Remote sensing image target detection method, system and equipment based on deep learning

Abstract

The present invention relates to the field of image analysis technology, and specifically to a remote sensing image target detection method, system and device based on deep learning. In order to solve the problem of low accuracy of remote sensing image target detection results in the prior art, the present invention first performs low-level semantic information aggregation processing on different depth features of the remote sensing image to be detected, and then performs cross-low-level semantic injection processing, and then performs high-level semantic information aggregation processing and cross-high-level semantic injection processing to obtain low-level target aggregation features, intermediate target aggregation features and high-level target aggregation features; then, the low-level target aggregation features, intermediate target aggregation features and high-level target aggregation features are subjected to linear mapping and full connection processing to separate small targets from the background, and finally, bounding box regression, bounding box classification recognition and non-maximum suppression processing are performed to obtain a target detection result with high accuracy.

Description

Technical Field

[0001] The present invention relates to the field of image analysis technology, and specifically to a remote sensing image target detection method, system, and device based on deep learning. Background Art

[0002] In recent years, the field of drones has developed rapidly. Remote sensing image target detection is one of the key technologies in the field of drones. In order to promote the development of drone technology, it is necessary to develop an efficient remote sensing image target detection algorithm.

[0003] Currently, target detection is often performed based on the YOLO series of models. However, when it comes to remote sensing images, YOLO always faces two major challenges: multi-scale targets and complex background information. This is because the scale of targets in remote sensing images is much smaller than that of targets in other detection tasks, and there are also background factors such as shadows and occlusions around the targets in remote sensing images. As a result, when using the YOLO series of models for remote sensing image target detection, the detection results of small targets are inaccurate and are prone to missed detections and false detections. Summary of the Invention

[0004] The purpose of the present invention is to provide a method, system and device for remote sensing image target detection based on deep learning.

[0005] The technical solutions of the present invention are as follows:

[0006] A remote sensing image target detection method based on deep learning includes the following operations:

[0007] S1. The remote sensing image to be inspected is subjected to image enhancement processing to obtain a remote sensing enhanced image; the remote sensing enhanced image is subjected to small target feature extraction processing once, twice, three times, and four times respectively to obtain shallow remote sensing features, middle remote sensing features, middle-deep remote sensing features, and deep remote sensing features; the small target feature extraction processing operation is specifically as follows: the input is subjected to downsampling and convolution processing to obtain convolution features; the convolution features are subjected to structural feature extraction and multi-scale texture enhancement respectively, and then spliced to obtain output;

[0008] S2, shallow remote sensing features, mid-level remote sensing features, mid-deep remote sensing features and deep remote sensing features are aggregated through low-level semantic information to obtain first semantic aggregate features; deep remote sensing features are processed through channel dimension reduction to obtain deep semantic features; deep semantic features, mid-deep remote sensing features, mid-level remote sensing features and first semantic aggregate features are injected through low-level semantics to obtain first low-level injected features; first low-level injected features, mid-level remote sensing features, shallow remote sensing features and first semantic aggregate features are injected through low-level semantics to obtain second low-level injected features; first low-level injected features, second low-level injected features and deep semantic features are aggregated through high-level semantic information to obtain second semantic aggregate features; second low-level injected features are processed through channel dimension reduction to obtain low-level target aggregate features; second semantic aggregate features, first low-level injected features and second low-level injected features are injected through high-level semantics to obtain mid-level target aggregate features; second semantic aggregate features, mid-level target aggregate features and deep semantic features are injected through high-level semantics to obtain high-level target aggregate features;

[0009] S3, low-level target aggregation features, intermediate target aggregation features and high-level target aggregation features are processed by linear mapping and full connection, and then bounding box regression, bounding box classification recognition and non-maximum suppression are performed in sequence to obtain the target detection results.

[0010] The specific operations of the multi-scale texture enhancement processing in S1 are as follows: the convolution features are respectively subjected to multi-scale convolution feature extraction and spatial context feature extraction to obtain multi-scale convolution features and spatial context features; the multi-scale convolution features and spatial context features are multiplied element by element, and then added element by element with the multi-scale convolution features to obtain initial texture enhancement features; the initial texture enhancement features are subjected to convolution processing to obtain multi-scale texture enhancement features.

[0011] The operation of spatial context feature extraction can be achieved through average pooling, convolution, several depth-separable convolutions, convolution and Sigmoid function.

[0012] The specific operations of S2 low-level semantic information aggregation processing are as follows: shallow remote sensing features and middle-level remote sensing features are respectively processed by average pooling to obtain shallow pooling features and middle-level pooling features; deep remote sensing features are processed by bilinear mapping, and then spliced with middle-deep remote sensing features, shallow pooling features and middle-level pooling features based on channel connection to obtain initial aggregation features; the initial aggregation features are processed by convolution, structural parameterization and convolution to obtain the first semantic aggregation features.

[0013] The specific operations for obtaining the first low-level injection feature in S2 are as follows: the deep semantic features, the mid-deep remote sensing features, and the mid-level remote sensing features are respectively processed by bilinear mapping, convolution, and average pooling, and then subjected to channel connection-based splicing and convolution to obtain the first local injection feature; the first semantic aggregation feature is processed by convolution, Sigmoid function, and average pooling, and then multiplied element-wise with the first local injection feature to obtain the first initial injection feature; the first semantic aggregation feature is convolved and average pooled, and then added element-wise with the first initial injection feature to obtain the first low-level injection feature.

[0014] The specific operations of high-level semantic information aggregation processing in S2 are as follows: the first low-level injection feature and the second low-level injection feature are respectively average-pooled, and then spliced and convolved with the deep semantic feature based on channel connection to obtain the first high-level semantic feature; the first high-level semantic feature is processed by depthwise separable convolution, and then superimposed with the first high-level semantic feature, and then batch channel normalization is performed to obtain the second high-level semantic feature; the second high-level semantic feature is processed by the feedforward neural network, and then superimposed with the second high-level semantic feature, and then layer normalization is performed to obtain the second semantic aggregation feature.

[0015] The specific operations for obtaining the intermediate target aggregate feature in S2 are as follows: the first low-level injection feature and the second low-level injection feature are respectively subjected to convolution and average pooling, and then subjected to channel connection-based splicing and convolution to obtain the second local injection feature; the second semantic aggregate feature is subjected to convolution, Sigmoid function processing and average pooling, and then multiplied element-wise with the second local injection feature to obtain the second initial injection feature; the second semantic aggregate feature is subjected to convolution and average pooling, and then added element-wise with the second initial injection feature to obtain the intermediate target aggregate feature.

[0016] A remote sensing image target detection system based on deep learning, used to implement the above-mentioned remote sensing image target detection method based on deep learning, comprising:

[0017] The multi-level remote sensing feature generation module is used to perform image enhancement processing on the remote sensing image to be inspected to obtain a remote sensing enhanced image; the remote sensing enhanced image is subjected to small target feature extraction processing once, twice, three times, and four times respectively to obtain shallow remote sensing features, medium remote sensing features, medium-deep remote sensing features, and deep remote sensing features; the small target feature extraction processing operation is specifically as follows: the input is subjected to downsampling and convolution processing to obtain convolution features; the convolution features are subjected to structural feature extraction and multi-scale texture enhancement respectively, and then spliced to obtain the output;

[0018] The multi-target aggregate feature generation module is used to aggregate shallow remote sensing features, mid-layer remote sensing features, mid-deep remote sensing features and deep remote sensing features through low-level semantic information to obtain first semantic aggregate features; deep remote sensing features are processed through channel dimension reduction to obtain deep semantic features; deep semantic features, mid-deep remote sensing features, mid-layer remote sensing features and first semantic aggregate features are processed through low-level semantic injection to obtain first low-level injection features; first low-level injection features, mid-layer remote sensing features, shallow remote sensing features and first semantic aggregate features are processed through low-level semantic injection to obtain second low-level injection features; first low-level injection features, second low-level injection features and deep semantic features are aggregated through high-level semantic information to obtain second semantic aggregate features; second low-level injection features are processed through channel dimension reduction to obtain low-level target aggregate features; second semantic aggregate features, first low-level injection features and second low-level injection features are processed through high-level semantic injection to obtain mid-level target aggregate features; second semantic aggregate features, mid-level target aggregate features and deep semantic features are processed through high-level semantic injection to obtain high-level target aggregate features;

[0019] The target detection result generation module is used to perform bounding box regression, bounding box classification recognition and non-maximum suppression processing on low-level target aggregation features, intermediate target aggregation features and high-level target aggregation features after linear mapping and full connection processing to obtain target detection results.

[0020] A remote sensing image target detection device based on deep learning includes a processor and a memory, wherein the processor implements the above-mentioned remote sensing image target detection method based on deep learning when executing a computer program stored in the memory.

[0021] A computer-readable storage medium for storing a computer program, wherein the computer program, when executed by a processor, implements the above-mentioned remote sensing image target detection method based on deep learning.

[0022] The beneficial effects of the present invention are:

[0023] The present invention provides a remote sensing image target detection method based on deep learning. First, the remote sensing image to be detected is subjected to image enhancement processing to improve the feature expression ability of the remote sensing image to be detected. Then, the remote sensing enhanced image is subjected to several small target feature extraction processings to obtain different depth features of the remote sensing enhanced image, and shallow remote sensing features, middle remote sensing features, middle-deep remote sensing features and deep remote sensing features with different local feature detail highlighting intensities are obtained. Then, the shallow remote sensing features, middle remote sensing features, middle-deep remote sensing features and deep remote sensing features are subjected to low-level semantic information aggregation processing and then to cross-low-level semantic injection processing to highlight local color differences and small contour details in the remote sensing image to obtain first low-level injection features and second low-level injection features. Then, the first low-level injection features and the second low-level injection features are injected into the remote sensing image. After the injected features and deep semantic features are subjected to high-level semantic information aggregation processing, high-level semantic injection processing is performed. On the basis of detail enhancement, the global features of the remote sensing image are improved and the difference between background information and small target details is enhanced to obtain intermediate target aggregation features and high-level target aggregation features; then the low-level target aggregation features, intermediate target aggregation features and high-level target aggregation features obtained based on the first low-level injected features are subjected to linear mapping and full connection processing, so as to better separate the small target from the background, and finally the target detection results are obtained by bounding box regression, bounding box classification recognition and non-maximum suppression processing in sequence; this detection method is applied to target detection in remote sensing images, which can effectively separate small target features from the background, thereby improving the accuracy of small target detection results and avoiding missed detection and false detection. DETAILED DESCRIPTION

[0024] This embodiment provides a method for remote sensing image target detection based on deep learning, including the following operations:

[0025] S1. The remote sensing image to be inspected is subjected to image enhancement processing to obtain a remote sensing enhanced image; the remote sensing enhanced image is subjected to small target feature extraction processing once, twice, three times, and four times respectively to obtain shallow remote sensing features, middle remote sensing features, middle-deep remote sensing features, and deep remote sensing features; the small target feature extraction processing operation is specifically as follows: the input is subjected to downsampling and convolution processing to obtain convolution features; the convolution features are subjected to structural feature extraction and multi-scale texture enhancement respectively, and then spliced to obtain output;

[0026] S2, shallow remote sensing features, mid-level remote sensing features, mid-deep remote sensing features and deep remote sensing features are aggregated through low-level semantic information to obtain first semantic aggregate features; deep remote sensing features are processed through channel dimension reduction to obtain deep semantic features; deep semantic features, mid-deep remote sensing features, mid-level remote sensing features and first semantic aggregate features are injected through low-level semantics to obtain first low-level injected features; first low-level injected features, mid-level remote sensing features, shallow remote sensing features and first semantic aggregate features are injected through low-level semantics to obtain second low-level injected features; first low-level injected features, second low-level injected features and deep semantic features are aggregated through high-level semantic information to obtain second semantic aggregate features; second low-level injected features are processed through channel dimension reduction to obtain low-level target aggregate features; second semantic aggregate features, first low-level injected features and second low-level injected features are injected through high-level semantics to obtain mid-level target aggregate features; second semantic aggregate features, mid-level target aggregate features and deep semantic features are injected through high-level semantics to obtain high-level target aggregate features;

[0027] S3, low-level target aggregation features, intermediate target aggregation features and high-level target aggregation features are processed by linear mapping and full connection, and then bounding box regression, bounding box classification recognition and non-maximum suppression are performed in sequence to obtain the target detection results.

[0028] S1. The remote sensing image to be inspected is subjected to image enhancement processing to obtain a remote sensing enhanced image; the remote sensing enhanced image is subjected to small target feature extraction processing once, twice, three times and four times respectively to obtain shallow remote sensing features, middle remote sensing features, middle-deep remote sensing features and deep remote sensing features.

[0029] The remote sensing image to be inspected is subjected to image enhancement processing to improve the feature expression ability of the remote sensing image to be inspected, and then the remote sensing enhanced image is subjected to several small target feature extraction processes to obtain different depth features of the remote sensing enhanced image, and shallow remote sensing features, middle remote sensing features, middle-deep remote sensing features and deep remote sensing features with different intensities of small target feature detail prominence are obtained, thereby obtaining rich feature information of the remote sensing image, which is beneficial to improving the accuracy of small target detection in subsequent remote sensing images.

[0030] First, to improve the feature expression capability of the remote sensing image to be inspected and increase the efficiency of subsequent processing, the remote sensing image to be inspected is subjected to image enhancement processing to obtain a remote sensing enhanced image. The image enhancement processing operation can be implemented in sequence through multi-scale depthwise separable convolution (which can be achieved by concatenating depthwise separable convolutions of different scales), batch normalization, GELU activation function, depthwise separable convolution, and batch normalization.

[0031] Then, in order to obtain different depth features of the remote sensing enhanced image and obtain rich feature information, which is conducive to improving the subsequent ability to obtain small targets in the remote sensing enhanced image and facilitating the capture of detailed features of the remote sensing enhanced image, the remote sensing enhanced image is processed by small target feature extraction once, twice, three times and four times respectively to obtain shallow remote sensing features, middle remote sensing features, middle-deep remote sensing features and deep remote sensing features.

[0032] The small target feature extraction process is as follows: the input is downsampled and convolved to obtain convolution features; the convolution features are subjected to structural feature extraction and multi-scale texture enhancement, respectively, to extract global structural features and texture detail features from the remote sensing image; and then a channel-connected splicing process is performed to enhance local small target features and obtain output. The input is a remote sensing enhanced image, or shallow remote sensing features, or mid-layer remote sensing features, or mid-deep remote sensing features, and the output is shallow remote sensing features, or mid-layer remote sensing features, or mid-deep remote sensing features, or deep remote sensing features.

[0033] The above structural feature extraction can be achieved through full connection, ReLu activation function and full connection layer in sequence.

[0034] The above-mentioned multi-scale texture enhancement processing operations are specifically as follows: the convolution features are respectively subjected to multi-scale convolution feature extraction and spatial context feature extraction to obtain multi-scale convolution features and spatial context features; the multi-scale convolution features and spatial context features are element-wise multiplied and then added element-by-element with the multi-scale convolution features to obtain initial texture enhancement features; the initial texture enhancement features are subjected to convolution processing to obtain multi-scale texture enhancement features.

[0035] The spatial context feature extraction operation can be implemented through average pooling, convolution, several depthwise separable convolutions, convolution, and sigmoid functions to extract spatial distance information from remote sensing images, which helps improve the accuracy of subsequent detection and recognition. The multi-scale convolution feature extraction operation is as follows: the convolution feature is processed by depthwise separable convolution (kernel is 3×3) to obtain the first depthwise separable convolution feature; the first depthwise separable convolution feature is processed by depthwise separable convolution of different scales (kernels are 3×3, 5×5, 7×7, 9×9, and 11×11), and then element-wise added to the first depthwise separable convolution feature, and then convolved to obtain the multi-scale convolution feature.

[0036] S2, shallow remote sensing features, mid-level remote sensing features, mid-deep remote sensing features and deep remote sensing features are aggregated with low-level semantic information to obtain the first semantic aggregate feature; deep remote sensing features are processed by channel dimension reduction to obtain deep semantic features; deep semantic features, mid-deep remote sensing features, mid-level remote sensing features and first semantic aggregate features are injected with low-level semantics to obtain the first low-level injected feature; the first low-level injected feature, mid-level remote sensing features, shallow remote sensing features and first semantic aggregate features are injected with low-level semantics to obtain the second low-level injected feature; the first low-level injected feature, the second low-level injected feature and deep semantic features are aggregated with high-level semantic information to obtain the second semantic aggregate feature; the second low-level injected feature is processed by channel dimension reduction to obtain the low-level target aggregate feature; the second semantic aggregate feature, the first low-level injected feature and the second low-level injected feature are injected with high-level semantics to obtain the mid-level target aggregate feature; the second semantic aggregate feature, the mid-level target aggregate feature and deep semantic features are injected with high-level semantics to obtain the high-level target aggregate feature.

[0037] After low-level semantic information aggregation processing of shallow remote sensing features, mid-level remote sensing features, mid-deep remote sensing features and deep remote sensing features, cross-low-level semantic injection processing is performed to achieve detail enhancement in different aspects, highlight local color differences and small contour details in remote sensing images, and obtain the first low-level injection feature and the second low-level injection feature; and then the first low-level injection feature, the second low-level injection feature and the deep semantic feature are subjected to high-level semantic information aggregation processing and then high-level semantic injection processing, on the basis of detail enhancement, the global features of the remote sensing image are improved and the difference between background information and small target details is enhanced, and medium-level target aggregation features and high-level target aggregation features are obtained.

[0038] First, shallow remote sensing features, mid-layer remote sensing features, mid-deep remote sensing features and deep remote sensing features are processed by low-level semantic information aggregation to achieve detail enhancement and obtain the first semantic aggregation feature.

[0039] The specific operations of low-level semantic information aggregation processing are as follows: shallow remote sensing features and middle-level remote sensing features are respectively processed by average pooling to obtain shallow pooling features and middle-level pooling features; deep remote sensing features are processed by bilinear mapping, and then spliced with middle-deep remote sensing features, shallow pooling features and middle-level pooling features based on channel connection to obtain initial aggregation features; the initial aggregation features are processed by convolution, structural parameterization (which can be achieved through the RepConv module) and convolution to obtain the first semantic aggregation features.

[0040] Then, the deep semantic features, mid-deep remote sensing features, mid-level remote sensing features, and first semantic aggregation features are processed with low-level semantic injection to highlight details such as local color differences in the remote sensing image, thereby obtaining the first low-level injection features. Specifically, the deep semantic features, mid-deep remote sensing features, and mid-level remote sensing features are processed with bilinear mapping, convolution, and average pooling, and then spliced and convolved based on channel connection to obtain the first local injection features. The first semantic aggregation features are processed with convolution, Sigmoid function, and average pooling, and then multiplied element-wise with the first local injection features to obtain the first initial injection features. The first semantic aggregation features are convolved and average pooled, and then added element-wise with the first initial injection features to obtain the first low-level injection features.

[0041] At the same time, the deep remote sensing features are processed by channel dimension reduction to obtain deep semantic features that can reflect the global characteristics of the remote sensing image.

[0042] Next, the first low-level injection feature is processed with the mid-level remote sensing feature, the shallow remote sensing feature, and the first semantic aggregation feature through low-level semantic injection to highlight the detailed features such as tiny contours in the remote sensing image, thereby obtaining the second low-level injection feature. Specifically, the first low-level injection feature, the mid-level remote sensing feature, and the shallow remote sensing feature are processed through bilinear mapping, convolution, and average pooling, and then through channel-connection-based splicing and convolution to obtain the second-first local injection feature; the first semantic aggregation feature is processed through convolution, Sigmoid function, and average pooling, and then multiplied element-wise with the second-first local injection feature to obtain the second-first initial injection feature; the first semantic aggregation feature is convolved and average pooled, and then added element-wise with the second-first initial injection feature to obtain the second low-level injection feature.

[0043] Subsequently, the first low-level injection features, the second low-level injection features and the deep semantic features are aggregated through high-level semantic information to enhance the global features of the remote sensing image on the basis of detail enhancement, and obtain the second semantic aggregated features.

[0044] Among them, the operation of high-level semantic information aggregation processing is specifically as follows: the first low-level injection feature and the second low-level injection feature are respectively average pooled, and then spliced and convolved with the deep semantic feature based on channel connection to obtain the first high-level semantic feature; the first high-level semantic feature is processed by depth-separable convolution, and then superimposed with the first high-level semantic feature, and then batch channel normalization is performed to obtain the second high-level semantic feature; the second high-level semantic feature is processed by the feedforward neural network, and then superimposed with the second high-level semantic feature, and then layer normalization is performed to obtain the second semantic aggregation feature.

[0045] At the same time, the second low-level injection features are processed by channel dimension reduction to obtain low-level target aggregation features.

[0046] Next, the second semantic aggregate feature is processed with the first and second low-level injection features through high-level semantic injection to enhance the differences between background information and details, thereby obtaining a mid-level target aggregate feature. Specifically, the first and second low-level injection features are convolved and average pooled, respectively, and then spliced and convolved based on channel connection to obtain the second local injection feature; the second semantic aggregate feature is convolved, Sigmoid function processed, and average pooled, and then element-wise multiplied with the second local injection feature to obtain the second initial injection feature; the second semantic aggregate feature is convolved and average pooled, and then element-wise added with the second initial injection feature to obtain the mid-level target aggregate feature.

[0047] Finally, the second semantic aggregate feature, the intermediate target aggregate feature, and the deep semantic feature are processed with high-level semantic injection. By further improving the global features, the difference between background information and small target details is enhanced to obtain the high-level target aggregate feature. Specifically, the intermediate target aggregate feature and the deep semantic feature are respectively convolved and average pooled, and then spliced and convolved based on channel connection to obtain the second local injection feature; the second semantic aggregate feature is convolved, processed with the Sigmoid function, and average pooled, and then multiplied element-wise with the second local injection feature to obtain the second initial injection feature; the second semantic aggregate feature is convolved and average pooled, and then added element-wise with the second initial injection feature to obtain the high-level target aggregate feature.

[0048] S3, low-level target aggregation features, intermediate target aggregation features and high-level target aggregation features are processed by linear mapping and full connection, and then bounding box regression, bounding box classification recognition and non-maximum suppression are performed in sequence to obtain the target detection results.

[0049] The low-level target aggregation features, intermediate target aggregation features and high-level target aggregation features are processed by linear mapping and full connection to better separate small targets from the background and obtain target remote sensing image features; the target remote sensing image features are processed in sequence by bounding box regression, bounding box classification recognition and non-maximum suppression to obtain a target detection map containing only one prediction box in the same area. The category corresponding to the maximum classification probability value on the prediction box corresponding to each area in the target detection map is taken as the recognition result of the corresponding area to obtain the target detection result.

[0050] This embodiment further provides a remote sensing image target detection system based on deep learning, which is used to implement the above-mentioned remote sensing image target detection method based on deep learning, including:

[0051] The multi-level remote sensing feature generation module is used to perform image enhancement processing on the remote sensing image to be inspected to obtain a remote sensing enhanced image; the remote sensing enhanced image is subjected to small target feature extraction processing once, twice, three times, and four times respectively to obtain shallow remote sensing features, medium remote sensing features, medium-deep remote sensing features, and deep remote sensing features; the small target feature extraction processing operation is specifically as follows: the input is subjected to downsampling and convolution processing to obtain convolution features; the convolution features are subjected to structural feature extraction and multi-scale texture enhancement respectively, and then spliced to obtain the output;

[0052] The multi-target aggregate feature generation module is used to aggregate shallow remote sensing features, mid-layer remote sensing features, mid-deep remote sensing features and deep remote sensing features through low-level semantic information to obtain first semantic aggregate features; deep remote sensing features are processed through channel dimension reduction to obtain deep semantic features; deep semantic features, mid-deep remote sensing features, mid-layer remote sensing features and first semantic aggregate features are processed through low-level semantic injection to obtain first low-level injection features; first low-level injection features, mid-layer remote sensing features, shallow remote sensing features and first semantic aggregate features are processed through low-level semantic injection to obtain second low-level injection features; first low-level injection features, second low-level injection features and deep semantic features are aggregated through high-level semantic information to obtain second semantic aggregate features; second low-level injection features are processed through channel dimension reduction to obtain low-level target aggregate features; second semantic aggregate features, first low-level injection features and second low-level injection features are processed through high-level semantic injection to obtain mid-level target aggregate features; second semantic aggregate features, mid-level target aggregate features and deep semantic features are processed through high-level semantic injection to obtain high-level target aggregate features;

[0053] The target detection result generation module is used to perform bounding box regression, bounding box classification recognition and non-maximum suppression processing on low-level target aggregation features, intermediate target aggregation features and high-level target aggregation features after linear mapping and full connection processing to obtain target detection results.

[0054] This embodiment also provides a remote sensing image target detection device based on deep learning, including a processor and a memory, wherein the processor implements the above-mentioned remote sensing image target detection method based on deep learning when executing a computer program stored in the memory.

[0055] This embodiment also provides a computer-readable storage medium for storing a computer program, wherein when the computer program is executed by a processor, the above-mentioned remote sensing image target detection method based on deep learning is implemented.

[0056] The present embodiment provides a remote sensing image target detection method based on deep learning. First, the remote sensing image to be detected is subjected to image enhancement processing to improve the feature expression ability of the remote sensing image to be detected. Then, the remote sensing enhanced image is subjected to several small target feature extraction processes to obtain different depth features of the remote sensing enhanced image, and shallow remote sensing features, middle remote sensing features, middle-deep remote sensing features and deep remote sensing features with different local feature detail highlighting intensities are obtained. Then, the shallow remote sensing features, middle remote sensing features, middle-deep remote sensing features and deep remote sensing features are subjected to low-level semantic information aggregation processing and then to cross-low-level semantic injection processing to highlight local color differences and small contour details in the remote sensing image, and obtain first low-level injection features and second low-level injection features. Then, the first low-level injection features and the second low-level injection features are injected into the remote sensing image. After the injected features and deep semantic features are subjected to high-level semantic information aggregation processing, high-level semantic injection processing is performed. On the basis of detail enhancement, the global features of the remote sensing image are improved and the difference between background information and small target details is enhanced to obtain intermediate target aggregation features and high-level target aggregation features; then the low-level target aggregation features, intermediate target aggregation features and high-level target aggregation features obtained based on the first low-level injected features are subjected to linear mapping and full connection processing, so as to better separate the small target from the background, and finally the target detection results are obtained by bounding box regression, bounding box classification recognition and non-maximum suppression processing in sequence; this detection method is applied to target detection in remote sensing images, which can effectively separate small target features from the background, thereby improving the accuracy of small target detection results and avoiding missed detection and false detection.

Claims

1. A remote sensing image target detection method based on deep learning, characterized in that: The following operations are included: S1. The remote sensing image to be inspected is subjected to image enhancement processing to obtain a remote sensing enhanced image; the remote sensing enhanced image is subjected to small target feature extraction processing once, twice, three times, and four times to obtain shallow remote sensing features, middle remote sensing features, middle-deep remote sensing features, and deep remote sensing features; The specific operation of small target feature extraction is as follows: the input is downsampled and convolved to obtain convolution features; the convolution features are subjected to structural feature extraction and multi-scale texture enhancement respectively, and then spliced to obtain output; S2, shallow remote sensing features, mid-layer remote sensing features, mid-deep remote sensing features and deep remote sensing features are aggregated through low-level semantic information to obtain first semantic aggregate features; deep remote sensing features are processed through channel dimension reduction to obtain deep semantic features; deep semantic features, mid-deep remote sensing features, mid-layer remote sensing features and first semantic aggregate features are injected through low-level semantics to obtain first low-level injected features; first low-level injected features, mid-layer remote sensing features, shallow remote sensing features and first semantic aggregate features are injected through low-level semantics to obtain second low-level injected features; The first low-level injection feature, the second low-level injection feature and the deep semantic feature are aggregated through high-level semantic information processing to obtain the second semantic aggregate feature; the second low-level injection feature is processed through channel dimension reduction to obtain the low-level target aggregate feature; The second semantic aggregate feature, the first low-level injection feature, and the second low-level injection feature are processed by high-level semantic injection to obtain the intermediate target aggregate feature; The second semantic aggregation feature, the intermediate target aggregation feature, and the deep semantic feature are processed by high-level semantic injection to obtain the high-level target aggregation feature; S3, low-level target aggregation features, intermediate target aggregation features and high-level target aggregation features are processed by linear mapping and full connection, and then bounding box regression, bounding box classification recognition and non-maximum suppression are performed in sequence to obtain the target detection results.

2. The method for remote sensing image target detection based on deep learning according to claim 1, characterized in that: In S1, the multi-scale texture enhancement processing operation is specifically as follows: The convolution features are subjected to multi-scale convolution feature extraction and spatial context feature extraction respectively to obtain multi-scale convolution features and spatial context features; the multi-scale convolution features and spatial context features are multiplied element by element and then added element by element with the multi-scale convolution features to obtain the initial texture enhancement features; The initial texture enhancement features are processed by convolution to obtain multi-scale texture enhancement features.

3. The method for remote sensing image target detection based on deep learning according to claim 2, characterized in that: The operation of spatial context feature extraction can be achieved through average pooling, convolution, several depth-separable convolutions, convolution and Sigmoid function.

4. The method for remote sensing image target detection based on deep learning according to claim 1, characterized in that: In S2, the operations of low-level semantic information aggregation processing are specifically as follows: The shallow remote sensing features and the middle remote sensing features are processed by average pooling respectively to obtain shallow pooling features and middle pooling features; the deep remote sensing features are processed by bilinear mapping, and then spliced with the middle and deep remote sensing features, shallow pooling features and middle pooling features based on channel connection to obtain the initial aggregation features; the initial aggregation features are processed by convolution, structural parameterization and convolution to obtain the first semantic aggregation features.

5. The method for remote sensing image target detection based on deep learning according to claim 1, characterized in that: In S2, the operation of obtaining the first low-level injection feature is specifically as follows: The deep semantic features, mid-deep remote sensing features, and mid-level remote sensing features are processed by bilinear mapping, convolution, and average pooling, and then spliced and convolved based on channel connection to obtain the first local injection feature; The first semantic aggregation feature is processed by convolution, Sigmoid function and average pooling, and then multiplied element-by-element with the first local injection feature to obtain the first initial injection feature; After convolution and average pooling, the first semantic aggregation feature is added element-by-element with the first initial injection feature to obtain the first low-level injection feature.

6. The method for remote sensing image target detection based on deep learning according to claim 1, characterized in that: In S2, the operations of high-level semantic information aggregation processing are specifically as follows: The first low-level injected features and the second low-level injected features are respectively average-pooled, and then concatenated and convolved with the deep semantic features based on channel connection to obtain the first high-level semantic features; The first high-level semantic feature is processed by depthwise separable convolution, superimposed with the first high-level semantic feature, and then batch channel normalization is performed to obtain the second high-level semantic feature; The second high-level semantic feature is processed by the feedforward neural network, superimposed with the second high-level semantic feature, and then layer normalization is performed to obtain the second semantic aggregation feature.

7. The method for remote sensing image target detection based on deep learning according to claim 1, characterized in that: In S2, the operation of obtaining the intermediate target aggregate features is specifically as follows: The first low-level injection feature and the second low-level injection feature are respectively convolved and average-pooled, and then spliced and convolved based on channel connection to obtain the second local injection feature; The second semantic aggregation feature is processed by convolution, Sigmoid function and average pooling, and then multiplied element-by-element with the second local injection feature to obtain the second initial injection feature; The second semantic aggregate feature is convolved and average pooled, and then added element-wise to the second initial injection feature to obtain the intermediate target aggregate feature.

8. A remote sensing image target detection system based on deep learning, used to implement the remote sensing image target detection method based on deep learning according to claim 1, characterized in that: include: The multi-level remote sensing feature generation module is used to perform image enhancement processing on the remote sensing image to be inspected to obtain a remote sensing enhanced image; the remote sensing enhanced image is subjected to small target feature extraction processing once, twice, three times, and four times respectively to obtain shallow remote sensing features, medium remote sensing features, medium-deep remote sensing features, and deep remote sensing features; the small target feature extraction processing operation is specifically as follows: the input is subjected to downsampling and convolution processing to obtain convolution features; the convolution features are subjected to structural feature extraction and multi-scale texture enhancement respectively, and then spliced to obtain the output; The multi-target aggregation feature generation module is used to aggregate shallow remote sensing features, mid-layer remote sensing features, mid-deep remote sensing features and deep remote sensing features through low-level semantic information to obtain a first semantic aggregation feature; the deep remote sensing features are processed by channel dimension reduction to obtain a deep semantic feature; the deep semantic features, mid-deep remote sensing features, mid-layer remote sensing features and the first semantic aggregation feature are processed by low-level semantic injection to obtain a first low-level injection feature; the first low-level injection feature, mid-layer remote sensing features, shallow remote sensing features and the first semantic aggregation feature are processed by low-level semantic injection to obtain a second low-level injection feature; the first low-level injection feature, the second low-level injection feature and the deep semantic feature are aggregated through high-level semantic information to obtain a second semantic aggregation feature; the second low-level injection feature is processed by channel dimension reduction to obtain a low-level target aggregation feature; The second semantic aggregate feature, the first low-level injection feature, and the second low-level injection feature are processed by high-level semantic injection to obtain the intermediate target aggregate feature; The second semantic aggregation feature, the intermediate target aggregation feature, and the deep semantic feature are processed by high-level semantic injection to obtain the high-level target aggregation feature; The target detection result generation module is used to perform bounding box regression, bounding box classification recognition and non-maximum suppression processing on low-level target aggregation features, intermediate target aggregation features and high-level target aggregation features after linear mapping and full connection processing to obtain target detection results.

9. A remote sensing image target detection device based on deep learning, characterized in that: The method comprises a processor and a memory, wherein when the processor executes the computer program stored in the memory, the method for remote sensing image target detection based on deep learning as described in any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium, characterized in that Used to store a computer program, wherein when the computer program is executed by a processor, the remote sensing image target detection method based on deep learning as described in any one of claims 1 to 7 is implemented.