Remote sensing target detection feature extraction algorithm based on depth multi-scale

Abstract

This invention proposes a feature extraction algorithm for remote sensing target detection based on deep multi-scale, comprising: S1, segmenting a downloaded public dataset; S2, constructing a deep multi-scale remote sensing target detection model; and S3, using the trained target detection algorithm model to detect the remote sensing image to be detected and generating the final detected image. Through convolutional operations of various sizes and channel attention mechanisms, features at different scales can be captured, and dynamic weighting enhances the model's focus on important features. This makes the network model more capable of handling complex visual tasks, especially when processing remote sensing images with detail and hierarchy. The proposed backbone network has been extensively tested on multiple remote sensing datasets, demonstrating excellent performance and effectively detecting various types of targets, proving its potential and feasibility in practical applications. This innovative network structure also brings new ideas to the field of remote sensing image processing.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image processing technology, specifically to a feature extraction algorithm for remote sensing target detection based on depth multi-scale. Background Technology

[0002] In the context of modern remote sensing technology, the acquisition of high-resolution remote sensing images has become an important research area. These images provide valuable data resources for various 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 cases of complex backgrounds, different scales, and varying target shapes. These challenges significantly affect the accuracy and reliability of detection.

[0003] Traditional object detection methods typically rely on manual feature extraction techniques. While these methods may perform well in specific scenarios, they often fall short when dealing with remotely sensed images exhibiting diverse and complex features. With the continuous development of deep learning technologies in computer vision, especially the introduction of Convolutional Neural Networks (CNNs), the ability to automatically extract features has been greatly enhanced. These deep learning models, through multi-level nonlinear transformations, can automatically learn and extract high-level features from raw data, demonstrating significant advantages, particularly in multi-scale feature extraction.

[0004] Targets in remote sensing images often have different scales and shapes, making it difficult for single-scale feature extraction methods to effectively capture information about all targets. Research shows that multi-scale feature extraction can effectively improve the performance of detection algorithms, especially when targets vary significantly in size. Therefore, combining deep learning with multi-scale feature extraction techniques 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 target detection in remote sensing images, these methods often neglect local details when dealing with large-scale scenes, resulting in insufficient detection capabilities for small targets. Furthermore, in scenes with complex backgrounds or dense targets, these models may face the challenge of target loss or false detection.

[0006] Therefore, a deep multi-scale feature extraction algorithm is proposed, providing a new approach for remote sensing target detection. This algorithm can capture more comprehensive details in images, especially significantly improving the detection capability of small targets, thereby enhancing overall detection performance. This method not only helps overcome the limitations of existing technologies but also opens up new research directions for remote sensing image analysis. Summary of the Invention

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

[0008] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0009] The feature extraction algorithm for remote sensing target detection based on depth multi-scale includes the following steps: S1, the downloaded public dataset is segmented to generate an image of size 1024×1024, and the number of overlapping pixels of the segmented images is 500.

[0010] S2. Construct a deep multi-scale remote sensing target detection model. Use the segmented remote sensing dataset to train the deep multi-scale remote sensing target detection algorithm to obtain the remote sensing target detection model.

[0011] S3. Use the trained target detection algorithm model to detect the remote sensing image to be detected and generate the final detection image.

[0012] The specific steps of S1 above are as follows:

[0013] S1.1 Download the dataset from the official website of the publicly available remote sensing dataset and decompress it. The decompressed dataset is divided into training set, validation set and test set.

[0014] S1.2. The decompressed dataset is cropped using a pre-written processing script. The number of overlapping pixels between images is 500. The cropped dataset is divided into training and validation sets, validation sets, and test sets.

[0015] In S2 above, the construction of the deep multi-scale remote sensing target detection algorithm is divided into the following four stages:

[0016] S2.1 First stage: For any input RGB three-color image, a downsampling block is first used to obtain feature map 102. Then, feature map 102 is used as input, and two proposed deep multi-scale feature extraction modules are stacked to obtain output feature map 105.

[0017] The network architecture at this stage is: input three-color RGB image → downsampling block 101 → feature map 102;

[0018] S2.2, Second stage, the construction process is the same as the first stage. The final output feature map 105 of the first stage is used as input. First, a downsampling block is used to obtain feature map 107. Then, feature map 107 is used as input, and two proposed deep multi-scale feature extraction modules are stacked to obtain the output feature map 110.

[0019] The network architecture at this stage is as follows:

[0020] Feature map 105 → downsampling block 106 → feature map 107;

[0021] Feature map 107 → Deep multi-scale feature extraction module 108 → Deep multi-scale feature extraction module 109 → Feature map 110;

[0022] 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 from 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 the four proposed deep multi-scale feature extraction modules are stacked to obtain the output feature map 117.

[0023] The network architecture at this stage is as follows:

[0024] Feature map 110 → downsampling block 111 → feature map 112;

[0025] 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;

[0026] 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, firstly using downsampling blocks to obtain feature map 1119, then using feature map 119 as input, stacking two proposed deep multi-scale feature extraction modules to obtain the output feature map 122;

[0027] The network architecture at this stage is as follows:

[0028] Feature map 117 → downsampling block 118 → feature map 119;

[0029] Feature map 119 → Deep multi-scale feature extraction module 120 → Deep multi-scale feature extraction module 121 → Feature map 122.

[0030] The sub-steps in the four stages of constructing the depth multi-scale remote sensing target detection algorithm described above are as follows:

[0031] The input feature 200 is processed by a 3*3 depthwise separable convolution D1 to obtain feature map 201;

[0032] The obtained feature map 201 is used as input and passed through a depthwise separable convolution D2 of size 5*5 to obtain feature map 202;

[0033] The obtained feature map 202 is used as input and passed through a depthwise separable convolution D3 of size 7*7 to obtain feature map 203;

[0034] The obtained feature map 204 is used as input and passed through a depthwise separable convolution D4 of size 9*9 to obtain feature map 204;

[0035] The four feature maps 201, 202, 203, and 204 are averaged at the channel level to obtain four different feature maps of size H×W×1. These four feature maps are then spliced ​​together at the channel level to obtain a spatial weight of size H×W×4.

[0036] The obtained spatial weights are mixed using a convolution of size 7*7 with padding of 4 to obtain a spatial weight of size H×W×4, 205.

[0037] The obtained spatial weights are then reweighted using the Sigmoid function;

[0038] The Sigmoid function is defined by the following formula:

[0039]

[0040] The obtained spatial weights are summed by channels and reweighted with the weight parameters learned by the network to obtain a weighted feature map 206. The weighted feature map 206 is then multiplied with the input feature map 200 to obtain the final output feature 207.

[0041] This invention provides a feature extraction algorithm for remote sensing target detection based on depth multi-scale, with the following beneficial effects:

[0042] (1) A depth-based 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 non-dilated convolution kernel design, 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 maintaining high accuracy.

[0043] (2) A plug-and-play deep multi-scale feature extraction module is proposed. Through convolution operations of various sizes and channel attention mechanism, it can capture features at different scales and enhance the model's attention to important features through dynamic weighting. This makes the network model more capable of handling complex visual tasks, especially when processing remote sensing images with details and layers.

[0044] (3) The proposed backbone network has been extensively tested on multiple remote sensing datasets and has shown excellent performance. 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. Attached Figure Description

[0045] The present invention will be further described below with reference to the accompanying drawings and embodiments:

[0046] Figure 1 This invention presents a multi-scale remote sensing target detection framework for depth.

[0047] Figure 2 This is the deep multi-scale feature extraction module proposed in this invention. Detailed Implementation

[0048] To make the objectives, technical solutions, and advantages of this invention clearer, the following will describe the specific technical solutions of this invention systematically and completely in conjunction with the accompanying drawings provided by this invention. Obviously, the described embodiments are only some embodiments of this invention, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0049] The remote sensing target detection feature extraction algorithm based on depth multi-scale includes the following steps: S1, the downloaded public dataset (such as DOTA, DIOR) is cut to generate an image of size 1024×1024, and the number of overlapping pixels of the cut images is 500.

[0050] S2. Construct a deep multi-scale remote sensing target detection model. Use the segmented remote sensing dataset to train the deep multi-scale remote sensing target detection algorithm to obtain the remote sensing target detection model.

[0051] S3. Use the trained target detection algorithm model to detect the remote sensing image to be detected and generate the final detection image.

[0052] The specific steps of S1 above are as follows:

[0053] S1.1 Download the dataset from the official website of the publicly available remote sensing dataset (taking DOTA as an example) and decompress it. The decompressed dataset is divided into training set, validation set and test set.

[0054] S1.2 The original dataset images are large, and direct training would require a lot of training resources. Therefore, the decompressed dataset needs to be cropped. The pre-written processing script is used to cut the images. The number of overlapping pixels in the images is 500. The cropped dataset is divided into training and validation sets, validation sets and test sets.

[0055] In S2 above, the construction of the deep multi-scale remote sensing target detection algorithm is divided into the following four stages:

[0056] S2.1 First stage: For any input RGB three-color image, a downsampling block is first used to obtain feature map 102. Then, feature map 102 is used as input, and two proposed deep multi-scale feature extraction modules are stacked to obtain output feature map 105.

[0057] The network architecture at this stage is: input three-color RGB image → downsampling block 101 → feature map 102;

[0058] S2.2, Second stage, the construction process is the same as the first stage. The final output feature map 105 of the first stage is used as input. First, a downsampling block is used to obtain feature map 107. Then, feature map 107 is used as input, and two proposed deep multi-scale feature extraction modules are stacked to obtain the output feature map 110.

[0059] The network architecture at this stage is as follows:

[0060] Feature map 105 → downsampling block 106 → feature map 107;

[0061] Feature map 107 → Deep multi-scale feature extraction module 108 → Deep multi-scale feature extraction module 109 → Feature map 110;

[0062] 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 from 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 the four proposed deep multi-scale feature extraction modules are stacked to obtain the output feature map 117.

[0063] The network architecture at this stage is as follows:

[0064] Feature map 110 → downsampling block 111 → feature map 112;

[0065] 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;

[0066] 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, firstly using downsampling blocks to obtain feature map 1119, then using feature map 119 as input, stacking two proposed deep multi-scale feature extraction modules to obtain the output feature map 122;

[0067] The network architecture at this stage is as follows:

[0068] Feature map 117 → downsampling block 118 → feature map 119;

[0069] Feature map 119 → Deep multi-scale feature extraction module 120 → Deep multi-scale feature extraction module 121 → Feature map 122.

[0070] The sub-steps in the four stages of constructing the depth multi-scale remote sensing target detection algorithm described above are as follows:

[0071] The input feature 200 is processed by a 3*3 depthwise separable convolution D1 to obtain feature map 201;

[0072] The obtained feature map 201 is used as input and passed through a depthwise separable convolution D2 of size 5*5 to obtain feature map 202;

[0073] The obtained feature map 202 is used as input and passed through a depthwise separable convolution D3 of size 7*7 to obtain feature map 203;

[0074] The obtained feature map 204 is used as input and passed through a depthwise separable convolution D4 of size 9*9 to obtain feature map 204;

[0075] The four feature maps 201, 202, 203, and 204 are averaged at the channel level to obtain four different feature maps of size H×W×1. These four feature maps are then spliced ​​together at the channel level to obtain a spatial weight of size H×W×4.

[0076] The obtained spatial weights are mixed using a convolution of size 7*7 with padding of 4 to obtain a spatial weight of size H×W×4, 205.

[0077] The obtained spatial weights are then reweighted using the Sigmoid function;

[0078] The Sigmoid function is defined by the following formula:

[0079]

[0080] The obtained spatial weights are summed by channels and reweighted with the weight parameters learned by the network to obtain a weighted feature map 206. The weighted feature map 206 is then multiplied with the input feature map 200 to obtain the final output feature 207.

[0081] Example 1:

[0082] A depth-based multi-scale remote sensing target detection algorithm includes the following steps:

[0083] Step S1: Data preprocessing; The downloaded public dataset, such as DOTA or DIOR, is segmented to generate images of size 1024×1024. The number of overlapping pixels in the segmented images is 500 to ensure the continuity of information between adjacent image patches, thereby better capturing the features of the target at different scales and in different contexts.

[0084] Step S2: Model Construction and Training; Construct the proposed deep multi-scale remote sensing target detection model. The model design considers multi-level feature fusion to enhance the ability to identify targets of different sizes. Train the deep multi-scale remote sensing target detection algorithm using the segmented remote sensing dataset. During training, data augmentation techniques, such as random flipping, rotation, and scaling, are employed to improve the model's generalization ability, ultimately yielding the remote sensing target detection model.

[0085] Step S3: Object Detection and Result Generation; The trained object detection algorithm model is used to detect the remote sensing image to be detected and generate the final detected image. During the detection process, the model analyzes the multi-scale features of the image, generates bounding boxes and category labels for the objects, and visualizes the detection results for further analysis and application.

[0086] The above steps complete the construction of a deep multi-scale remote sensing target detection network.

[0087] like Figure 1 As shown, when constructing a deep multi-scale remote sensing target detection algorithm, it is necessary to first construct a deep multi-scale feature extraction module. The detailed construction steps are as follows:

[0088] The input feature map 200 has a size of C×H×W. A separable depthwise convolution D1 with a kernel size of 3 and padding of 1 is applied to the input feature map, resulting in an output feature map 201 with the same size of C×H×W. A separable depthwise convolution D2 with a kernel size of 5 and padding of 2 is then applied to the resulting feature map, resulting in an output feature map 202 with the same size of C×H×W. A separable depthwise convolution D3 with a kernel size of 7 and padding of 3 is then applied to the resulting feature map, resulting in an output feature map 203 with the same size of C×H×W. Finally, a separable depthwise convolution D4 with a kernel size of 9 and padding of 4 is applied to the resulting feature map, resulting in an output feature map 204 with the same size of C×H×W.

[0089] The four feature maps are then fused using weights. Compared to the usual multi-scale approach, we introduce learnable dynamic parameters from the network for fusion. The specific fusion steps are as follows:

[0090] The four feature maps 201, 202, 203, and 204 are averaged at the channel level to obtain four feature maps of different sizes H×W×1. These four feature maps are then concatenated at the channel level to obtain spatial weights of size H×W×4. These spatial weights are then mixed using a 7*7 convolution with padding of 4 to obtain spatial weight 205 of size H×W×4. Finally, these spatial weights are reweighted using the Sigmoid function, which is defined as follows:

[0091] The obtained spatial weights are summed by channels and then reweighted with the weight parameters learned by the network to obtain a weighted feature map 206. The weighted feature map 206 is multiplied with the input feature map 200 to obtain the final output feature 207.

[0092] Constructing a deep multi-scale feature extraction module facilitates the extraction of multi-scale feature information. By stacking separable convolutional layers of different scales and calculating spatial weights, feature maps are weighted and fused, improving the network's ability to understand and represent image features. Re-weighting using the sigmoid function helps strengthen the expressive power of the feature maps and suppress noise. Finally, the weighted feature map is multiplied by the input features to obtain the final output features, preserving effective information and improving network performance and accuracy.

[0093] The overall network construction process is as follows:

[0094] Input a three-color RGB image → downsampling block 101 → feature map 102;

[0095] Feature map 102 → Deep multi-scale feature extraction module 103 → Deep multi-scale feature extraction module 104 → Feature map 105;

[0096] Feature map 105 → downsampling block 106 → feature map 107;

[0097] Feature map 107 → Deep multi-scale feature extraction module 108 → Deep multi-scale feature extraction module 109 → Feature map 110;

[0098] Feature map 110 → downsampling block 111 → feature map 112;

[0099] 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;

[0100] Feature map 117 → downsampling block 118 → feature map 119;

[0101] Feature map 119 → Deep multi-scale feature extraction module 120 → Deep multi-scale feature extraction module 121 → Feature map 122;

[0102] The construction process for the deep multi-scale feature extraction module is as follows:

[0103] Input feature 200 → D1 (3×3 depthwise separable convolution) → feature map 201;

[0104] Feature map 201 → D2 (5×5 depthwise separable convolution) → Feature map 202;

[0105] Feature map 202 → D3 (7×7 depthwise separable convolution) → Feature map 203;

[0106] Feature map 203 → D4 (9×9 depthwise separable convolution) → Feature map 204;

[0107] Feature maps 201, 202, 203, 204 → Average value of channels → Four feature maps of size H×W×1 → Channel splicing → Spatial weight 205;

[0108] Spatial weight 205 → 7*7 convolution (padding with 4) → spatial weight 205;

[0109] Spatial weight 205 → Sigmoid function → reweighting;

[0110] Spatial weight 205 → Channel summation → Weighted feature map 206;

[0111] Weighted feature map 206 × input feature 200 → output feature 207.

[0112] The main contributions of this invention to the field of remote sensing target detection are reflected in algorithm innovation and performance improvement. First, by introducing a deep multi-scale feature extraction module combined with a multi-level feature fusion mechanism, the model can better identify targets at different scales, improving detection accuracy and robustness. Second, in the data preprocessing and model training stages, the use of publicly available dataset segmentation and data augmentation techniques, along with deep separable convolutional layers for feature extraction, effectively enhances the model's generalization ability and performance. Furthermore, the network's learnable dynamic parameters and reweighting mechanism enhance its ability to understand and represent image features, making the algorithm more practical and versatile.

[0113] In summary, the depth multi-scale remote sensing target detection algorithm proposed in this invention comprehensively applies multiple techniques, such as multi-scale feature extraction, data processing, and network optimization. Experimental verification demonstrates good detection accuracy and efficiency, showing broad application prospects and promotional value. This comprehensive algorithm design approach and methodological innovation brings new inspiration and breakthroughs to the research fields of remote sensing image processing and target detection, and is of great significance to related research and practical applications.

[0114] Example 2:

[0115] To explain the feasibility of this scheme, further explanation and illustration will be provided below with experimental data.

[0116] (1) Experimental Setup: The code for this invention is based on the PyTorch framework, and the hardware uses six 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. Furthermore, the proposed deep multi-scale remote sensing target detection network is first pre-trained on ImageNet1K for 300 epochs as pre-training weights. During training, the input image size is set to 1024×1024, and data augmentation strategies such as random rotation are employed.

[0117] (2) Dataset Setup: To verify the universality of the method of this invention, training and validation were performed on three mainstream datasets: The DOTA dataset, Dota-v1.0, is a large-scale open-domain remote sensing image dataset containing 2806 remote sensing images, covering multiple geographical environments in different cities across China. This dataset contains 188,282 target instances of different categories, mainly including aircraft, ships, storage tanks, basketball courts, etc. The DOTA dataset is characterized by small target scale, dense distribution, and irregular shapes, making it suitable for tasks such as target detection and target recognition.

[0118] DOTA-v1.5 is an updated version of the DOTA dataset, containing 2806 remote sensing images and 188,282 target instances of different categories. Compared to previous versions, DOTA-v1.5 has made several improvements and optimizations. First, more remote sensing images have been added, covering more diverse geographical environments in Chinese cities, making the dataset more diverse and comprehensive. Second, the target categories have been expanded, introducing new categories such as bridges and parking lots, increasing the richness of the dataset. Furthermore, annotation accuracy has been optimized, improving the accuracy and reliability of target detection and recognition tasks.

[0119] The DIOR dataset (Detection in ORiented scenes Dataset) is a remote sensing image dataset for detecting rotating targets, containing 2800 high-resolution aerial remote sensing images. The targets in the DIOR dataset mainly include vehicles, ships, and aircraft, which typically exhibit irregular shapes such as rotation or tilt. The characteristic of the DIOR dataset is the rotation of the targets, making it suitable for testing the performance of target detection algorithms in real-world scenes with rotating targets.

[0120] (3) Parameter settings: The AdamW optimizer was used to optimize the network. The network momentum was 0.9 and the weight decay factor was 0.05. The warm-up method was used for warm-up, 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.

[0121] (4) Evaluation index: The main index in this experiment is AP50. AP50 refers to the average precision (AP) calculated using 0.5 as the threshold for IoU (Intersection over Union), which can better reflect the accuracy of the detector.

[0122] Final experimental results:

[0123] Table 1: Quantitative evaluation of state-of-the-art methods on DOTA v1.0

[0124] 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

[0125] Table 2: Quantitative evaluation on DOTAv1.5 using state-of-the-art methods

[0126]

[0127] Table 3: Quantitative evaluation on DIOR-R using state-of-the-art methods

[0128]

[0129] Experimental results on multiple remote sensing target detection datasets demonstrate that the proposed DHCF network significantly outperforms many existing state-of-the-art methods, particularly in small target detection. Firstly, in the DOTAv1.0 dataset evaluation, DHCF achieves a detection accuracy of 81.26 for small vehicles, 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's ability to effectively capture the details of small targets demonstrates its effectiveness in complex scenes.

[0130] On the DOTAv1.5 dataset, DHCF also excels in the small vehicle category, achieving an AP value of 52.61. Although its overall AP50 (73.67) is slightly lower than some top-tier methods, its outstanding performance in specific target categories further demonstrates DHCF's superiority in small target detection. Finally, in comparisons on the DIOR-R dataset, DHCFNet, with an AP50 of 71.95, still surpasses several classic methods, showcasing its reliability in complex remote sensing images.

[0131] In summary, the DHCF network, with its relatively small parameter count and high accuracy, demonstrates powerful capabilities, particularly in small target detection, showcasing the effectiveness of deep multi-scale feature fusion. These results not only validate the design principles of DHCF but also provide strong support and new directions for further research in the field of remote sensing target detection.

Claims

1. A feature extraction algorithm for remote sensing target detection based on depth multi-scale, characterized in that, Includes the following steps: S1. Cut the downloaded public dataset to generate images of size 1024×1024, with the number of overlapping pixels of the cut images being 500. S2. Construct a deep multi-scale remote sensing target detection model. Use the segmented remote sensing dataset to train the deep multi-scale remote sensing target detection algorithm to obtain the remote sensing target detection model. It is divided into the following four stages: S2.1 First stage: For any input RGB three-color image, a downsampling block is first used to obtain feature map 102. Then, feature map 102 is used as input, and two proposed deep multi-scale feature extraction modules are stacked to obtain output feature map 105. S2.2, Second stage, the construction process is the same as the first stage. The final output feature map 105 of the first stage is used as input. First, a downsampling block is used to obtain feature map 107. Then, feature map 107 is used as input, and two proposed deep multi-scale feature extraction modules are stacked to obtain the output 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 from 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 the four proposed deep multi-scale feature extraction modules are stacked to obtain the output 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, firstly using downsampling blocks to obtain feature map 119, then using feature map 119 as input, stacking two proposed deep multi-scale feature extraction modules to obtain the output feature map 122; The four stages of constructing the deep multi-scale remote sensing target detection algorithm include the following sub-steps: The input feature 200 is processed by a 3*3 depthwise separable convolution D1 to obtain feature map 201; The obtained feature map 201 is used as input and passed through a depthwise separable convolution D2 of size 5*5 to obtain feature map 202; The obtained feature map 202 is used as input and passed through a depthwise separable convolution D3 of size 7*7 to obtain feature map 203; The obtained feature map 204 is used as input and passed through a depthwise separable convolution D4 of size 9*9 to obtain feature map 204; The four feature maps 201, 202, 203, and 204 are averaged at the channel level to obtain four different feature maps of size H×W×1. These four feature maps are then spliced ​​together 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 with padding of 4 to obtain a spatial weight of size H×W×4, 205. The obtained spatial weights are then reweighted using the Sigmoid function; The Sigmoid function is defined by the following formula: The obtained spatial weights are summed by channels and reweighted with the weight parameters learned by the network to obtain a weighted feature map 206. The weighted feature map 206 is multiplied with the input feature map 200 to obtain the final output feature 207. 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 depth multi-scale as described in claim 1, characterized in that, The specific steps of S1 are as follows: S1.1 Download the dataset from the official website of the publicly available remote sensing dataset and decompress it. The decompressed dataset is divided into training set, validation set and test set. S1.

2. The decompressed dataset is cropped using a pre-written processing script. The number of overlapping pixels between images is 500. The cropped dataset is divided into training and validation sets, validation sets, and test sets.

Images