module is used to perform convolution processing on the input multi-scale feature map, and then use the softmax function to process it to obtain a first result map, in which the pixel positions greater than the first set threshold in the first result map are set to 1, and the remaining pixel positions are set to 0, to obtain an image occlusion mask map; As a further improvement of the present invention, the mapping module is used to perform dot product processing on the feature map and the image occlusion mask map, hide the area where there is no occlusion, add the corresponding dot product result and the feature map and input it into the feature mapping layer to obtain the mapping feature; The mapping module is further provided with an occlusion detection module and a non-occlusion detection module; the occlusion detection module is used to convolve the mapping features to obtain the prediction results of the classification and the detection frame offset, and to perform positive sample determination according to the prediction results to obtain the detection results of the occlusion area; the detection frame is a preset anchor frame, which is used to add the offset result to obtain the corresponding detection result; The non-occlusion detection module is used to obtain the prediction results of classification, detection frame and center degree, perform positive sample determination according to the prediction results, and obtain the detection results of the non-occlusion area.

[0008] As a further improvement of the present invention, a detection module is further connected to the occlusion detection module and the non-occlusion detection module; the detection module respectively receives positive samples of all occlusion detection modules and non-occlusion detection modules, and after scale transformation, performs non-maximum suppression processing on the positive samples to obtain occlusion detection results and non-occlusion detection results; for the occlusion detection results, the IoU is calculated with the non-occlusion detection results, and the occlusion detection results with IoU greater than 0.7 are retained, and finally the retained occlusion detection results and the non-occlusion detection results are integrated as the final detection results.

[0009] As a further improvement of the present invention, the remote sensing image data is preprocessed, and the preprocessing step includes standardizing and normalizing the remote sensing image data.

[0010] As a further improvement of the present invention, after the remote sensing image data is preprocessed, the remote sensing image data is further enhanced, specifically including horizontal flipping and random cropping: The horizontal flipping specifically includes: flipping the remote sensing image along the vertical center, making the image symmetrical left and right, and mirror-reversing the annotation frame corresponding to the remote sensing image; The random cropping specifically includes: randomly cropping a sub-region in the original remote sensing image as a training sample, detecting whether the corresponding annotation box exceeds the image boundary through the training sample, and if so, aligning the annotation box to the edge of the image.

[0011] As a further improvement of the present invention, the target detection model sets different learning rates and weight decay parameters during the training process, sets the number of training rounds and hyperparameters; Adamw is used as the model optimizer and poly is used as the learning rate strategy; the DIOR remote sensing dataset is used to iteratively train the target inspection model; According to the value of the loss function, the corresponding parameters of the target detection model are updated through back propagation until the optimal model parameters are obtained.

[0012] As a further improvement of the present invention, the loss function is:

[0013] Where L is the loss function, represents the classification loss, represents the regression loss.

[0014] In a second aspect, the present invention provides a remote sensing image target detection system for implementing the above remote sensing image target detection method, comprising: A data acquisition unit, used for acquiring remote sensing images to be detected; The target detection unit is used to input the remote sensing image to be detected into the trained target detection model, perform target detection on the remote sensing image, and obtain a detection result; The target detection model includes a backbone network module, an adaptive occlusion perception module and a mapping module. The backbone network module is used to perform feature processing on the input remote sensing image to obtain a multi-scale feature map; the adaptive occlusion perception module is used to perform occlusion processing on the multi-scale feature map to obtain an image occlusion mask map; the features are optimized according to the image occlusion mask map; the mapping module is used to perform feature mapping to obtain mapping features, obtain the detection result of the occluded area according to the mapping features, obtain the detection result of the non-occluded area according to the features in the multi-scale feature map, and finally obtain the detection result according to the occluded area detection result and the non-occluded area detection result.

[0015] In a third aspect, the present invention provides a computer-readable storage medium storing one or more programs, wherein the one or more programs include instructions, which, when executed by a computing device, enable the computing device to perform the above-mentioned remote sensing image target detection method.

[0016] In a fourth aspect, the present invention provides a computing device, comprising: One or more processors, a memory and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, and the one or more programs include steps for executing the above-mentioned remote sensing image target detection method.

[0017] The beneficial effects of the present invention are as follows: the remote sensing image target detection method provided by the present invention realizes adaptive perception of the target occlusion problem through an adaptive occlusion perception module. By performing occlusion processing on the multi-scale feature map to obtain an image occlusion mask map, the occlusion situation in the image can be effectively perceived, so as to solve the occlusion problem in a targeted manner in the feature optimization process, reduce false detection and missed detection caused by occlusion, and improve the accuracy of target detection. At the same time, the adaptive occlusion perception module can automatically adjust the occlusion processing and alignment mark processing according to different remote sensing images, adapt to the occlusion situation in different scenes, and improve the adaptability of the model. By performing feature mapping and quickly obtaining the detection result through the mapping module, the calculation time can be reduced, the efficiency of target detection can be improved, and the needs of real-time or large-scale remote sensing image detection can be met. The multi-module structure of the model enables it to comprehensively utilize the advantages of different modules for processing when facing complex remote sensing image data, enhances the robustness of the model, and can stably perform target detection under various environmental conditions. The target detection model of the present invention effectively takes into account both the computational complexity and the occlusion problem. Improve the efficiency and accuracy of model detection.

[0018] Furthermore, the backbone network module adopts the ResNet network structure design which is relatively efficient. It can reduce the amount of calculation and the number of parameters while ensuring the effect of feature extraction. The target detection model can calculate more quickly when processing remote sensing images, improve the detection efficiency, especially for large-scale remote sensing image data sets, which can save a lot of computing time and resources. The convolution processing of the feature pyramid submodule usually adopts an efficient convolution algorithm, which further improves the computing efficiency and can process more remote sensing images in a limited time, meeting the real-time requirements in practical applications.

[0019] Furthermore, in the adaptive occlusion perception module, the multi-scale feature map is convolved and the first result map is obtained using the softmax function, which can effectively classify the features and estimate the probability. By setting the first set threshold and binarizing the pixel position to obtain the image occlusion mask map, the model can accurately identify the occluded area in the image, providing clear guidance for subsequent feature optimization processing. This precise occlusion mask generation helps to reduce target detection errors caused by occlusion and improve detection accuracy. The processing method using convolution and softmax functions can automatically learn the occlusion patterns in different images, adapt to various complex occlusion situations, and improve the generalization ability of the model. Processing the paired boxes whose values ​​in the IoU matrix are greater than the set value can effectively screen out target pairs that may have occlusion relationships and reduce unnecessary calculations and misjudgments.

[0020] Furthermore, by preprocessing the remote sensing image data input to the target detection model, we can prevent certain features from dominating the model training due to excessively large or small values. This allows the model to learn the importance of each feature more fairly and improve the accuracy and stability of target detection.

[0021] Furthermore, the remote sensing images are presented from different angles through horizontal flipping operations, which increases the diversity of the data. At the same time, the model learns more universal features and improves the model's ability to detect targets in different directions. After horizontal flipping and random cropping, the model will be exposed to a variety of image variants during training, thereby learning to be more robust to different image changes. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0023] Figure 1is a schematic diagram of a remote sensing image target detection method according to an embodiment of the present invention; Figure 2 is a schematic diagram of the structure of a target detection model in an embodiment of the present invention; Figure 3 is a schematic diagram of the structure of a remote sensing image target detection system in an embodiment of the present invention; Figure 4 It is a schematic diagram of the structure of an electronic device in an embodiment of the present invention. DETAILED DESCRIPTION

[0024] In order to make the purpose and technical solution of the present invention clearer and easier to understand, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. The specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0025] Explanation of relevant terms: Anchor-free: A method in the field of object detection that does not rely on a predefined anchor box (AnchorBox). It predicts the position and shape of the target directly on the image or feature map instead of the offset relative to the anchor box.

[0026] FPN (Feature Pyramid Network): Feature Pyramid Network; a network structure used for target detection. It aims to solve the multi-scale problem in target detection, that is, targets of different sizes need to be detected on feature maps of different scales.

[0027] ResNet (Residual Network): Residual network; a deep neural network structure, the core of which is to solve the gradient vanishing and gradient exploding problems in deep network training by introducing residual connections, so that the network can be trained deeper and have stronger performance.

[0028] ASS: Adaptive Occlusion Awareness Module; DIOR: A large-scale benchmark dataset for object detection in optical remote sensing images.

[0029] Threshold: threshold.

[0030] GIoU (Generalized Intersection over Union): Generalized intersection over union, an indicator used to measure the degree of overlap between two shapes (mainly the predicted box and the real box in target detection).

[0031] GTbox (GroundTruthbox): Ground Truthbox.

[0032] IoU: Intersection over Union function.

[0033] The technical solution of the present invention will be clearly and completely described below in conjunction with the accompanying drawings and specific embodiments, wherein the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments.

[0034] Example 1 like Figure 1 , Figure 2 As shown, this embodiment provides a remote sensing image target detection method, and the following is a specific implementation method.

[0035] Obtain a remote sensing image to be detected; input the remote sensing image to be detected into a trained target detection model, perform target detection on the remote sensing image, and obtain a detection result.

[0036] Among them, the data set used to train the target detection model in this embodiment is a remote sensing image data set of RGB three channels. In this embodiment, the DIOR data set is used as the basic data set, where the DIOR data set contains 23463 images and 192472 instances, covering 20 object classes. The construction process of this data set specifically includes: The image size of the dataset is 800×800 pixels, and the spatial resolution is 0.5m~30m. The dataset is divided into a training set (11,725 ​​images) and a test set (11,738 images).

[0037] The remote sensing images in the dataset are labeled images. The image labels are divided into 20 categories. The box annotation type uses a horizontal box, that is, the coordinates of the upper left corner and the lower right corner of the annotation box. The images and annotation files are stored in jpg and xml formats respectively.

[0038] The remote sensing images in the dataset are preprocessed images, and the specific preprocessing steps include standardization and normalization.

[0039] The standardization process specifically includes: scaling the input image data to a range according to the calculated mean and standard deviation so that it satisfies the normal distribution. The standardization formula is expressed as:

[0040] Here, μ and σ represent the mean and variance of the data set, respectively.

[0041] Normalization processing specifically includes: scaling the data so that it falls into a small specific interval. By default, the maximum and minimum normalization is used to scale it to [0,1]. The normalization formula is expressed as:

[0042] Among them, x is the original data, min(x) and max(x) represent the minimum and maximum values ​​of the graph respectively.

[0043] In addition, in order to highlight local feature learning and increase data diversity, this embodiment also performs data enhancement processing on the remote sensing image. The specific processing method includes horizontal flipping and random cropping processing: The horizontal flipping specifically includes: flipping the remote sensing image along the vertical center, mirroring it left and right, and at the same time, mirroring the annotation box while processing the image symmetrically.

[0044] Random cropping specifically includes: randomly cropping a sub-region in the original remote sensing image as a training sample, detecting whether the corresponding annotation box exceeds the image boundary, and if so, aligning the annotation box to the edge of the image.

[0045] Among them, the target detection model includes a backbone network module, an adaptive occlusion perception module and a mapping module.

[0046] The backbone network module is used to perform feature processing on the input remote sensing image to obtain a multi-scale feature map. The backbone network module uses the ResNet network as the network architecture. The backbone network module includes a backbone submodule and a feature pyramid submodule. The backbone submodule extracts features from the input remote sensing image to obtain a first feature map; the feature pyramid submodule is used to perform convolution processing on the features in the first feature map to obtain a multi-scale feature map.

[0047] The backbone submodule is constructed based on the ResNet network. The enhanced image is input into the backbone submodule, and the first feature map is obtained after four stages. The first feature map contains four layers of feature vectors: C1, C2, C3, and C4, where the first feature map is a multi-scale feature map.

[0048] Select the last three layers of features C2, C3, and C4 in the first feature map and input them into the pyramid submodule (FPN submodule) to obtain a multi-scale feature map, where the multi-scale feature map includes features P2, P3, P4, and P5. The specific formula is as follows:

[0049]

[0050]

[0051] in, is the network convolutional layer.

[0052] The adaptive occlusion perception module is used to perform convolution processing on the input multi-scale feature map, and then use the softmax function to process it to obtain a first result map, in which the pixel positions greater than the first set threshold are set to 1, and the remaining pixel positions are set to 0 to obtain an image occlusion mask map; The adaptive occlusion perception module is used to perform occlusion processing on the multi-scale feature map to obtain an image occlusion mask map; and optimize the features according to the image occlusion mask map; Specifically, the multi-scale feature maps P2, P3, P4, and P5 output by the backbone network module are input into the parallel adaptive occlusion perception module (ASS) to obtain the image occlusion mask maps M2, M3, M4, and M5.

[0053] Specifically, the parallel adaptive occlusion perception module uses parameter-independent 1×1 convolution for feature maps of different scales, and then processes them using the softmax function. Finally, the positions in the result map that are greater than the threshold are set to 1, and the other positions are set to 0, thus obtaining the required image occlusion mask map.

[0054] The mapping module (FMM) is used to perform dot product processing on the feature map and the image occlusion mask map, hide the area where there is no occlusion, add the corresponding dot product result and the feature map, and input it into the feature mapping layer to obtain the mapping feature, and then input the mapping feature into the subsequent module.

[0055] An occlusion detection module and a non-occlusion detection module are also provided after the mapping module. The occlusion detection module is used to convolve the mapping features output by the mapping module to obtain the prediction results of the classification and detection frame offset, and make a positive sample judgment based on the prediction results to obtain the detection results of the occluded area. The non-occlusion detection module is used to obtain the prediction results of the classification, detection frame and center according to the features in the multi-scale feature map, and make a positive sample judgment based on the prediction results to obtain the detection results of the non-occluded area. Among them, the weights of the occlusion detection module and the non-occlusion detection module are independent. Among them, the detection frame of the occlusion module is a preset anchor frame, which is used to add the offset result to obtain the corresponding detection result; and the non-occlusion module directly obtains the detection result.

[0056] In addition, the occlusion detection module and the non-occlusion detection module are connected to a detection module, which receives positive samples of all occlusion detection modules and non-occlusion detection modules respectively. After scale transformation, these results are subjected to non-maximum suppression processing to obtain occlusion detection results and non-occlusion detection results. For the occlusion detection results, the IoU is calculated with the non-occlusion detection results, and the occlusion results with IoU greater than 0.7 are retained. Finally, all the results are integrated as the final detection results.

[0057] Specifically, the feature mapping layer is composed of 3×3 convolution and group normalization operations stacked twice. In this embodiment, the stride of the 3×3 convolution is set to 1, and the element padding is set to 1.

[0058] For the occlusion detection module, this embodiment uses 1×1 convolution to obtain the prediction results of classification and detection box offset, where the number of classification result channels is 20, which is the number of categories in the data set. The number of detection box channels is 4×3, corresponding to the three preset aspect ratio anchor boxes. The anchor box is preset to an area of ​​256 pixels and uses aspect ratios of [1:2, 1:1, 2:1] respectively. The preset anchor frames are resized to the corresponding scales, and the prediction results of the anchor frame and box offset are added to obtain the prediction results of the occluded object detection frame. Finally, the prediction results of the occlusion module are judged as positive samples: if there is a real box (GTbox) and the IoU of the detection box is greater than the set value (for example, 0.7), it is regarded as a positive sample, otherwise it is regarded as a negative sample.

[0059] For the non-occlusion detection module, this embodiment uses parameter-independent 3×3 convolution to obtain the prediction results of classification, box and center respectively, where the stride of the 3×3 convolution is set to 1 and the pad is set to 1; the number of classification result channels is the number of pre-defined categories; the number of detection box channels is 4; and the number of center channels is 1. Finally, the prediction results of the non-occlusion detection module are judged as positive samples: if the point (x, y) is within the (c x -rs,c y -rs,c x +rs,c y +rs) range, it is considered a positive sample, otherwise it is considered a negative sample. x ,c y ) is the center point of the real box, s is the step size of the feature map relative to the original image, and r is a hyperparameter controlling the distance from the center of GT.

[0060] In the process of training the target detection function, this embodiment sets different learning rates and weight decay parameters according to the training results, sets the number of training rounds and hyperparameters, and the hyperparameters include parameters such as the number of categories and thresholds.

[0061] Adamw is used as the model optimizer, and poly is used as the learning rate strategy; the DIOR remote sensing dataset is used to iteratively train the target inspection model.

[0062] In this embodiment, the number of training rounds is set to 50 epochs, and the formula is as follows:

[0063] Among them, base_lr defaults to 6e-5 (i.e. 6×10 -5 ), num_epoch, power defaults to 0.9.

[0064] According to the value of the loss function, the corresponding parameters of the target detection model are updated through back propagation until the optimal model parameters are obtained.

[0065] Save the optimal model parameters. After each epoch, obtain the mIoU and mAP of each category of the model on the validation set, calculate their mean, and select the model parameters with the highest mAP index to save.

[0066] Among them, the loss function of the target detection model is:

[0067] Among them, L is the loss function, L cls represents the classification loss, L reg Represents regression loss. The loss function calculation formula is as follows:

[0068] Among them, L reg represents the regression loss, using the GIoU loss function, T i is the predicted value of the i-th box, T * i is the true value of the corresponding box. cls Represents classification loss, using binary cross entropy loss function, S i is the category prediction of the i-th box, S * i is the true value of the category of the corresponding box. When judging the corresponding box, we use the annotated box with the largest IoU with the predicted box as the corresponding box to match it.

[0069] In summary, this embodiment performs feature processing on remote sensing images through the backbone network module to obtain a multi-scale feature map, which can extract image features from different levels and angles, provide richer information for subsequent target detection, and help to identify targets more accurately. The adaptive occlusion perception module obtains an image occlusion mask map by performing occlusion processing on the multi-scale feature map, which can effectively perceive the occlusion situation in the image, thereby solving the occlusion problem in a targeted manner in the feature optimization process, reducing false detection and missed detection caused by occlusion, and improving the accuracy of target detection. The adaptive occlusion perception module can automatically adjust the occlusion processing and alignment mark processing according to different remote sensing images, adapt to the occlusion situation in different scenes, and improve the adaptability of the model. The mapping module performs feature mapping and quickly obtains the detection results, which can reduce the calculation time, improve the efficiency of target detection, and meet the needs of real-time or large-scale remote sensing image detection.

[0070] Example 2 like Figure 3 As shown, this embodiment provides a remote sensing image target detection system for implementing the remote sensing image target detection method described in Example 1. The system includes a data acquisition unit and a target detection unit, wherein the target detection unit includes a backbone network module, an adaptive occlusion perception module, and a mapping module; specifically: A data acquisition unit, used for acquiring remote sensing images to be detected; The target detection unit is used to input the remote sensing image to be detected into the trained target detection model, perform target detection on the remote sensing image, and obtain a detection result; The target detection model includes a backbone network module, an adaptive occlusion perception module and a mapping module. The backbone network module is used to perform feature processing on the input remote sensing image to obtain a multi-scale feature map; the adaptive occlusion perception module is used to perform occlusion processing on the multi-scale feature map to obtain an image occlusion mask map; the features are optimized according to the image occlusion mask map; the mapping module is used to perform feature mapping to obtain mapping features, obtain occlusion area detection results according to the mapping features, obtain non-occlusion area detection results according to the features in the multi-scale feature map, and finally obtain the detection results according to the occlusion area detection results and the non-occlusion area detection results.

[0071] The multi-module structure of the model enables it to comprehensively utilize the advantages of different modules to process complex remote sensing image data, enhances the robustness of the model, and enables stable target detection under various environmental conditions.

[0072] Example 3 In another embodiment of the present invention, the present invention also provides a storage medium, specifically a computer-readable storage medium (Memory), which is a memory device in a terminal device for storing programs and data. It is understandable that the computer-readable storage medium here can include both the built-in storage medium in the terminal device and the extended storage medium supported by the terminal device, and can be any tangible medium containing or storing a program, which can be used by an instruction execution system, device or device or used in combination with it. The computer-readable storage medium provides a storage space, which stores the operating system of the terminal. In addition, one or more instructions suitable for being loaded and executed by the processor are also stored in the storage space, and these instructions can be one or more computer programs (including program codes). It should be noted that more specific examples (non-exhaustive list) of the computer-readable storage medium here include: an electrical connection with one or more wires, a portable disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above.

[0073] Computer readable storage media also include data signals propagated in baseband or as part of a carrier wave, which carry readable program codes. Such propagated data signals can take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. The readable storage medium can also be any readable medium other than a readable storage medium, which can send, propagate, or transmit a program for use by or in conjunction with an instruction execution system, device, or device. The program code contained on the readable storage medium can be transmitted using any appropriate medium, including but not limited to wireless, wired, optical cable, RF, etc., or any suitable combination of the above.

[0074] Program code for performing the operations of the present invention may be written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Java, C++, etc., and conventional procedural programming languages ​​such as "C" or similar programming languages. The program code may be executed entirely on the user computing device, partially on the user device, as a separate software package, partially on the user computing device and partially on a remote computing device, or entirely on a remote computing device or server. In the case of a remote computing device, the remote computing device may be connected to the user computing device through any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computing device (e.g., through the Internet using an Internet service provider).

[0075] The processor may load and execute one or more instructions stored in the computer-readable storage medium to implement the corresponding steps of the remote sensing image target detection method in the above embodiment; the processor may load and execute the following steps: Acquire a remote sensing image to be detected; Inputting the remote sensing image to be detected into the trained target detection model, performing target detection on the remote sensing image, and obtaining a detection result; The target detection model includes a backbone network module, an adaptive occlusion perception module and a mapping module. The backbone network module is used to perform feature processing on the input remote sensing image to obtain a multi-scale feature map; the adaptive occlusion perception module is used to perform occlusion processing on the multi-scale feature map to obtain an image occlusion mask map; the features are optimized according to the image occlusion mask map; the mapping module is used to perform feature mapping to obtain mapping features, obtain occlusion area detection results according to the mapping features, obtain non-occlusion area detection results according to the features in the multi-scale feature map, and finally obtain the detection results according to the occlusion area detection results and the non-occlusion area detection results.

[0076] Figure 4 A schematic diagram of a computer device provided by an embodiment of the present invention.

[0077] See also Figure 4 , the terminal device is a computer device, and the computer device 60 of this embodiment includes: a processor 61, a memory 62, and a computer program 63 stored in the memory 62 and executable on the processor 61. When the computer program 63 is executed by the processor 61, the remote sensing image target detection method in the embodiment is implemented. To avoid repetition, it is not described one by one here. Alternatively, when the computer program 63 is executed by the processor 61, the functions of each model / unit in Embodiment 1 are implemented. To avoid repetition, it is not described one by one here.

[0078] The computer device 60 may be a computing device such as a desktop computer, a notebook, a PDA, or a cloud server. The computer device 60 may include, but is not limited to, a processor 61 and a memory 62. Those skilled in the art will appreciate that Figure 4 This is only an example of the computer device 60 and does not constitute a limitation of the computer device 60. It may include more or fewer components than shown in the figure, or a combination of certain components, or different components. For example, the computer device may also include input and output devices, network access devices, buses, etc.

[0079] The processor 61 may be a central processing unit (CPU), other general-purpose processors, central processing units, graphics processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, data processing logic based on quantum computing, discrete hardware components, etc. A general-purpose processor may be a microprocessor or the processor may be any conventional processor, etc.

[0080] The memory 62 may be an internal storage unit of the computer device 60, such as a hard disk or memory of the computer device 60. The memory 62 may also be an external storage device of the computer device 60, such as a plug-in hard disk, a smart memory card (SmartMediaCard, SMC), a secure digital (SecureDigital, SD) card, a flash card (FlashCard), etc. equipped on the computer device 60.

[0081] Furthermore, the memory 62 may include both an internal storage unit of the computer device 60 and an external storage device. The memory 62 is used to store computer programs and other programs and data required by the computer device. The memory 62 may also be used to temporarily store data that has been output or is to be output.

[0082] Any reference to a memory, database or other medium used in the embodiments provided in this application may include at least one of a non-volatile and a volatile memory. Non-volatile memory may include a read-only memory (ROM), a magnetic tape, a floppy disk, a flash memory, an optical memory, a high-density embedded non-volatile memory, a resistive random access memory (ReRAM), a magnetic random access memory (MRAM), a ferroelectric random access memory (FRAM), a phase change memory (PCM), a graphene memory, etc. Volatile memory may include a random access memory (RAM) or an external cache memory, etc. As an illustration and not limitation, RAM may be in various forms, such as a static random access memory (SRAM) or a dynamic random access memory (DRAM), etc.

[0083] The database involved in each embodiment provided in this application may include at least one of a relational database and a non-relational database. The non-relational database may include a distributed database based on blockchain, etc., but is not limited thereto. The processor involved in each embodiment provided in this application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., but is not limited thereto.

Claims

1. A remote sensing image target detection method, characterized in that: include: Acquire a remote sensing image to be detected; Inputting the remote sensing image to be detected into the trained target detection model, performing target detection on the remote sensing image, and obtaining a detection result; The target detection model includes a backbone network module, an adaptive occlusion perception module and a mapping module. The backbone network module is used to perform feature processing on the input remote sensing image to obtain a multi-scale feature map; the adaptive occlusion perception module is used to perform occlusion processing on the multi-scale feature map to obtain an image occlusion mask map; The features are optimized according to the image occlusion mask map; the mapping module is used to perform feature mapping to obtain mapping features, obtain occlusion area detection results according to the mapping features, obtain non-occlusion area detection results according to the features in the multi-scale feature map, and finally obtain the detection results according to the occlusion area detection results and the non-occlusion area detection results.

2. The remote sensing image target detection method according to claim 1, characterized in that: The backbone network module adopts a ResNet network; the backbone network module includes a backbone submodule and a feature pyramid submodule, the backbone submodule performs feature extraction on the input remote sensing image to obtain a first feature map; the feature pyramid submodule is used to perform convolution processing on the features in the first feature map to obtain a multi-scale feature map.

3. The remote sensing image target detection method according to claim 1, characterized in that: The adaptive occlusion perception module is used to perform convolution processing on the input multi-scale feature map, and then use the softmax function to process it to obtain a first result map, in which the pixel positions greater than the first set threshold in the first result map are set to 1, and the remaining pixel positions are set to 0 to obtain an image occlusion mask map.

4. The remote sensing image target detection method according to claim 1, characterized in that: The mapping module is used to perform dot product processing on the feature map and the image occlusion mask map, hide the area where there is no occlusion, add the corresponding dot product result and the feature map, and input it into the feature mapping layer to obtain the mapping feature; The mapping module is further provided with an occlusion detection module and a non-occlusion detection module; the occlusion detection module is used to convolve the mapping features to obtain the prediction results of the classification and the detection frame offset, and to perform positive sample determination according to the prediction results to obtain the detection results of the occlusion area; the detection frame is a preset anchor frame, which is used to add the offset result to obtain the corresponding detection result; The non-occlusion detection module is used to obtain the prediction results of classification, detection frame and center degree, perform positive sample determination according to the prediction results, and obtain the detection results of the non-occlusion area.

5. The remote sensing image target detection method according to claim 4, characterized in that: A detection module is further connected to the occlusion detection module and the non-occlusion detection module; the detection module receives positive samples of all occlusion detection modules and non-occlusion detection modules respectively, and after scale transformation, performs non-maximum suppression processing on the positive samples respectively to obtain occlusion detection results and non-occlusion detection results; For the occlusion detection results, the IoU is calculated with the non-occlusion detection results, and the occlusion detection results with IoU greater than 0.7 are retained. Finally, the retained occlusion detection results and the non-occlusion detection results are integrated as the final detection results.

6. The remote sensing image target detection method according to claim 1, characterized in that: The remote sensing image data is preprocessed, and the preprocessing step includes standardizing and normalizing the remote sensing image data.

7. The remote sensing image target detection method according to claim 6, characterized in that: After the remote sensing image data is preprocessed, the remote sensing image data is further enhanced, specifically including horizontal flipping and random cropping: The horizontal flipping specifically includes: flipping the remote sensing image along the vertical center, making the image symmetrical left and right, and mirror-reversing the annotation frame corresponding to the remote sensing image; The random cropping specifically includes: randomly cropping a sub-region in the original remote sensing image as a training sample, detecting whether the corresponding annotation box exceeds the image boundary through the training sample, and if so, aligning the annotation box to the edge of the image.

8. The remote sensing image target detection method according to any one of claims 1 to 5, characterized in that: The target detection model sets different learning rates and weight decay parameters during the training process, sets the number of training rounds and hyperparameters; Adamw is used as the model optimizer and poly is used as the learning rate strategy; the DIOR remote sensing dataset is used to iteratively train the target inspection model; According to the value of the loss function, the corresponding parameters of the target detection model are updated through back propagation until the optimal model parameters are obtained.

9. The remote sensing image target detection method according to claim 8, characterized in that: The loss function is: Where L is the loss function, represents the classification loss, represents the regression loss.

10. A remote sensing image target detection system, used to implement the remote sensing image target detection method according to any one of claims 1 to 9, characterized in that: include: A data acquisition unit, used for acquiring remote sensing images to be detected; The target detection unit is used to input the remote sensing image to be detected into the trained target detection model, perform target detection on the remote sensing image, and obtain a detection result; The target detection model includes a backbone network module, an adaptive occlusion perception module and a mapping module. The backbone network module is used to perform feature processing on the input remote sensing image to obtain a multi-scale feature map; the adaptive occlusion perception module is used to perform occlusion processing on the multi-scale feature map to obtain an image occlusion mask map; the features are optimized according to the image occlusion mask map; the mapping module is used to perform feature mapping to obtain mapping features, obtain the detection result of the occluded area according to the mapping features, obtain the detection result of the non-occluded area according to the features in the multi-scale feature map, and finally obtain the detection result according to the occluded area detection result and the non-occluded area detection result.

Images