Remote sensing target detection feature extraction algorithm based on depth multiple scales

Abstract

The invention provides a remote sensing target detection feature extraction algorithm based on depth multiple scales. The remote sensing target detection feature extraction algorithm comprises the following steps: S1, cutting a downloaded public data set; s2, constructing a deep multi-scale remote sensing target detection model; and S3, detecting a to-be-detected remote sensing image by using the trained target detection algorithm model, and generating a final detection image. Through convolution operations of various sizes and a channel attention mechanism, features of different scales can be captured, and attention of the model to important features is enhanced through dynamic weighting. Therefore, the network model has stronger capability in processing complex visual tasks, especially in processing remote sensing images with details and hierarchies. Through wide tests on a plurality of remote sensing data sets, the provided backbone network shows an excellent effect, various types of targets can be effectively detected, the potential and feasibility of the backbone network in practical application are proved, and the innovative network structure also brings a new thought for the field of remote sensing image processing.

Description

Technical Field

[0001] The present invention relates to the technical field of remote sensing image processing, and in particular to a remote sensing target detection feature extraction algorithm based on deep multi-scale. Background Art

[0002] In the context of modern remote sensing technology, the acquisition of high-resolution remote sensing images has become an important research field. These images provide valuable data resources for multiple applications such as geographic information systems (GIS), environmental monitoring, urban planning, agricultural management, and disaster assessment. However, with the widespread application of remote sensing images, the difficulty of target detection and recognition has also increased, especially in the case of complex backgrounds, different scales, and changes in target shape. These challenges significantly affect the accuracy and reliability of detection.

[0003] Traditional target detection methods usually rely on manual feature extraction techniques. These methods may perform well in specific scenarios, but they often fail to cope with remote sensing images with diverse and complex features. With the continuous development of deep learning technology in the field of computer vision, especially the introduction of convolutional neural networks (CNNs), the ability of automatic feature extraction has been greatly improved. These deep learning models can automatically learn and extract high-level features from raw data through multi-level nonlinear transformations, especially showing significant advantages in multi-scale feature extraction.

[0004] Targets in remote sensing images often have different scales and shapes, which makes it difficult for a single-scale feature extraction method to effectively capture the information of all targets. Studies have shown that multi-scale feature extraction can effectively improve the performance of detection algorithms, especially when the target sizes vary greatly. Therefore, combining deep learning with multi-scale feature extraction technology can better adapt to the complex target features in remote sensing images, thereby improving the accuracy and robustness of target detection.

[0005] Although some deep learning-based models, such as Faster R-CNN, YOLO, and RetinaNet, have been widely used for object detection in remote sensing images, these methods tend to ignore local details when dealing with large-scale scenes, resulting in insufficient detection capabilities for small objects. In addition, in scenes with complex backgrounds or dense objects, these models may face challenges such as target loss or false detection.

[0006] Therefore, the proposed deep multi-scale feature extraction algorithm provides a new idea for remote sensing target detection. This algorithm can capture the detailed information in the image more comprehensively, especially significantly improve the detection ability of small targets, thereby enhancing the overall detection performance. This method not only helps to overcome the limitations of existing technologies, but also opens up a new research direction for remote sensing image analysis. Summary of the invention

[0007] The present invention provides a remote sensing target detection feature extraction algorithm based on deep multi-scale, aiming to solve the problems of false detection and missed detection caused by insufficient feature extraction due to complex background in the target detection process in remote sensing images, especially for the detection of small targets.

[0008] In order to solve the above technical problems, the technical solution adopted by the present invention is: The remote sensing target detection feature extraction algorithm based on deep multi-scale includes the following steps: S1, cutting the downloaded public data set to generate images of size 1024×1024, and the number of overlapping pixels of the cut images is 500; S2. Build a deep multi-scale remote sensing target detection model, and use the cut remote sensing data set to train a deep multi-scale remote sensing target detection algorithm to obtain a remote sensing target detection model; S3. Use the trained target detection algorithm model to detect the remote sensing image to be detected and generate the final detection image.

[0009] The specific steps of S1 above are: S1.1. Download and decompress the dataset from the official website of the public remote sensing dataset. The decompressed dataset is divided into training set, validation set and test set. S1.2. Crop the decompressed data set using the written processing script. The number of overlapping pixels of the images during cutting is 500. The data set after cutting is divided into training validation set, validation set and test set.

[0010] In S2 above, the construction of a deep multi-scale remote sensing target detection algorithm is divided into the following four stages: S2.1, in the first stage, for any input RGB three-color image, a downsampling block is first used to obtain a feature map 102, and then the feature map 102 is used as input to stack two proposed deep multi-scale feature extraction modules to obtain an output feature map 105; The network architecture at this stage is: input three-color RGB image → downsampling block 101 → feature map 102; S2.2, the second stage, the construction process is the same as the first stage, using the final output feature map 105 of the first stage as input, first using the downsampling block to obtain the feature map 107, then using the feature map 107 as input, stacking two proposed deep multi-scale feature extraction modules to obtain the output feature map 110; The network architecture at this stage is: Feature map 105 → downsampling block 106 → feature map 107; Feature map 107 → deep multi-scale feature extraction module 108 → deep multi-scale feature extraction module 109 → feature map 110; S2.3, the third stage, the overall process is the same as the first and second stages. In the third stage, the deep multi-scale feature extraction module is stacked four times, the feature map 110 output by the second stage is used as input, and the downsampling block is used to obtain the feature map 112. Then, the feature map 112 is used as input, and four proposed deep multi-scale feature extraction modules are stacked to obtain the output feature map 117. The network architecture at this stage is: Feature map 110 → downsampling block 111 → feature map 112; Feature map 112 → deep multi-scale feature extraction module 113 → deep multi-scale feature extraction module 114 → deep multi-scale feature extraction module 115 → deep multi-scale feature extraction module 116 → feature map 117; S2.4, the fourth stage, the construction process is the same as the first stage, using the final output feature map 117 of the third stage as input, first using the downsampling block to obtain the feature map 1119, then using the feature map 119 as input, stacking two proposed deep multi-scale feature extraction modules to obtain the output feature map 122; The network architecture at this stage is: Feature map 117 → downsampling block 118 → feature map 119; Feature map 119 →deep multi-scale feature extraction module 120 →deep multi-scale feature extraction module 121 →feature map 122.

[0011] In the above four stages of constructing a deep multi-scale remote sensing target detection algorithm, the sub-steps of construction are: The input feature 200 is subjected to a depthwise separable convolution D1 of size 3*3 to obtain a feature map 201; The obtained feature map 201 is used as input and subjected to a depthwise separable convolution D2 of size 5*5 to obtain a feature map 202; The obtained feature map 202 is used as input and subjected to a depthwise separable convolution D3 of size 7*7 to obtain a feature map 203; The obtained feature map 204 is used as input and subjected to a depthwise separable convolution D4 of size 9*9 to obtain the feature map 204; The four feature maps 201, 202, 203, and 204 are respectively averaged at the channel level to obtain four feature maps of different sizes H×W×1, and these four feature maps are concatenated at the channel level to obtain a spatial weight of size H×W×4; The obtained spatial weights are mixed using a convolution of size 7*7 and padding 4 to obtain a spatial weight 205 of size H×W×4; The obtained spatial weight is re-weighted using the Sigmoid function; The Sigmoid function is defined by the following formula:

[0012] The obtained spatial weights are channel-summed and re-weighted with the network self-learned weight parameters to obtain a weighted feature map 206. The weighted feature map 206 is multiplied by the input feature map 200 to obtain the final output feature 207.

[0013] The present invention provides a remote sensing target detection feature extraction algorithm based on deep multi-scale, and the beneficial effects are as follows: (1) A deep multi-scale remote sensing target detection network framework is proposed to improve the performance of target detection in remote sensing images. The framework adopts a convolution kernel design without dilated convolution, which not only effectively reduces the number of model parameters and computational complexity, but also avoids excessive smoothing of information and loss of details while maintaining efficient feature extraction. In addition, the overall network structure has been carefully optimized to ensure that the model's running speed is improved while ensuring high accuracy.

[0014] (2) A plug-and-play deep multi-scale feature extraction module is proposed, which can capture features of different scales through convolution operations of various sizes and channel attention mechanism, and enhance the model's attention to important features through dynamic weighting. This makes the network model more capable in handling complex visual tasks, especially in processing remote sensing images with details and layers.

[0015] (3) The proposed backbone network has been extensively tested on multiple remote sensing datasets and has demonstrated excellent results. It can effectively detect various types of targets, proving its potential and feasibility in practical applications. This innovative network structure has also brought new ideas to the field of remote sensing image processing. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] The present invention will be further described below in conjunction with the accompanying drawings and embodiments: Figure 1 The deep multi-scale remote sensing target detection framework proposed by the present invention; Figure 2 This is the deep multi-scale feature extraction module proposed in the present invention. DETAILED DESCRIPTION

[0017] In order to make the purpose, technical solutions and advantages of the present invention clearer, the following content will systematically and completely describe the specific technical solutions of the present invention in combination with the drawings provided according to the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0018] The remote sensing target detection feature extraction algorithm based on deep multi-scale includes the following steps: S1, cutting the downloaded public data set (such as DOTA, DIOR) to generate a picture of size 1024×1024, and the number of overlapping pixels of the cut picture is 500; S2. Build a deep multi-scale remote sensing target detection model, and use the cut remote sensing data set to train a deep multi-scale remote sensing target detection algorithm to obtain a remote sensing target detection model; S3. Use the trained target detection algorithm model to detect the remote sensing image to be detected and generate the final detection image.

[0019] The specific steps of S1 above are: S1.1. Download and decompress the dataset from the official website of the public remote sensing dataset (taking DOTA as an example). The decompressed dataset is divided into training set, validation set and test set; S1.2. The original dataset images are large, and direct training requires more training resources. Therefore, the decompressed dataset needs to be cropped and cut using a written processing script. The number of overlapping pixels during cutting is 500. The cut dataset is divided into a training validation set, a validation set, and a test set.

[0020] In S2 above, the construction of a deep multi-scale remote sensing target detection algorithm is divided into the following four stages: S2.1, in the first stage, for any input RGB three-color image, a downsampling block is first used to obtain a feature map 102, and then the feature map 102 is used as input to stack two proposed deep multi-scale feature extraction modules to obtain an output feature map 105; The network architecture at this stage is: input three-color RGB image → downsampling block 101 → feature map 102; S2.2, the second stage, the construction process is the same as the first stage, using the final output feature map 105 of the first stage as input, first using the downsampling block to obtain the feature map 107, then using the feature map 107 as input, stacking two proposed deep multi-scale feature extraction modules to obtain the output feature map 110; The network architecture at this stage is: Feature map 105 → downsampling block 106 → feature map 107; Feature map 107 → deep multi-scale feature extraction module 108 → deep multi-scale feature extraction module 109 → feature map 110; S2.3, the third stage, the overall process is the same as the first and second stages. In the third stage, the deep multi-scale feature extraction module is stacked four times, the feature map 110 output by the second stage is used as input, and the downsampling block is used to obtain the feature map 112. Then, the feature map 112 is used as input, and four proposed deep multi-scale feature extraction modules are stacked to obtain the output feature map 117. The network architecture at this stage is: Feature map 110 → downsampling block 111 → feature map 112; Feature map 112 → deep multi-scale feature extraction module 113 → deep multi-scale feature extraction module 114 → deep multi-scale feature extraction module 115 → deep multi-scale feature extraction module 116 → feature map 117; S2.4, the fourth stage, the construction process is the same as the first stage, using the final output feature map 117 of the third stage as input, first using the downsampling block to obtain the feature map 1119, then using the feature map 119 as input, stacking two proposed deep multi-scale feature extraction modules to obtain the output feature map 122; The network architecture at this stage is: Feature map 117 → downsampling block 118 → feature map 119; Feature map 119 →deep multi-scale feature extraction module 120 →deep multi-scale feature extraction module 121 →feature map 122.

[0021] In the above four stages of constructing a deep multi-scale remote sensing target detection algorithm, the sub-steps of construction are: The input feature 200 is subjected to a depthwise separable convolution D1 of size 3*3 to obtain a feature map 201; The obtained feature map 201 is used as input and subjected to a depthwise separable convolution D2 of size 5*5 to obtain a feature map 202; The obtained feature map 202 is used as input and subjected to a depthwise separable convolution D3 of size 7*7 to obtain a feature map 203; The obtained feature map 204 is used as input and subjected to a depthwise separable convolution D4 of size 9*9 to obtain the feature map 204; The four feature maps 201, 202, 203, and 204 are respectively averaged at the channel level to obtain four feature maps of different sizes H×W×1, and these four feature maps are concatenated at the channel level to obtain a spatial weight of size H×W×4; The obtained spatial weights are mixed using a convolution of size 7*7 and padding 4 to obtain a spatial weight 205 of size H×W×4; The obtained spatial weight is re-weighted using the Sigmoid function; The Sigmoid function is defined by the following formula:

[0022] The obtained spatial weights are channel-summed and re-weighted with the network self-learned weight parameters to obtain a weighted feature map 206. The weighted feature map 206 is multiplied by the input feature map 200 to obtain the final output feature 207.

[0023] Embodiment 1: A remote sensing target detection algorithm based on deep multi-scale includes the following steps: Step S1: Data preprocessing: The downloaded public datasets, such as DOTA and DIOR, are cut to generate images of size 1024 × 1024. The number of overlapping pixels of the cut images is 500 to ensure the continuity of information between adjacent image blocks, so as to better capture the characteristics of the target at different scales and contexts.

[0024] Step S2: Model construction and training; The proposed deep multi-scale remote sensing target detection model is constructed. The model design considers multi-level feature fusion to enhance the recognition ability of targets of different sizes. The deep multi-scale remote sensing target detection algorithm is trained using the cut remote sensing dataset. Data enhancement techniques such as random flipping, rotation and scaling are used during the training process to improve the generalization ability of the model, and finally a remote sensing target detection model is obtained.

[0025] Step S3: Target detection and result generation: Use the trained target detection algorithm model to detect the remote sensing image to be detected and generate the final detection image. During the detection process, the model analyzes the multi-scale features of the image, generates the bounding box and category label of the target, and visualizes the detection results for further analysis and application.

[0026] The above steps complete the construction of the deep multi-scale remote sensing target detection network.

[0027] like Figure 1 As shown in the figure, when building a deep multi-scale remote sensing target detection algorithm, it is necessary to first build a deep multi-scale feature extraction module. The detailed construction steps are as follows: The size of the input feature 200 is C×H×W. The depthwise separable convolution D1 with a convolution kernel size of 3 and a padding of 1 is used to convolve the input feature map. The output feature after convolution is 201, and the size is still C×H×W; the depthwise separable convolution D2 with a convolution kernel size of 5 and a padding of 2 is used to convolve the obtained feature map. The output feature after convolution is 202, and the size is still C×H×W; the depthwise separable convolution D3 with a convolution kernel size of 7 and a padding of 3 is used to convolve the obtained feature map. The output feature after convolution is 203, and the size is still C×H×W; the depthwise separable convolution D4 with a convolution kernel size of 9 and a padding of 4 is used to convolve the obtained feature map. The output feature after convolution is 204, and the size is still C×H×W.

[0028] The four feature maps are subsequently fused with weights. Compared with the common multi-scale approach, we introduce dynamic parameters that can be learned by the network for fusion. The specific steps of fusion are as follows: The obtained four feature maps 201, 202, 203, and 204 are respectively averaged at the channel level to obtain four feature maps of different sizes of H×W×1, and then the four feature maps are spliced ​​at the channel level to obtain a spatial weight of size H×W×4; the obtained spatial weight is mixed using a convolution of size 7*7 and padding 4 to obtain a spatial weight 205 of size H×W×4; the obtained spatial weight is re-weighted using a Sigmoid function, where the Sigmoid function is defined as follows:

[0029] The obtained spatial weights are channel-summed and re-weighted with the weight parameters self-learned by the network to obtain a weighted feature map 206; the weighted feature map 206 is multiplied by the input feature map 200 to obtain the final output feature 207.

[0030] Constructing a deep multi-scale feature extraction module is conducive to extracting multi-scale feature information. By stacking deep separable convolutional layers of different scales and performing spatial weight calculations, the weighting and fusion of feature maps are achieved, improving the network's ability to understand and represent image features. Re-weighting using the Sigmoid function helps to enhance the expressiveness of feature maps and suppress noise. Finally, the final output feature is obtained by multiplying the weighted feature map with the input feature, retaining effective information and improving network performance and accuracy.

[0031] The overall network construction process is as follows: Input three-color RGB image → downsampling block 101 → feature map 102; Feature map 102 → deep multi-scale feature extraction module 103 → deep multi-scale feature extraction module 104 → feature map 105; Feature map 105 → downsampling block 106 → feature map 107; Feature map 107 → deep multi-scale feature extraction module 108 → deep multi-scale feature extraction module 109 → feature map 110; Feature map 110 → downsampling block 111 → feature map 112; Feature map 112 → deep multi-scale feature extraction module 113 → deep multi-scale feature extraction module 114 → deep multi-scale feature extraction module 115 → deep multi-scale feature extraction module 116 → feature map 117; Feature map 117 → downsampling block 118 → feature map 119; Feature map 119 → deep multi-scale feature extraction module 120 → deep multi-scale feature extraction module 121 → feature map 122; When constructing a deep multi-scale feature extraction module, the construction process is as follows: Input feature 200 → D1 (3×3 depthwise separable convolution) → feature map 201; Feature map 201 → D2 (5×5 depthwise separable convolution) → feature map 202; Feature map 202 → D3 (7×7 depthwise separable convolution) → feature map 203; Feature map 203 → D4 (9×9 depthwise separable convolution) → feature map 204; Feature maps 201, 202, 203, 204 → channel averaging → four feature maps of size H×W×1 → channel concatenation → spatial weight 205; Spatial weight 205 → 7*7 convolution (padding 4) → spatial weight 205; Spatial weight 205 → Sigmoid function → re-weighting; Spatial weight 205 → channel summation → weighted feature map 206; Weighted feature map 206 × input feature 200 → output feature 207.

[0032] The contribution of the present invention in the field of remote sensing target detection is mainly reflected in the innovation of the algorithm and the improvement of performance. First, by introducing a deep multi-scale feature extraction module and combining a multi-level feature fusion mechanism, the model can better identify targets of different scales, thereby improving the accuracy and robustness of detection. Secondly, in the data preprocessing and model training stages, the cutting of public data sets and data enhancement techniques, as well as deep separable convolutional layers for feature extraction, effectively improve the generalization ability and performance of the model. In addition, through the network's learnable dynamic parameters and re-weighting mechanism, the ability to understand and characterize image features is enhanced, making the algorithm more practical and versatile.

[0033] In general, the deep multi-scale remote sensing target detection algorithm proposed in this paper comprehensively applies a number of technical means, such as multi-scale feature extraction, data processing, network optimization, etc., and has shown good detection accuracy and efficiency through experimental verification, and has broad application prospects and promotion value. This comprehensive algorithm design idea and method innovation has brought new inspiration and breakthroughs to the research field of remote sensing image processing and target detection, and is of great significance to the research and practical application of related fields.

[0034] Embodiment 2: In order to illustrate the feasibility of this scheme, further explanation is given below in combination with experiments.

[0035] (1) Experimental setup: The code of this invention is based on the Pytorch framework experiment, and the hardware uses 6 NVIDIA 3090Ti GPUs to train the network model. This experiment uses Oriented-RCNN as the overall network framework, and embeds the proposed deep multi-scale remote sensing target detection algorithm as the backbone. In addition, the deep multi-scale remote sensing target detection network proposed in this invention is first pre-trained on ImageNet1K for 300 epochs as pre-training weights. During training, the size of the input image is set to 1024×1024, and data enhancement strategies such as random rotation are used.

[0036] (2) Dataset setting: In order to verify the versatility of the method of the present invention, training and verification are carried out on three mainstream datasets: DOTA dataset Dota-v1.0 is a large-scale open-domain remote sensing image dataset, containing 2806 remote sensing images, covering multiple geographical environments of different cities across China. The dataset contains 188,282 target instances of different categories, mainly including airplanes, ships, storage tanks, basketball courts, etc. The DOTA dataset is characterized by small target scale, dense distribution and irregular shape, which is suitable for tasks such as target detection and target recognition.

[0037] DOTA-v1.5 is an updated version of the DOTA dataset, which contains 2806 remote sensing images and 188,282 target instances of different categories. Compared with previous versions, DOTA-v1.5 has made some improvements and optimizations. First, more remote sensing images have been added, covering more different geographical environments of Chinese cities, making the dataset more diverse and comprehensive. Secondly, the target categories have been expanded, and new categories such as bridges and parking lots have been introduced, increasing the richness of the dataset. In addition, the annotation accuracy has also been optimized, which improves the accuracy and reliability of target detection and recognition tasks.

[0038] The DIOR dataset (Detection in ORiented scenes Dataset) is a remote sensing image dataset for detecting rotating targets, containing 2,800 high-resolution aerial remote sensing images. The targets in the DIOR dataset mainly include vehicles, ships, airplanes, etc. These targets usually present irregular shapes such as rotation and tilt. The DIOR dataset is characterized by target rotation, which is suitable for testing the performance of target detection algorithms in actual scenes with rotating targets.

[0039] (3) Parameter setting: The AdamW optimizer was used for optimization. The network momentum was 0.9, the weight decay factor was 0.05, the warm-up method was used for preheating, and cosine annealing was used to update the learning rate. The number of training rounds was 24, and the initial learning rate was 0.0001.

[0040] (4) Evaluation indicators: The main indicator of this experiment is AP50. AP50 refers to the use of 0.5 as the IoU (Intersection over Union) threshold when calculating the average precision (AP), which can better reflect the accuracy of the detector.

[0041] Final experimental results: Table 1: Quantitative evaluation with state-of-the-art methods on DOTAv1.0 Method Venue #Params Small-Vehicle swimming-pool AP50 GWD ICML2021 47.4M 73.31 70.15 72.61 RoI Trans CVPR2019 55.1M 74.43 71.99 74.05 R3Det AAAI2021 41.9M 76.18 62.39 69.70 S2ANet TGRS2021 38.6M 78.11 69.13 74.12 O-RCNN ICCV2021 41.1M 78.93 68.84 75.81 LSKNet ICCV2023 31.0M 74.95 74.47 77.49 PKINet CVPR2024 30.8M 80.25 73.62 78.39 DHCF(ours) ---- 31.8M 81.26 81.51 79.59 Table 2: Quantitative evaluation with state-of-the-art methods on DOTAv1.5

[0042] Table 3: Quantitative evaluation with state-of-the-art methods on DIOR-R

[0043] Experimental results on multiple remote sensing target detection datasets show that the proposed DHCF network significantly outperforms many existing state-of-the-art methods, especially in small target detection. First, in the evaluation of the DOTAv1.0 dataset, DHCF achieved a detection accuracy of 81.26 for small vehicles, which is significantly higher than other methods such as S2ANet and O-RCNN. This result highlights its advantages in fine-grained feature extraction and target localization. Small target detection has always been a major challenge in remote sensing image analysis, and DHCF can effectively capture the details of small targets, indicating its effectiveness in complex scenarios.

[0044] On the DOTAv1.5 dataset, DHCF also performed well in the small vehicle category, achieving an AP value of 52.61. Although the overall AP50 (73.67) was slightly lower than some top methods, the outstanding performance in specific target categories further demonstrated the superiority of DHCF in small target detection. Finally, in the comparison of the DIOR-R dataset, DHCFNet still surpassed multiple classic methods with an AP50 of 71.95, showing its reliability in complex remote sensing images.

[0045] In summary, the DHCF network has a small number of parameters and high accuracy, especially in small target detection, showing the effectiveness of deep multi-scale feature fusion. These results not only verify the design concept of DHCF, but also provide strong support and new directions for further research in the field of remote sensing target detection.

Claims

1. A remote sensing target detection feature extraction algorithm based on deep multi-scale, characterized in that: The following steps are involved: S1. Split the downloaded public dataset to generate images of size 1024×1024, with the number of overlapping pixels of the split images being 500; S2. Build a deep multi-scale remote sensing target detection model, and use the cut remote sensing data set to train a deep multi-scale remote sensing target detection algorithm to obtain a remote sensing target detection model; S3. Use the trained target detection algorithm model to detect the remote sensing image to be detected and generate the final detection image.

2. The remote sensing target detection feature extraction algorithm based on deep multi-scale according to claim 1 is characterized in that: The specific steps of S1 are: S1.

1. Download and decompress the dataset from the official website of the public remote sensing dataset. The decompressed dataset is divided into training set, validation set and test set. S1.

2. Crop the decompressed data set using the written processing script. The number of overlapping pixels of the images during cutting is 500. The data set after cutting is divided into training validation set, validation set and test set.

3. The remote sensing target detection feature extraction algorithm based on deep multi-scale according to claim 2 is characterized in that: In S2, the construction of a deep multi-scale remote sensing target detection algorithm is divided into the following four stages: S2.1, in the first stage, for any input RGB three-color image, a downsampling block is first used to obtain a feature map 102, and then the feature map 102 is used as input to stack two proposed deep multi-scale feature extraction modules to obtain an output feature map 105; The network architecture at this stage is: input three-color RGB image → downsampling block 101 → feature map 102; S2.2, the second stage, the construction process is the same as the first stage, using the final output feature map 105 of the first stage as input, first using the downsampling block to obtain the feature map 107, then using the feature map 107 as input, stacking two proposed deep multi-scale feature extraction modules to obtain the output feature map 110; The network architecture at this stage is: Feature map 105 → downsampling block 106 → feature map 107; Feature map 107 → deep multi-scale feature extraction module 108 → deep multi-scale feature extraction module 109 → feature map 110; S2.3, the third stage, the overall process is the same as the first and second stages. In the third stage, the deep multi-scale feature extraction module is stacked four times, the feature map 110 output by the second stage is used as input, and the downsampling block is used to obtain the feature map 112. Then, the feature map 112 is used as input, and four proposed deep multi-scale feature extraction modules are stacked to obtain the output feature map 117. The network architecture at this stage is: Feature map 110 → downsampling block 111 → feature map 112; Feature map 112 → deep multi-scale feature extraction module 113 → deep multi-scale feature extraction module 114 → deep multi-scale feature extraction module 115 → deep multi-scale feature extraction module 116 → feature map 117; S2.4, the fourth stage, the construction process is the same as the first stage, using the final output feature map 117 of the third stage as input, first using the downsampling block to obtain the feature map 1119, then using the feature map 119 as input, stacking two proposed deep multi-scale feature extraction modules to obtain the output feature map 122; The network architecture at this stage is: Feature map 117 → downsampling block 118 → feature map 119; Feature map 119 →deep multi-scale feature extraction module 120 →deep multi-scale feature extraction module 121 →feature map 122.

4. The remote sensing target detection feature extraction algorithm based on deep multi-scale according to claim 3 is characterized in that: In the four stages of constructing a deep multi-scale remote sensing target detection algorithm, the sub-steps of construction are: The input feature 200 is subjected to a depthwise separable convolution D1 of size 3*3 to obtain a feature map 201; The obtained feature map 201 is used as input and subjected to a depthwise separable convolution D2 of size 5*5 to obtain a feature map 202; The obtained feature map 202 is used as input and subjected to a depthwise separable convolution D3 of size 7*7 to obtain a feature map 203; The obtained feature map 204 is used as input and subjected to a depthwise separable convolution D4 of size 9*9 to obtain the feature map 204; The four feature maps 201, 202, 203, and 204 are respectively averaged at the channel level to obtain four feature maps of different sizes H×W×1, and these four feature maps are concatenated at the channel level to obtain a spatial weight of size H×W×4; The obtained spatial weights are mixed using a convolution of size 7*7 and padding 4 to obtain a spatial weight 205 of size H×W×4; The obtained spatial weight is re-weighted using the Sigmoid function; The Sigmoid function is defined by the following formula: The obtained spatial weights are channel-summed and re-weighted with the network self-learned weight parameters to obtain a weighted feature map 206. The weighted feature map 206 is multiplied by the input feature map 200 to obtain the final output feature 207.

Images