Remote sensing target detection method based on multi-scale feature collaboration and semantic complementation

Abstract

The invention discloses a remote sensing target detection method based on multi-scale feature collaboration and semantic complementation, and the method comprises the steps: S1, extracting the features of a given input remote sensing image through employing a backbone network, and generating a multi-level feature map after down-sampling; s2, collecting and mining global semantic information in the multi-level feature map; s3, combining the global multi-scale features and the local hierarchical features, and fusing semantic and spatial context information; s4, fusing significant information of adjacent layers in the input feature map to enhance semantic representation of the output feature map; s5, the classification and positioning decoupling detection head predicts the category and the detection frame of the target from the enhanced feature map; and S6, selecting Distribrute Focus Loss and Complete IoU Loss as regression loss to be used for bounding box regression of the target, wherein the Distribrute Focus Loss and the Complete IoU Loss are used as the regression loss. The method provided by the invention is beneficial to modeling the relationship between the object and the background in the remote sensing image, reduces the influence of the complex background on the accurate target characteristics, adapts to the large-scale change of different remote sensing targets, and improves the detection accuracy.

Description

Technical Field

[0001] The present invention relates to the technical field of remote sensing image recognition and target detection, and in particular to a remote sensing image target detection method based on multi-scale feature collaboration and semantic complementarity. Background Art

[0002] With the rapid development of remote sensing observation technology, many high-resolution remote sensing satellites around the world have acquired a large number of multi-geomorphic, multi-dimensional, and high-resolution remote sensing images. High-resolution remote sensing images contain rich detailed features of objects and have richer spatial structures and topological relationships. Nowadays, remote sensing images have been widely used in meteorology, agriculture, environment, and disaster monitoring, and have extensive practical value. In recent years, the rapid development of deep learning has provided favorable technical support for intelligent information extraction from remote sensing images. In order to effectively promote remote sensing image target detection methods based on deep learning, many institutions and researchers have also publicly released a variety of large-scale remote sensing image recognition-related data sets.

[0003] Compared with common natural scenes, remote sensing and aerial images usually capture objects of arbitrary levels or directions from a bird's-eye view, coupled with different shooting angles and imaging conditions, which makes the detection of targets in multi-scene remote sensing images still challenging. Specifically, there are the following two main challenges that need to be addressed.

[0004] Complex background interference problem: The foreground target in the remote sensing image usually accounts for a small proportion and is easily interfered by the surrounding complex background such as buildings and vegetation, which makes the features of the target object blurred and is not conducive to algorithm recognition.

[0005] The problem of large-scale changes in targets: Remote sensing images are usually captured from satellites or drones with a wide field of view, so the target scale varies with the image resolution. Huge scale changes often hinder accurate target detection.

[0006] In order to solve the problem of complex background interfering with foreground targets, many current works focus on exploring different methods to improve detection performance by using semantic information. Semantic information helps to model the relationship between objects and backgrounds in remote sensing images and reduce the impact of complex backgrounds on precise target features. Currently, many methods extract semantic information through feature fusion, semantic encoding or attention mechanism. However, these methods do not take into account that the semantic gap between adjacent feature layers is small, and the semantic information of the current output layer can be supplemented by aggregating the significant features of adjacent layers.

[0007] To address the large-scale variation of targets, most current research focuses on obtaining multi-scale information by focusing on feature pyramids and their variant structures, spatial and channel attention, and cross-scale feature fusion. However, there are too many paths in a network using a feature pyramid structure, and non-adjacent layers can only interact indirectly, resulting in information loss during cross-layer information interaction, which prevents the network from better information fusion. Summary of the invention

[0008] The purpose of the present invention is to solve the technical problems that the detection accuracy of remote sensing image target detection is easily affected by the interference of complex background on foreground targets and the target scale usually changes with the image resolution in existing remote sensing image target detection methods, and to propose a remote sensing target detection method based on multi-scale feature collaboration and semantic complementarity.

[0009] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0010] A remote sensing target detection method based on multi-scale feature collaboration and semantic complementarity includes the following steps:

[0011] Step S1: Use the backbone network to extract the given input remote sensing image Features, generate downsampled multi-level feature maps where i∈{1, 2, 3, 4, 5}, H and W represent the height and width of the input image, H i ×W i Represents the feature map S i Dimensions (H / 2 ∧i ×W / 2 ∧i );

[0012] Step S2: Collect and mine the multi-level feature maps through GMI. 2 ,S 3 ,S 4 ,S 5}, thereby obtaining the multi-scale feature F containing global information fuse ;

[0013] Step S3: Use SAM to coordinate global multi-scale features F fuse and local level features S i , after exchanging fusion information, {P 2 ,P 3 ,P 4 ,P 5};

[0014] Step S4: Fusion of feature maps by SCM {P 2 ,P 3 ,P 4,P 5}, the semantic representation of the output feature map is enhanced by the significant information of the adjacent layers, and the output feature map {O 3 ,O 4 ,O 5};

[0015] Step S5: The detection head decoupled from classification and localization predicts the category and detection box of the target from the enhanced feature map respectively;

[0016] Step S6: Use cross entropy loss as the classification loss for target classification, and select Distributed FocalLoss and Complete IoU Loss as the regression loss for target bounding box regression.

[0017] In step S2, the following steps are included:

[0018] Step S21: The three-layer feature map {S 2 ,S 3 ,S 4} by downsampling to S 5 After the same unified size is concatenated in the channel dimension and scaled by r times, the global feature F is obtained; the process uses the following formula:

[0019] S' i = downsample(Si),i∈{2,3,4}

[0020] F=Conv 1×1 ([S' 2 ,S' 3 ,S' 4 ,S 5 ])

[0021] Where downsample(·) represents the downsampling operation, which is implemented by average pooling, and [·,·] represents the cascade operation along the channel dimension;

[0022] Step S22: Using two 3×3 convolutions and two 5×5 convolutions with expansion rates R={1,2,2,3} in parallel on the global feature F, feature maps with different receptive field sizes are obtained respectively; the feature maps with different receptive fields are fused to obtain a richer multi-scale feature F. ms ; The process uses the following formula:

[0023] F ms =Conv 1×1 ([F,F' 1 ,F' 2 ,F' 3 ,F' 4 ]);

[0024] Step S23: Introduce the transformer layer to transform the multi-scale features F ms The long-distance global dependency of F is modeled to further extract and fuse feature information. First, the linear layer is replaced by a 1×1 convolutional layer. ms The projection operation is performed to obtain the query Q, key K and value V; the process uses the following formula:

[0025] Q, K, V = Flatten (F ms );

[0026] Step S24: Then perform the multi-head self-attention operation and residual connection in the transformer layer on Q, K, and V to obtain F attn , the multi-head self-attention operation can weight the features to improve the representation ability of the model; the process uses the following formula:

[0027] F attn =Residual(Attn(Q,K,V));

[0028] Step S25: Finally, in F attn Perform feedforward network and residual connection operations to enhance the nonlinear extraction capability of the network and adjust the feature dimension back to obtain the global multi-scale feature F fuse ; The process uses the following formula:

[0029] F fuse =Reshape(Residual(MLP(F attn )))

[0030] Where MLP(·) represents multi-head self-attention and Residual(·) represents residual connection.

[0031] In step S3, the following steps are included:

[0032] Step S31: First use Conv 1×1 The F obtained in step S2 fuse The channel is enlarged according to the scaling factor r given in step S21, and then the local feature {S 2 ,S 3 ,S 4 ,S 5} is divided into G i = {G 2 ,G 3 ,G 4 ,G 5}; The process uses the following formula:

[0033] [G 2 ,G 3,G 4 ,G 5 ]=Split(Conv 1×1 (F fuse ));

[0034] Step S32: Then G i Upsample to S in step S1 i The size of G′ i , let G′ i and The process uses the following formula:

[0035] G′ i =BiLinear(G i )

[0036] Where BiLinear(·) represents the upsampling operation based on bilinear interpolation;

[0037] Step S33: Finally, using G′ i As a weight to measure cross-scale interactions, this weight is based on the backbone network feature S i The auxiliary residual of P is obtained by operation. i :

[0038]

[0039] in and represent element-wise multiplication and element-wise addition, respectively. As the output of the semantic alignment module, it participates in subsequent operations.

[0040] In step S3, the following steps are included:

[0041] Step S41: The semantic complementation module receives the output feature map P of step S33 i = {P 2 ,P 3 ,P 4 ,P 5} as input;

[0042] Step S42: First, fuse the highest-level feature map P 4 and feature map P 5 Get the feature map O 5 , where the feature map P 5 As the global information adjustment channel, it is used as the weight filtering feature map P 4 The redundant information of the network is further extracted by using the Repblock convolutional block. The process uses the following formula:

[0043]

[0044] Step S43: Next, the Fusion structure is used to hierarchically fuse the features of adjacent levels. 4 ,P 3 ,O 5}Get feature map O 4 , and then merge {P 3 ,P 2 ,O 4}Get feature map O 3 ; The process uses the following formula:

[0045] O i =Fusion(O i+1 ,P i ,P i-1 ),i={3,4}

[0046] Step S44: Using the current level feature P i and the next level feature P i-1 Generate P as a local feature local , and then use the O of the previous level Fusion block i+1 As global information, we get O global ; The process uses the following formula::

[0047] P local =Conv 1×1 ([P i ,downsample(P i-1 )])

[0048] O global =BiLinear(Conv 1×1 (O i+1 ))

[0049] Step S45: Use attention operation to fuse global information O global and local information P local , filter out the interference information of local features, and finally further extract and fuse information through the re-parameterized convolution block RepBlock; the information fusion process can be calculated as follows:

[0050] O act =BiLinear(Sigmoid(Conv 1×1 (O i+1 )))

[0051]

[0052] Among them, act represents the attention weight, and Represent element-wise multiplication and element-wise addition, respectively.

[0053] In step S6, Distributed Focal Loss and Complete IoU Loss are selected as regression losses for bounding box regression of the target, and the loss function used is:

[0054]

[0055] In order to balance the loss, a corresponding weight is assigned to each loss, expressed as λ 1 ,λ 2 and λ 3 ; L cls represents the classification loss, L dfl and L ciou They represent Distributed Focal Loss and Complete IoU Loss respectively, which together constitute the regression loss.

[0056] Compared with the prior art, the present invention has the following technical effects:

[0057] 1) The present invention provides a new end-to-end single-stage target detection network for remote sensing images, which can make full use of multi-scale information and contextual information of adjacent features to achieve high-precision detection of remote sensing targets;

[0058] 2) The present invention proposes a global multi-scale interaction structure for large-scale changes in remote sensing targets. The structure collects and mines the global multi-scale semantic information of the multi-level feature graph of the backbone network, realizes the information exchange of global and local context information, and can accurately detect targets of different scales in remote sensing images;

[0059] 3) The present invention proposes a semantic complementation module, which utilizes the semantic similarity of adjacent feature maps and enhances the semantic representation of the output feature map by fusing the significant information of adjacent layer features. This greatly overcomes the redundant interference of complex background in remote sensing images on foreground targets and further refines target features. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] The present invention will be further described below in conjunction with the accompanying drawings and embodiments:

[0061] Figure 1 is a schematic diagram of a flow chart of a remote sensing image target detection method according to an embodiment of the present invention;

[0062] Figure 2 It is the overall structure diagram of the remote sensing target detection model based on multi-scale feature collaboration and semantic complementarity proposed in the present invention;

[0063] Figure 3 It is a structural diagram of Fusion in the semantic complementarity module mentioned in the present invention;

[0064] Figure 4 is a schematic diagram of a flow chart of a global multi-scale interaction structure according to an embodiment of the present invention;

[0065] Figure 5 is a flow chart of a semantic alignment module according to an embodiment of the present invention;

[0066] Figure 6 It is a flowchart of the semantic complementation module of an embodiment of the present invention. DETAILED DESCRIPTION

[0067] like Figure 1 As shown, a remote sensing target detection method based on multi-scale feature collaboration and semantic complementarity includes the following steps:

[0068] S1. Using CSPDarknet53 backbone network to extract given input remote sensing image features and generate multi-level feature maps after 2, 4, 8, 16 and 32 times downsampling where i∈{1, 2, 3, 4, 5}, H and W represent the height and width of the input image (640×640), H i ×W i Represents the feature map S i Dimensions (H / 2 ∧i ×W / 2 ∧i );

[0069] S2. The four stages after GMI collects and mines the above multi-level feature maps {S 2 ,S 3 ,S 4 ,S 5}, thereby obtaining multi-scale features containing global information, and finally scaling F through 1×1 convolution channels. fuse , reduce the amount of calculation of the model;

[0070] S3. Using SAM to coordinate global multi-scale features F fuse and local level features S i , after exchanging fusion information, {P 2 ,P 3 ,P 4 ,P 5}, where P i and The semantic alignment module can alleviate the semantic gap before and after the multi-scale information operation;

[0071] S4. Fusion of feature maps by SCM {P 2 ,P 3 ,P 4 ,P 5}, enhance the semantic representation of the output feature map, obtain more robust and discriminative features with global semantic relationships, and reduce background interference. The output feature map {O 3 ,O 4 ,O 5};

[0072] S5. The detection head decoupled from classification and localization predicts the target category and detection box from the enhanced feature map respectively;

[0073] S6. Cross entropy loss is used as the classification loss for target classification, and Distributed Focal Loss and Complete IoU Loss are selected as regression losses for target bounding box regression.

[0074] like Figure 2 As shown in the figure, at a macro level, the remote sensing target detection network based on multi-scale feature collaboration and semantic complementarity consists of a backbone network, a global multi-scale interaction structure (GMI), a semantic alignment module (SAM), a semantic complementary module (SCM), and a decoupled detection head. After the remote sensing image is input into the backbone network, the GMI structure collects the last four layers of features of the backbone network to extract global multi-scale information, and then SAM collaborates with global features and local features to exchange information. Next, the SCM module aggregates features of adjacent layers to supplement semantic information. Finally, the classification and localization decoupled detection head predicts the category and detection box of the target to obtain the final detection result.

[0075] In a preferred embodiment, the process of step S2 is as follows: Figure 4 As shown, the following steps are included:

[0076] S21: The three-layer feature map of the backbone network {S 2 ,S 3 ,S 4} by downsampling to S 5 After the same unified size is concatenated in the channel dimension and scaled by r times, the global feature F is obtained:

[0077] S' i = downsample(S i ),i∈{2,3,4}

[0078] F=Conv 1×1 ([S' 2 ,S' 3 ,S'4 ,S 5 ])

[0079] Here, downsample(·) represents the downsampling operation, which is implemented by average pooling, and [·,·] represents the cascade operation along the channel dimension.

[0080] S22: On the global feature F, two 3×3 convolutions and two 5×5 convolutions with expansion rates R = {1, 2, 2, 3} are used in parallel to obtain feature maps {F′ of different receptive field sizes. 1 ,F′ 2 ,F′ 3 ,F′ 4}. The feature maps of different receptive fields are fused to obtain richer multi-scale features F ms :

[0081] F ms =Conv 1×1 ([F,F' 1 ,F' 2 ,F' 3 ,F' 4 ])

[0082] S23: Introduce the transformer layer to multi-scale features F ms The long-distance global dependency of F is modeled to further extract and fuse feature information. First, the linear layer is replaced by a 1×1 convolutional layer. ms Perform a projection operation to get the query (Q), key (K), and value (V):

[0083] Q, K, V = Flatten (F ms )

[0084] S24: Then perform the multi-head self-attention operation and residual connection in the transformer layer on Q, K, V to obtain F attn , the multi-head self-attention operation can weight the features to improve the representation ability of the model. The calculation process is described as follows:

[0085] F attn =Residual(Attn(Q,K,V))

[0086] S25: Final fight with F attn Execute the feedforward network and residual connection to enhance the nonlinear extraction ability of the network, and adjust the dimension of the feature back to obtain the global multi-scale feature F fuse :

[0087] F fuse =Reshape(Residual(MLP(F attn)))

[0088] Where MLP(·) represents multi-head self-attention and Residual(·) represents residual connection.

[0089] In a preferred embodiment of the present invention, Figure 5 As shown, step S3 includes the following steps:

[0090] S31: First use Conv 1×1 The F obtained in step S2 above fuse The channel is enlarged by the scaling factor r given in S21, and then the local feature {S 2 ,S 3 ,S 4 ,S 5} is divided into G i = {G 2 ,G 3 ,G 4 ,G 5}:

[0091] [G 2 ,G 3 ,G 4 ,G 5 ]=Split(Conv 1×1 (F fuse ))

[0092] S32: Then G i Upsample to S1 step S i The size of G′ i , let G′ i and The calculation is as follows:

[0093] G' i =BiLinear(G i )

[0094] Where BiLinear(·) represents an upsampling operation based on bilinear interpolation.

[0095] S33: Finally, using G′ i As a weight to measure cross-scale interactions, this weight is based on the backbone network feature S i The auxiliary residual of P is obtained by operation. i :

[0096]

[0097] in and represent element-wise multiplication and element-wise addition, respectively. As the output of the semantic alignment module, it participates in subsequent operations.

[0098] In a preferred embodiment, the step S4 is as follows: Figure 6 The following steps are shown:

[0099] S41: The semantic complementation module receives the output P of step S33 i = {P 2 ,P 3 ,P 4 ,P 5} as input, where

[0100] S42: First merge the top layer P 4 and P 5 Get O 5 , where P 5 As global information adjustment channel and then as weight filter P 4 The redundant information of the network is further extracted by using the Repblock convolutional block. The overall calculation is as follows:

[0101]

[0102] S43: Next, the Fusion structure is used to hierarchically fuse the features of adjacent levels. 4 ,P 3 ,O 5}Get O 4 , and then merge {P 3 ,P 2 ,O 4}Get O 3 The formula is as follows:

[0103] O i =Fusion(O i+1 ,P i ,P i-1 ),i={3,4}

[0104] S44: Fusion structure Figure 3 As shown, using the current level feature P i and the next level feature P i-1 Generate P as a local feature local , and then use the O of the previous level Fusion block i+1 As global information, we get O global The formula is as follows:

[0105] P local =Conv 1×1 ([Pi ,downsample(P i-1 )])

[0106] O global =BiLinear(Conv 1×1 (O i+1 ))

[0107] S45: Finally, the attention operation is used to fuse the global information O global and local information P local , filter out the interference information of local features, and finally further extract and fuse information through the re-parameterized convolution block (RepBlock). The information fusion process can be calculated as follows:

[0108] O act =BiLinear(Sigmoid(Conv 1×1 (O i+1 )))

[0109]

[0110] Among them, act represents the attention weight, and Represent element-wise multiplication and element-wise addition, respectively.

[0111] The loss function in step S6 is:

[0112]

[0113] In order to balance the loss, a corresponding weight is assigned to each loss, expressed as λ 1 ,λ 2 and λ 3 ; L cls represents the classification loss, L dfl and L ciou They represent Distributed Focal Loss and Complete IoU Loss respectively, which together constitute the regression loss;

[0114] In order to verify the effectiveness of the method, this patent was experimented and verified on the DIOR, HRRSD and NWPU VHR-10 datasets. This patent uses three evaluation indicators: precision (P), recall (R), and mean average precision (mAP) to reflect the effectiveness of the patent method. Precision (P) reflects the proportion of the number of positive samples correctly predicted by the model to all predicted positive samples, and recall (R) represents the proportion of the number of correctly predicted positive samples to all actual positive samples. The corresponding calculation formula is as follows:

[0115]

[0116]

[0117]

[0118]

[0119] Among them, TP represents the number of positive samples predicted correctly by the model, FP represents the number of positive samples predicted incorrectly, TN represents the number of negative samples predicted correctly, FN represents the number of negative samples predicted incorrectly, and AP can be defined as the integral area of ​​the precision value from 0 to 1 in all recall values. For multi-class targets, mAP represents the mean of AP under all categories. The experimental results of this patent are as follows:

[0120]

[0121] From the above results, we can see that good detection results were achieved on the DIOR, HRRSD and NWPU VHR-10 datasets.

Claims

1. A remote sensing target detection method based on multi-scale feature collaboration and semantic complementarity, characterized in that: The following steps are involved: Step S1: Use the backbone network to extract the given input remote sensing image Features, generate downsampled multi-level feature maps where i∈{1, 2, 3, 4, 5}, H and W represent the height and width of the input image, H i ×W i Represents the feature map S i Dimensions (H / 2 ∧i ×W / 2 ∧i ); Step S2: Collect and mine the global semantic information in the four stages {S2, S3, S4, S5} after the multi-level feature map to obtain a multi-scale feature F containing global information fuse ; Step S3: Use the semantic alignment module SAM to coordinate the global multi-scale feature F fuse and local level features S i , after exchanging fusion information, {P2, P3, P4, P5} is generated; Step S4: The semantic representation of the output feature map is enhanced by fusing the significant information of adjacent layers in the feature map {P2, P3, P4, P5} through the semantic complementation module SCM to obtain the output feature map {O3, O4, O5}; Step S5: The detection head decoupled from classification and localization predicts the category and detection box of the target from the enhanced feature map respectively; Step S6: Use cross entropy loss as the classification loss for target classification, and select Distributed Focal Loss and Complete IoU Loss as regression losses for target bounding box regression.

2. The method according to claim 1, characterized in that In step S2, the following steps are included: Step S21: The three-layer feature maps {S2, S3, S4} of the backbone network are downsampled to the same uniform size as S5, concatenated in the channel dimension, and scaled by r times to obtain the global feature F; the process uses the following formula: S′ i =downsample(S i ),i∈{2,3,4}; F=Conv 1×1 ([S′2,S′3,S′4,S5]); Where downsample(·) represents the downsampling operation, which is implemented by average pooling, and [·,·] represents the cascade operation along the channel dimension; Step S22: Using two 3×3 convolutions and two 5×5 convolutions with expansion rates R={1,2,2,3} in parallel on the global feature F, feature maps with different receptive field sizes are obtained respectively; the feature maps with different receptive fields are fused to obtain a richer multi-scale feature F. ms ; The process uses the following formula: F ms =Conv 1×1 ([F,F′1,F′2,F′3,F′4]); Step S23: Introduce the transformer layer to transform the multi-scale features F ms The long-distance global dependency of F is modeled to further extract and fuse feature information. First, the linear layer is replaced by a 1×1 convolutional layer. ms The projection operation is performed to obtain the query Q, key K and value V; the process uses the following formula: Q、K、V=Flatten(F ms ); Step S24: Then perform the multi-head self-attention operation and residual connection in the transformer layer on Q, K, and V to obtain F attn , the multi-head self-attention operation can weight the features to improve the representation ability of the model; the process uses the following formula: F attn =Residual(Attn(Q,K,V)); Step S25: Finally, in F attn Perform feedforward network and residual connection operations to enhance the nonlinear extraction capability of the network and adjust the feature dimension back to obtain the global multi-scale feature F fuse ; The process uses the following formula: F fuse =Reshape(Residual(MLP(F attn ))); Where MLP(·) represents multi-head self-attention and Residual(·) represents residual connection.

3. The method according to claim 2, characterized in that In step S3, the following steps are included: Step S31: First use Conv 1×1 The F obtained in step S2 fuse The channel is enlarged according to the scaling factor r given in step S21, and then divided according to the channel dimension of the local features {S2, S3, S4, S5} to obtain G i ={G2,G3,G4,G5}; The process uses the following formula: [G2,G3,G4,G5]=Split(Conv 1×1 (F fuse )); Step S32: Then G i Upsample to S in step S1 i The size of G′ i , let G′ i and The process uses the following formula: G′ i =BiLinear(G i ); Where BiLinear(·) represents the upsampling operation based on bilinear interpolation; Step S33: Finally, using G′ i As a weight to measure cross-scale interactions, this weight is based on the backbone network feature S i The auxiliary residual of P is obtained by operation. i : in and represent element-wise multiplication and element-wise addition, respectively. As the output of the semantic alignment module, it participates in subsequent operations.

4. The method according to claim 3, characterized in that In step S3, the following steps are included: Step S41: The semantic complementation module receives the output feature map P of step S33 i ={P2,P3,P4,P5} as input; Step S42: First, the highest-level feature map P4 and feature map P5 are fused to obtain feature map O5, wherein feature map P5 is used as a global information adjustment channel and then as a weight to filter the redundant information of feature map P4, and the Repblock convolution block is used to further extract information. The process uses the following formula: Step S43: Next, the Fusion structure is used to hierarchically fuse the features of adjacent levels {P4, P3, O5} to obtain the feature map O4, and then fuse {P3, P2, O4} to obtain the feature map O3; the process uses the following formula: O i =Fusion(O i+1 ,P i ,P i-1 ),i={3,4}; Step S44: Using the current level feature P i and the next level feature P i-1 Generate P as a local feature local , and then use the O of the previous level Fusion block i+1 As global information, we get O global ; The process uses the following formula:: P local =Conv 1×1 ([P i ,downsample(P i-1 )]); O global =BiLinear(Conv 1×1 (O i+1 )); Step S45: Use attention operation to fuse global information O global and local information P local , filter out the interference information of local features, and finally further extract and fuse information through the re-parameterized convolution block RepBlock; the information fusion process can be calculated as follows: O act =BiLinear(Sigmoid(Conv 1×1 (O i+1 ))); Among them, act represents the attention weight, and Represent element-wise multiplication and element-wise addition, respectively.

5. The method according to claim 1, characterized in that In step S6, Distributed FocalLoss and Complete IoU Loss are selected as regression losses for bounding box regression of the target, and the loss function used is: In order to balance the loss, each loss is assigned a corresponding weight, expressed as λ1, λ2 and λ3; L cls represents the classification loss, L dfl and L ciou They represent Distributed Focal Loss and Complete IoU Loss respectively, which together constitute the regression loss; the classification loss L cls In the above equation, L represents label confidence and P represents prediction confidence. In addition, L dfl The y in i and i+1 Indicates the closest value on the left and right sides of the predicted value y, satisfying y i <y<y i+1 , S i and S i+1 The calculation is as follows: Finally, L ciou The IoU in the equation represents the intersection over union of the target prediction box and the real box, b and b gt Represent the center points of the predicted box and the real box respectively, ρ is used to measure the Euclidean distance between the two center points, c represents the diagonal distance of the minimum closure area that contains both the predicted box and the real box, and v is used to measure the similarity of the aspect ratio between the two boxes, which can be expressed as: Among them, w, h and w gt 、h gt Represent the width and height of the predicted box and the true box respectively.

Images