Small target detection method in remote sensing images based on resolution saving and network utilization

Abstract

This invention relates to a method for small target detection in remote sensing images based on resolution preservation and utilization networks. The method involves acquiring the original remote sensing image and sequentially performing downsampling layer processing at different feature extraction stages to obtain a downsampling layer-processed image. Convolutional layers with multi-branch, different-scale convolutional kernels are used to extract feature information from each preliminary intermediate feature map, and all extracted feature information is treated as multi-channel information. Then, convolution with a preset-scale convolutional kernel is used to integrate this multi-channel information to obtain integrated multi-scale information. A preset resolution preservation and utilization network module is used to process the integrated multi-scale information to obtain the overall output of the module. Finally, this overall output is processed to obtain the detected small target in the remote sensing image. This method achieves cross-correlation of multi-scale features in the original remote sensing image, refines the representation of small target features, and achieves stable and accurate extraction of small target features from the original remote sensing image.

Description

Technical Field

[0001] This invention relates to the field of optical remote sensing image processing, and in particular to a method for small target detection in remote sensing images based on resolution preservation and utilization networks. Background Technology

[0002] With the rapid development of digitalization and informatization, remote sensing technology has become one of the important means of acquiring information about the Earth's surface. Optical remote sensing images (hereinafter also referred to as optical remote sensing images) play a crucial role in many fields due to their intuitive and rich information characteristics. Among them, small target detection, as an important application scenario of target detection in optical remote sensing images, has received widespread research attention. For example, small target detection in remote sensing images plays an important role in fields such as marine monitoring, ecological monitoring, post-disaster emergency response, and military applications.

[0003] Small target detection in computer vision utilizes pattern recognition techniques and algorithms to accurately detect and identify small, low-contrast, blurred, or partially occluded targets (also known as small targets) in complex environments. Under large-scale optical remote sensing imaging conditions, these small targets face numerous challenges. Compared to medium- and large-scale targets, small targets have extremely small pixel areas, occupying only a handful of points in remote sensing images, making it difficult to effectively extract their features. Furthermore, in large-scale imaging, remote sensing images are easily affected by factors such as atmospheric scattering and uneven illumination, introducing significant noise interference. This noise further obscures the foreground signal of small targets, resulting in extremely poor representation of them.

[0004] However, existing optical remote sensing image target detection methods are mostly based on the identification and matching of salient features of the target. When faced with the detection of small targets with weak foreground signals and indistinct features, they often fail to play an effective role, and their detection and recognition performance drops significantly.

[0005] Therefore, how to accurately detect small targets in optical remote sensing images has become a pressing technical problem that needs to be solved in the field of optical remote sensing image processing. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide a method for detecting small targets in remote sensing images based on resolution saving and network utilization, which can improve the detection accuracy of the above-mentioned prior art.

[0007] The technical solution adopted by this invention to solve the above-mentioned technical problems is: a method for small target detection in remote sensing images based on resolution saving and network utilization, characterized by comprising the following steps:

[0008] Step 1: Acquire the raw remote sensing images that have been collected;

[0009] Step 2: Perform downsampling layer processing on the acquired original remote sensing images at different feature extraction stages in sequence to obtain the images after downsampling layer processing at each feature extraction stage, thereby achieving spatial resolution downsampling under shallow feature extraction.

[0010] Step 3: Use depthwise separable convolution to process the images after downsampling layer processing obtained in each feature extraction stage, and obtain the preliminary intermediate feature maps of the base layer corresponding to each feature extraction stage, so as to extract the effective spatial context information within the minimum neighborhood range related to the small target.

[0011] Step 4: Use convolutional layers with different scale convolutional kernels of multiple branches to extract the feature information of each preliminary intermediate feature map and use all the extracted feature information as multi-channel information. Then, use convolution with a pre-set scale convolutional kernel to integrate the multi-channel information to obtain the integrated multi-scale information.

[0012] Step 5: The integrated multi-scale information is processed by the resolution preservation and utilization network module formed by the encoder-decoder path, the three-source resolution preservation module path and the residual connection path to obtain the overall output of the resolution preservation and utilization network module.

[0013] Step 6: Save the obtained resolution and process the overall output of the network module to obtain the small target detection of the remote sensing image.

[0014] Improved, in the remote sensing image small target detection method based on resolution preservation and network utilization, the downsampling layer in each feature extraction stage includes small depth-separable convolutional layers and regular convolutional layers connected in series.

[0015] Furthermore, in the remote sensing image small target detection method based on resolution preservation and network utilization, in step 2, the calculation method for the downsampling layer processing corresponding to the original remote sensing image is as follows: ;in, Represents the original remote sensing image X The output feature map after processing by the downsampling layer. This indicates a depthwise separable convolution with a kernel size of 3×3. This represents a regular convolution with a kernel size of 1×1. This represents a regular convolution with a kernel size of 3×3, in the original remote sensing image. X The altitude is H, and the original remote sensing image is shown. X The width is W, the original remote sensing image. X The number of channels is 3.

[0016] Further, in the aforementioned method for detecting small targets in remote sensing images based on resolution preservation and network utilization, the expressions for the downsampling layer processing methods at different feature extraction stages are as follows: ; 1≤i≤I; where, This represents the output feature map corresponding to the downsampling layer after the i-th feature extraction stage. The input feature map and output feature map represent the result of the downsampling layer processing in the i-th feature extraction stage. The spatial resolution is the input feature map Half the spatial resolution, output feature map The number of channels is C i Input feature map The number of channels is C i-1 C i =2C i-1 I represents the total number of feature extraction stages.

[0017] Furthermore, in the remote sensing image small target detection method based on resolution preservation and network utilization, in step 3, the corresponding base layers at different feature extraction stages are connected in parallel through multiple medium-sized depth separable convolutional layers to achieve multi-layer feature acquisition and information integration, so as to obtain the multi-scale information.

[0018] Furthermore, in the remote sensing image small target detection method based on resolution preservation and utilization network, the calculation method for the preliminary intermediate feature maps corresponding to different base layers at different feature extraction stages is as follows: ; 1≤n≤ N BL ;in, This represents the preliminary intermediate feature map corresponding to the nth base layer in the i-th feature extraction stage. This represents the input feature map corresponding to the nth base layer in the i-th feature extraction stage. N BL denoted as the number of base layers in the i-th feature extraction stage.

[0019] Improved, in the remote sensing image small target detection method based on resolution saving and network utilization, in step 4, the calculation formula for each channel information in the multi-channel information is as follows:

[0020] ;in, This represents the output feature map of the nth base layer in the i-th feature extraction stage. This represents the input feature map of the nth base layer in the i-th feature extraction stage.

[0021] Furthermore, in the remote sensing image small target detection method based on resolution saving and network utilization, in the encoder-decoder path, the encoder and decoder have a symmetrical structure and extract deep semantic features by repeatedly stacking the encoder and decoder. Each encoder contains one downsampling step, each decoder contains one upsampling step, and the number of encoders and decoders is equal.

[0022] In the path of the three-source resolution preservation module:

[0023] The feature map processed by the base layer is used as the input image, and the input image is decomposed into a first processing stream as a source of receptive field expansion, a second processing stream as a source of spatial preservation, and a third processing stream as a source of downsampling.

[0024] A conventional convolutional layer with a kernel size of 3×3 is used. Expanding the receptive field of a single feature point in the first processing stream generates a first feature processing result; wherein, the first feature processing result is labeled as... ，N × M This indicates the matrix dimension corresponding to the result of the first feature processing;

[0025] A conventional convolutional layer with a kernel size of 1×1 is used. Furthermore, a second feature processing result is generated using the channel information of a single feature point in the second processing stream; wherein, the second feature processing result is labeled as... ;

[0026] A conventional convolutional layer with a kernel size of 3×3 is used. Furthermore, the sampling step size is set to twice the original sampling step size to reduce the feature map spatial size and compress some redundant spatial information, resulting in a third feature map; wherein, the third feature map is labeled as ;

[0027] Use transposed convolution The third feature map The feature map spatial size is increased to match the spatial size of the original input image, resulting in the third feature processing result corresponding to the third feature map; whereby the third feature processing result is labeled as... .

[0028] Furthermore, in the remote sensing image small target detection method based on resolution preservation and utilization network, the path processing procedure of the three-source resolution preservation module is as follows:

[0029] Attention mining processing is performed on the first feature processing result and the second feature processing result to generate a first spatial attention weight; wherein, the first spatial attention weight is labeled as W 1:

[0030] ; ;in, Indicates the feature map Perform channel average pooling. Indicates the feature map Perform max pooling. F sm This represents the sigmoid activation function; This represents a convolutional layer with a size of 1×1;

[0031] Based on the first feature processing result, the second feature processing result, and the first spatial attention weight, an intermediate feature map corresponding to the first spatial attention weight is obtained; wherein, the intermediate feature map corresponding to the first spatial attention weight is labeled as... :

[0032] ;

[0033] The obtained intermediate feature map The result of the third feature processing Perform a fusion process to obtain a fused feature map; the fused feature map is labeled as follows. , ;

[0034] Attention mining is performed on the obtained fused feature map to generate second spatial attention weights; wherein, the second spatial attention weights are labeled as W 2:

[0035] ;

[0036] Based on the obtained intermediate feature map, the third feature processing result, and the second spatial attention weight W 2. Obtain the corresponding attention weights in the second space. W 2. Intermediate feature map; where, corresponding to the second spatial attention weights. W The intermediate feature map of 2 is labeled as : ;

[0037] Attention mining is performed on the obtained intermediate feature map to generate third-space attention weights; wherein, the third-space attention weights are labeled as W 3:

[0038] ;

[0039] Based on the obtained third space attention weight W 2 and input image X N×MThe final output feature map of the three-source resolution-preserving module path is obtained through processing; wherein, the final output feature map is labeled as Y N×M :

[0040] .

[0041] Improved, in the remote sensing image small target detection method based on resolution preservation and utilization network, the output calculation method corresponding to the resolution preservation and utilization network module in each feature extraction stage is as follows:

[0042] ;in, To preserve and utilize the processed output feature map at a higher resolution, σ1 and σ2 are coefficients controlling the three source weights, respectively. φ T This indicates that deep semantic features are fused through pseudo-downsampling and feature associations between multiple scales are established using a three-source attention mechanism. φ E Indicates encoding processing, φ D This indicates decoding processing.

[0043] Furthermore, in the remote sensing image small target detection method based on resolution preservation and utilization network, the overall output calculation method of the resolution preservation and utilization network module is as follows:

[0044] i∈(1,2,…, I );in, Y I To preserve and utilize the overall output of the network module at the required resolution, I This indicates the total number of feature extraction stages.

[0045] Compared with existing technologies, the advantages of this invention are as follows: The method for small target detection in remote sensing images based on resolution preservation and utilization networks acquires raw remote sensing images and sequentially performs downsampling layer processing at different feature extraction stages on these raw images, obtaining images after downsampling layer processing at each feature extraction stage. Convolutional layers with multi-branch convolutional kernels of different scales are used to extract feature information from each preliminary intermediate feature map, and all extracted feature information is used as multi-channel information. Then, convolution with a preset scale kernel is used to integrate this multi-channel information, resulting in integrated multi-scale information. A resolution preservation and utilization network module based on encoder-decoder paths, three-source resolution preservation module paths, and residual connection paths is used to perform resolution preservation and utilization processing on the integrated multi-scale information, obtaining the overall output of the resolution preservation and utilization network module. Finally, this overall output is processed to obtain the detected small target in the remote sensing image. In this way, the cross-correlation of multi-scale features in the raw remote sensing image is realized, gradually refining the small target feature representation and achieving stable and accurate extraction of small target features from the raw remote sensing image. Attached Figure Description

[0046] Figure 1 This is a schematic diagram of the process of a remote sensing image small target detection method based on resolution saving and network utilization in an embodiment of the present invention;

[0047] Figure 2 This diagram illustrates the detection results of small target objects—small vehicles—in a densely distributed scene using both traditional small target detection methods and the remote sensing image small target detection method described in this embodiment of the invention. Detailed Implementation

[0048] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0049] This embodiment provides a method for small target detection in remote sensing images based on resolution saving and network utilization. Specifically, see [link to relevant documentation]. Figure 1 As shown, the method for small target detection in remote sensing images based on resolution saving and network utilization in this embodiment includes the following steps 1 to 6:

[0050] Step 1: Acquire the raw remote sensing imagery; wherein, in this embodiment, the raw remote sensing imagery is labeled as... X Original remote sensing images X The altitude is H, and the original remote sensing image is shown. X The width is W, the original remote sensing image. X The number of channels is 3;

[0051] Step 2 involves sequentially applying downsampling layer processing at different feature extraction stages to the acquired raw remote sensing images, resulting in images processed by downsampling layers at each feature extraction stage. This achieves spatial resolution downsampling under shallow feature extraction. Each downsampling layer at each feature extraction stage comprises a series of small depthwise separable convolutional layers and a regular convolutional layer. Preferably, the downsampling layer is first applied directly to the raw remote sensing image to reduce the overall network parameter count and computational cost. Of course, with a deeper network, the downsampling layer can also be applied to a specific feature extraction stage. In this embodiment, the small depthwise separable convolutional layer is a 3×3 depthwise separable convolutional layer. Regular convolution is a regular convolution with a kernel size of 1×1. Or / and regular convolutions with a kernel size of 3×3. ;

[0052] For example, in step 2, the calculation method for processing the downsampling layer corresponding to the original remote sensing image is as follows: ;in, Represents the original remote sensing image X The output feature map after processing by the downsampling layer. This indicates a depthwise separable convolution with a kernel size of 3×3. This represents a regular convolution with a kernel size of 1×1. This represents a regular convolution with a kernel size of 3×3, in the original remote sensing image. X The altitude is H, and the original remote sensing image is shown. X The width is W, the original remote sensing image. X The number of channels is 3.

[0053] It should be noted that, for different feature extraction stages, in this embodiment, the expressions for the downsampling layer processing methods at different feature extraction stages are as follows: ; 1≤i≤I; where, This represents the output feature map corresponding to the downsampling layer after the i-th feature extraction stage. The input feature map and output feature map represent the result of the downsampling layer processing in the i-th feature extraction stage. The spatial resolution is the input feature map Half the spatial resolution, output feature map The number of channels is C i Input feature map The number of channels is C i-1 C i =2C i-1 , I This represents the total number of feature extraction stages.

[0054] Step 3: The images processed by the downsampling layers obtained in each feature extraction stage are processed using depthwise separable convolutions to obtain preliminary intermediate feature maps of the base layers corresponding to each feature extraction stage. This allows for the extraction of effective spatial context information within the smallest neighborhood related to the small target. For example, in this embodiment, the base layers corresponding to different feature extraction stages are connected in parallel through multiple small to medium-sized depthwise separable convolutional layers to achieve multi-layer feature acquisition and information integration, thereby obtaining the aforementioned multi-scale information.

[0055] Specifically, regarding the calculation method of the preliminary intermediate feature map in step 3 above, in this embodiment, the calculation methods of the preliminary intermediate feature maps corresponding to different base layers under different feature extraction stages are as follows: ; 1≤n≤ N BL ;in, This represents the preliminary intermediate feature map corresponding to the nth base layer in the i-th feature extraction stage. This represents the input feature map corresponding to the nth base layer in the i-th feature extraction stage. N BL denoted as the number of base layers in the i-th feature extraction stage.

[0056] By having the base layer compute preliminary intermediate feature maps, effective spatial context information within the smallest neighborhood range related to small targets can be extracted. Of course, the settings and processing of the base layer do not change the spatial size of the input feature map. The base layer can be reused repeatedly in a single feature extraction stage to improve the multi-scale feature extraction capability of each stage.

[0057] Step 4: Use convolutional layers with different scale convolutional kernels of multiple branches to extract the feature information of each preliminary intermediate feature map and use all the extracted feature information as multi-channel information. Then, use convolution with a pre-set scale convolutional kernel to integrate the multi-channel information to obtain the integrated multi-scale information.

[0058] Specifically, in step 4, the calculation formula for each channel information in the multi-channel information is as follows: ; This represents the output feature map of the nth base layer in the i-th feature extraction stage. This represents the input feature map of the nth base layer in the i-th feature extraction stage.

[0059] Step 5: The resolution preservation and utilization network module formed by the encoder-decoder path, the three-source resolution preservation module path, and the residual connection path is used to perform resolution preservation and utilization processing on the integrated multi-scale information to obtain the overall output of the resolution preservation and utilization network module; wherein, the resolution preservation and utilization network module is set after the base layer, and the number of stacked resolution preservation and utilization network modules is equal to the total number of feature extraction stages in the overall network.

[0060] Step 6: Save the obtained resolution and process the overall output of the network module to obtain the small target detection of the remote sensing image.

[0061] Regarding step 5 above, in the aforementioned encoder-decoder path, the encoder and decoder have a symmetrical structure and extract deep semantic features by repeatedly stacking the encoder and decoder. Each encoder contains one downsampling step, each decoder contains one upsampling step, and the number of encoders and decoders is equal.

[0062] Regarding step 5 above, the processing method in the aforementioned three-source resolution preservation module path is set as follows: steps a1~a5:

[0063] Step a1: The feature map processed by the base layer is used as the input image and the input image is decomposed into a first processing stream as a source of receptive field expansion, a second processing stream as a source of spatial preservation, and a third processing stream as a source of downsampling. Each feature extraction stage contains multiple base layers. The function of the base layers is to enrich the multi-scale information extracted by the network and realize multi-scale feature acquisition and information integration.

[0064] Step a2: Use a conventional convolutional layer with a kernel size of 3×3. Expanding the receptive field of a single feature point in the first processing stream generates a first feature processing result, so as not to reduce the spatial size of the feature map and to effectively utilize local spatial information; wherein, the first feature processing result is labeled as ，N × M This represents the matrix dimension corresponding to the result of the first feature processing.

[0065] Step a3: Use a conventional convolutional layer with a kernel size of 1×1. Furthermore, the second feature processing result is generated using the channel information of a single feature point in the second processing stream, so as not to reduce the spatial size of the feature map and effectively preserve the original spatial information; wherein, the second feature processing result is labeled as ;

[0066] Step a4: Use a conventional convolutional layer with a kernel size of 3×3. Furthermore, the sampling step size is set to twice the original sampling step size to reduce the feature map spatial size and compress some redundant spatial information, resulting in a third feature map; wherein, the third feature map is labeled as ;

[0067] Step a5, use transposed convolution The third feature map The feature map spatial size is increased to match the spatial size of the original input image, resulting in the third feature processing result corresponding to the third feature map; whereby the third feature processing result is labeled as... .

[0068] It should be noted that conventional convolutional layers The setting aims to preserve the spatial information of the feature map to the greatest extent possible, unlike conventional convolutional layers. The setting is to slightly expand the effective receptive field of feature points without changing the size of the feature map space; and to achieve this through a combination of downsampling and upsampling (i.e., conventional convolutional layers). With transposed convolution This can enhance the feature response density and improve the response distribution to local features of large-size remote sensing images.

[0069] Regarding step 5 above, the path processing procedure for the three-source resolution preservation module is as follows: steps b1 to b7:

[0070] Step b1: Perform attention mining processing on the first feature processing result and the second feature processing result to generate a first spatial attention weight; wherein, the first spatial attention weight is labeled as... W 1 :

[0071] ; ;

[0072] in, Indicates the feature map Perform channel average pooling. Indicates the feature map Perform max pooling. F sm This represents the sigmoid activation function; This represents a convolutional layer with a size of 1×1;

[0073] Step b2: Based on the first feature processing result, the second feature processing result, and the first spatial attention weight, obtain the intermediate feature map corresponding to the first spatial attention weight; wherein, the intermediate feature map corresponding to the first spatial attention weight is labeled as... : ;

[0074] Step b3, obtain the intermediate feature map The result of the third feature processing Perform a fusion process to obtain a fused feature map; the fused feature map is labeled as follows. , ;

[0075] Step b4, based on the obtained fused feature map Attention mining is performed to generate second-space attention weights; wherein, these second-space attention weights are labeled as... W 2 :

[0076] ;

[0077] Step b5, based on the obtained intermediate feature map Third feature processing results and second-space attention weights W 2 This yields the corresponding attention weights in the second space. W 2 The intermediate feature map; where, corresponds to the second spatial attention weights. W 2 The intermediate feature map is labeled as : ;

[0078] Step b6: Perform attention mining processing based on the obtained intermediate feature map to generate third-space attention weights; wherein, the third-space attention weights are labeled as... W 3 :

[0079] ;

[0080] Step b7, based on the obtained third-space attention weights W 3 and input image W N×M The final output feature map of the three-source resolution-preserving module path is obtained through processing; wherein, the final output feature map is labeled as Y N×M : .

[0081] Based on the aforementioned encoder-decoder path and three-source resolution preservation module path processing, the output label corresponding to the resolution preservation and utilization network module in each feature extraction stage of this embodiment is set as follows: ;Should The expression is: ;in, To preserve and utilize the processed output feature map at a higher resolution, σ1 and σ2 are coefficients controlling the three source weights, respectively. φ T This indicates that deep semantic features are fused through pseudo-downsampling and feature associations between multiple scales are established using a three-source attention mechanism. φ E Indicates encoding processing, φ D This indicates decoding processing. Specifically, here... φ T It acts as a lateral link, enhancing semantic expression with minimal loss of spatial information, and achieving a balance between positioning accuracy and semantic depth.

[0082] Additionally, it should be noted that the resolution preservation and utilization of the overall output calculation method of the network module are as follows: i∈(1,2,…, I );in, Y I To preserve and utilize the overall output of the network module at the required resolution, I This indicates the total number of feature extraction stages.

[0083] In step 6 of this embodiment, the obtained resolution is saved and the overall output of the network module is processed to obtain the small target detection process of the remote sensing image as follows: steps c1~c2:

[0084] Step c1: Design a 5-layer stacked preset processing module; wherein each preset processing module includes a 3D convolutional group and a ReLU activation function located after the 3D convolutional group; wherein in each preset processing module, the number of channels of the 3D convolutional group is 3×3 and the stride of the 3D convolutional group is 1.

[0085] Step c2: Pre-set the overall objective function of the object detection model and the corresponding deep network; wherein, the overall objective function is set as follows:

[0086] L= L ds ( p, p* )+ L reg (t,t * );

[0087] t=(t x , t y , t w , t h , t θ ) ;

[0088] t x =(xx a ) / wa , t y =(yy a ) / h a ;

[0089] t w =log(w / w a ), t h =log(h / h a ),

[0090] t θ =tan(θ-θ a );

[0091] in, L This represents the loss function value, p is the classification confidence of the corresponding category in the classification branch, and also represents the predicted label of the small target bounding box; p* is the label of the anchor box, 1 for positive samples and 0 for negative samples; L ds (p, p* ) represents the predicted label p of the small target prediction box and the anchor box label. p* The classification loss is between the two; t is the offset vector representing the predicted small target box relative to the anchor box, t * This is the offset vector of the manually labeled box relative to the anchor box. L reg (t,t * ) represents the regression loss for predicting bounding box vectors for small targets, t x This represents the x-coordinate of the center point of the predicted bounding box for the small target. a w represents the x-coordinate of the center point of the anchor frame a t represents the width of the anchor frame. y This represents the y-coordinate of the center point of the predicted bounding box for the small target. a h represents the y-coordinate of the center point of the anchor frame. a t represents the height of the anchor frame. w t represents the width of the predicted bounding box for small targets. h t represents the height of the predicted bounding box for the small target. θ θ represents the angle parameter of the small target prediction box. a This indicates the angle parameter of the anchor frame.

[0092] To verify the detection performance of the remote sensing image small target detection method based on resolution preservation and network utilization in this embodiment, the detection results of this remote sensing image small target detection method are compared with those of traditional remote sensing image small target detection methods. See [link to documentation]. Figure 2 ,lie in Figure 2 The detection result diagram on the left side of the middle image is obtained using traditional remote sensing image small target detection methods, located in... Figure 2The detection effect diagram on the right side shows the detection results obtained using the remote sensing image small target detection method based on resolution saving and network utilization in this embodiment. It can be seen that in scenes with relatively dense target distribution, the remote sensing image small target detection method provided in this embodiment successfully detected a small vehicle next to a large tree (the rectangle in the circle represents the vehicle). It can be seen that its outer edge outline is not particularly clear, but the remote sensing image small target detection method in this embodiment still successfully detected this type of target.

[0093] Although preferred embodiments of the present invention have been described in detail above, it should be clearly understood that various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for detecting small targets in remote sensing images based on resolution preservation and network utilization, characterized in that, Includes the following steps: Step 1: Acquire the raw remote sensing images that have been collected; Step 2: Perform downsampling layer processing on the acquired original remote sensing images at different feature extraction stages in sequence to obtain the images after downsampling layer processing at each feature extraction stage, thereby achieving spatial resolution downsampling under shallow feature extraction. Step 3: Use depthwise separable convolution to process the images after downsampling layer processing obtained in each feature extraction stage, and obtain the preliminary intermediate feature maps of the base layer corresponding to each feature extraction stage, so as to extract the effective spatial context information within the minimum neighborhood range related to the small target. Step 4: Use convolutional layers with different scale convolutional kernels of multiple branches to extract the feature information of each preliminary intermediate feature map and use all the extracted feature information as multi-channel information. Then, use convolution with a pre-set scale convolutional kernel to integrate the multi-channel information to obtain the integrated multi-scale information. Step 5: The integrated multi-scale information is processed by the resolution preservation and utilization network module formed by the encoder-decoder path, the three-source resolution preservation module path and the residual connection path to obtain the overall output of the resolution preservation and utilization network module. Step 6: Save the obtained resolution and process the overall output of the network module to obtain the small target detection in the remote sensing image; wherein: In the encoder-decoder path, the encoder and decoder have a symmetrical structure and extract deep semantic features by repeatedly stacking the encoder and decoder. Each encoder contains one downsampling step, each decoder contains one upsampling step, and the number of encoders and decoders is equal. In the path of the three-source resolution preservation module: The feature map processed by the base layer is used as the input image, and the input image is decomposed into a first processing stream as a source of receptive field expansion, a second processing stream as a source of spatial preservation, and a third processing stream as a source of downsampling. A conventional convolutional layer with a kernel size of 3×3 is used to expand the receptive field of a single feature point in the first processing stream to generate the first feature processing result; A conventional convolutional layer with a kernel size of 1×1 is used, and the channel information of a single feature point in the second processing stream is used to generate the second feature processing result; A conventional convolutional layer with a kernel size of 3×3 is used, and the sampling stride is set to twice the original sampling stride to reduce the feature map spatial size and compress some redundant spatial information to obtain the third feature map. Transposed convolution is used to increase the spatial size of the third feature map to be consistent with the spatial size of the original input image, thus obtaining the third feature processing result of the corresponding third feature map. 2.The resolution-preserving and network-based remote sensing image small target detection method according to claim 1, wherein, The downsampling layer in each feature extraction stage consists of small, depthwise separable convolutional layers and regular convolutional layers connected in series.

3. The method for small target detection in remote sensing images based on resolution saving and network utilization according to claim 2, characterized in that, In step 2, the calculation method of the original remote sensing image corresponding to the down-sampling layer processing is as follows: ; in, Represents the original remote sensing image X The output feature map after processing by the downsampling layer, This indicates a depthwise separable convolution with a kernel size of 3×3. This represents a regular convolution with a kernel size of 1×1. This represents a regular convolution with a kernel size of 3×3, in the original remote sensing image. X The altitude is H, and the original remote sensing image is shown. X The width is W, the original remote sensing image. X The number of channels is 3.

4. The method for small target detection in remote sensing images based on resolution saving and network utilization according to claim 3, characterized in that, The expressions for the downsampling layer processing methods at different feature extraction stages are as follows: ; 1≤i≤I; in, This represents the output feature map corresponding to the downsampling layer after the i-th feature extraction stage. The input feature map and output feature map represent the result of the downsampling layer processing in the i-th feature extraction stage. The spatial resolution is the input feature map Half the spatial resolution, output feature map The number of channels is C i Input feature map The number of channels is C i-1 C i =2C i-1 I represents the total number of feature extraction stages.

5. The method for small target detection in remote sensing images based on resolution saving and utilization networks according to claim 4, characterized in that, In step 3, the base layers corresponding to different feature extraction stages are connected in parallel through multiple small and medium-sized depth separable convolutional layers to achieve multi-layer feature acquisition and information integration, so as to obtain the multi-scale information.

6. The method for small target detection in remote sensing images based on resolution saving and utilization networks according to claim 5, characterized in that, The calculation methods for the preliminary intermediate feature maps corresponding to different base layers in different feature extraction stages are as follows: ; 1≤n≤ N BL ; in, This represents the preliminary intermediate feature map corresponding to the nth base layer in the i-th feature extraction stage. This represents the input feature map corresponding to the nth base layer in the i-th feature extraction stage. N BL denoted as the number of base layers in the i-th feature extraction stage.

7. The method for small target detection in remote sensing images based on resolution saving and utilization networks according to claim 5, characterized in that, In step 4, the calculation formula for each channel information in the multi-channel information is as follows: ； in, This represents the output feature map of the nth base layer in the i-th feature extraction stage. This represents the input feature map of the nth base layer in the i-th feature extraction stage.

Images