A detection method for boundary enhancement and small target perception of a problem map

Abstract

The present application relates to the technical field of target detection, and especially provides a detection method for boundary enhancement and small target perception of a problem map. The method comprises the following steps: constructing a double-branch attention convolution module fusing convolution and an attention mechanism, and enhancing boundary perception ability through a global-local feature cooperative enhancement mechanism; designing a multi-path feature aggregation mechanism, dynamically reconstructing multi-scale feature representation through bidirectional path interaction and adaptive weighted fusion; constructing a global context fusion module to balance global information and local information; and refining a target query through a prediction head and generating a category label and a boundary box. The method enhances the overall boundary perception ability, ensures the sufficiency and effectiveness of feature aggregation, improves the positioning accuracy of the target, and improves the accuracy of border area identification.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of target detection, and particularly relates to a detection method for boundary enhancement and small target perception of a problem map. BACKGROUND

[0002] In recent years, target detection technology has been applied to the detection of problem maps. However, due to the special nature of map data, detection still faces many challenges and it is difficult to achieve ideal detection results. Specifically, scale changes and complex background interference can cause discontinuity of boundary features (such as national boundary lines), thereby reducing recognition accuracy. In addition, small targets such as islands are prone to be ignored or confused with the background due to the small number of pixels, local texture blur, and unclear edge details, resulting in missed detection or false detection. Therefore, it is necessary to design a detection model for the characteristics of map data to effectively extract and enhance edge information, thereby improving the detection accuracy of complex boundaries and small targets.

[0003] Attention mechanisms play a key role in enhancing edge representations, but traditional methods are difficult to adapt to complex boundary shapes and suppress background noise, often leading to incomplete or unstable edge extraction. In addition, single-scale feature extraction cannot handle targets of different sizes and shapes simultaneously, which limits the performance of the model in target recognition tasks. Furthermore, single-scale methods also have shortcomings in capturing the overall structure of large targets, leading to decreased positioning accuracy. Multi-scale feature fusion is one of the key technologies to address target detection challenges such as insufficient feature representation and low model positioning accuracy. In existing technologies, feature pyramid networks are used to integrate features of different levels layer by layer through upsampling to achieve multi-scale fusion. However, a single top-down path may ignore the feedback effect of low-level detail features on high-level semantic features. In addition, high-resolution networks proposed in existing technologies achieve multi-resolution feature fusion through feature interaction, but in the fusion process, low-resolution feature maps may introduce redundant information, affecting fusion performance.

[0004] Methods combining scale-oriented modeling and attention mechanisms also show certain advantages in enhancing small target features. For example, existing methods propose adaptive feature fusion, which dynamically adjusts features between different scales through learnable weights, thereby enhancing the representation ability of small targets. Meanwhile, methods that apply dilated convolution to expand the receptive field while maintaining high resolution effectively preserve the fine detail information of small targets. However, these methods still face challenges in practical applications, such as alignment errors in feature maps and the grid effect caused by dilated convolution, which can cause holes and breaks in feature maps, thereby affecting the complete detection and feature extraction of small targets. SUMMARY

[0005] Therefore, the application provides a detection method for boundary enhancement and small target perception of a problem map, which can enhance the overall boundary perception ability, ensure the sufficiency and effectiveness of feature aggregation, improve the positioning accuracy of the target and the accuracy of the identification of the national border area.

[0006] In a first aspect, the application provides a detection method for boundary enhancement and small target perception of a problem map, which comprises the following steps:

[0007] Step 1: a double-branch attention convolution module fusing convolution and attention mechanism is constructed, and a global-local feature cooperative enhancement mechanism is used to enhance the boundary perception ability;

[0008] Step 2: according to step 1, a multi-path feature aggregation mechanism is designed, and a bidirectional path interaction and adaptive weighted fusion are used to dynamically reconstruct a multi-scale feature representation;

[0009] Step 3: according to step 2, a global context fusion module is constructed to balance the global information and the local information;

[0010] Step 4: according to step 3, a target query is refined by a prediction head, and a category label and a boundary box are generated.

[0011] Optionally, the step 1 comprises:

[0012] A global correlation modeling path: firstly, 1×1 and 3×3 convolution operations are used on the input feature to generate query Q, key K and value V tensors; then, an attention map is calculated by a weighted filtering strategy to realize the interaction between the local features; subsequently, a weighted product operation is performed on the attention map and the value V to obtain a final output feature map , which is combined with a dynamic convolution to optimize the modeling effect of the edge; wherein, The expression of

[0013] (1)

[0014] wherein, X∈ represents an input feature map, ∈ are 3×3 depth separable convolution kernels of the query Q, the key K and the value V, respectively; is a scaling factor, and = C ; wherein, H represents the height of the feature map, W represents the width of the feature map, and C represents the channel number of the feature map;

[0015] Local channel interaction path: first, the input features are compressed and reconstructed by 1x1 convolution operation; then, the input features are divided into N subgroups along the channel dimension, and each subgroup extracts fine-grained local information through depth separable convolution; then, channel rearrangement is introduced to rearrange the channel index, and dynamic convolution and random dropout are combined;

[0016] Finally, the results of the global correlation modeling path and the local channel interaction path are fused after element-wise addition and summation to obtain the output features.

[0017] Optionally, the step 2 comprises:

[0018] The multi-path feature aggregation mechanism passes deep semantic information to shallow features through expression (2) to enhance semantic perception ability, and the form of expression (2) is:

[0019] (2);

[0020] wherein, represents the output feature map of the i-th layer after fusion; represents the feature of the i-th layer; and represents the weight for adjusting the contribution ratio of the up-sampling feature and the current layer feature respectively; function represents up-sampling the feature of the i+1-th layer to match the resolution of the i-th layer feature;

[0021] At the same time, the fine-grained information in the shallow feature is fed back to the deep feature by propagating the feature information from bottom to top, and the process is expressed as:

[0022] (3);

[0023] wherein, function represents down-sampling the feature of the i-1-th layer to match the resolution of the i-th layer feature;

[0024] In the part of bidirectional path interaction, the multi-path feature aggregation mechanism introduces adaptive weighted fusion; through expression (4), the weights from two different scale features are dynamically adjusted to realize feature fusion, and finally the fusion output O is obtained, and the form of expression (4) is:

[0025] (4);

[0026] wherein, represents the feature map of the first scale path; represents the feature map of the second scale path; represents a constant, which takes the value of 1x10 -4 .

[0027] Optionally, the step 3 comprises:

[0028] The global context fusion module models the features of different levels through a multi-branch structure. First, the global context fusion module combines the global average pooling GAP and the global maximum pooling GMP to extract the global information of the input features through expression (5), which is in the form of expression (5):

[0029] (5);

[0030] wherein, , ; H represents the height of the feature map, W represents the width of the feature map, represents the feature value at the spatial position (i, j); and respectively represent the weights of the linear transformation of the features extracted by the global average pooling and the global maximum pooling; represents the global distribution feature extracted by the global average pooling, represents the significant region feature extracted by the global maximum pooling; after weighted summation by the attention weights generated by the activation function Softmax, the global feature output is obtained.

[0031] Secondly, the global context fusion module introduces a self-attention mechanism, generates query Q and key K through 1×1 convolution, and performs nonlinear mapping through an activation function to improve the feature expression capability, while the value V is directly obtained from the input features; then, the query Q, the key K and the value V are multiplied element by element to generate the enhanced global feature representation, and the process is expressed as expression (6):

[0032] (6);

[0033] wherein, represents the global enhanced feature output generated by the self-attention mechanism; , are 1×1 convolution kernels of the query Q and the key K, respectively;

[0034] Finally, the processed by the 1×1 convolution is fused with for prediction.

[0035] In a second aspect, an embodiment of the present application provides a computer readable storage medium, the computer readable storage medium comprising a stored program, wherein the program, when executed, controls a device in which the computer readable storage medium is located to perform the method for boundary enhancement and small target perception detection of a problem map in the first aspect or any possible implementation manner of the first aspect.

[0036] In a third aspect, an embodiment of the present application provides an electronic device, comprising: one or more processors; a memory; and one or more computer programs, wherein the one or more computer programs are stored in the memory, and the one or more computer programs comprise instructions that, when executed by the device, cause the device to perform the method for boundary enhancement and small target perception detection of a problem map in the first aspect or any possible implementation manner of the first aspect.

[0037] The technical solution provided by the present application comprises the following steps: constructing a double-branch attention convolution module fusing convolution and attention mechanism, and enhancing boundary perception ability through global-local feature collaborative enhancement mechanism; designing a multi-path feature aggregation mechanism, dynamically reconstructing multi-scale feature representation through bidirectional path interaction and adaptive weighted fusion; constructing a global context fusion module to balance global information and local information; refining the target query through a prediction head, and generating a category label and a bounding box. The method enhances the overall boundary perception ability, ensures the sufficiency and effectiveness of feature aggregation, improves the positioning accuracy of the target and the accuracy of the border area identification. BRIEF DESCRIPTION OF DRAWINGS

[0038] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed in the embodiments will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0039] Figure 1 The flowchart of the method for boundary enhancement and small target perception detection of a problem map provided by the embodiment of the present application;

[0040] Figure 2 The architecture diagram of the method for boundary enhancement and small target perception detection of a problem map provided by the embodiment of the present application;

[0041] Figure 3 The architecture diagram of the double-branch attention convolution module provided by the embodiment of the present application;

[0042] Figure 4 The architecture diagram of the global context fusion module provided by the embodiment of the present application;

[0043] Figure 5 A schematic diagram of an electronic device is provided for embodiments of the present application. DETAILED DESCRIPTION

[0044] In order to make the objects, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are some but not all of the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0045] It should be clear that the described embodiments are only some but not all of the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0046] The terms used in the embodiments of the present application are only for the purpose of describing the specific embodiments and are not intended to limit the present application. The singular forms “a”, “an” and “the” used in the embodiments of the present application are also intended to include the plural forms, unless the context clearly indicates otherwise.

[0047] It should be understood that the term “and / or” used herein only describes an association relationship of associated objects, which means that there can be three relationships, for example, A and / or B, which can represent the three cases of A alone, A and B together, and B alone. In addition, the character “ / ” herein generally represents an “or” relationship between the front and rear associated objects.

[0048] Depending on the context, the word “if” as used herein can be interpreted as “when” or “upon” or “in response to determining” or “in response to detecting”. Similarly, depending on the context, the phrase “if it is determined” or “if (a stated condition or event) is detected” can be interpreted as “when it is determined” or “in response to determining” or “when (a stated condition or event) is detected” or “in response to detecting (a stated condition or event)”.

[0049] The present application provides a detection method for boundary enhancement and small target perception of a problem map, as shown in Figure 1 and Figure 2 The method comprises the following steps.

[0050] The method of the present application first extracts features through a three-stage backbone network, obtains multiple different scale features, and transmits these different scale features to the neck for further processing. In the backbone network, the input picture is processed by a convolutional block with batch normalization and Swish (CBS) in turn, a plurality of cross-stage partial block (CSP) stacking structures and a spatial pyramid pooling-fast (SPPF) module, and a plurality of feature maps with different resolutions are gradually generated. These modules work together to not only realize multi-level feature extraction of the image, but also retain rich spatial information and semantic information. In the neck, the dual-branch attention convolution module (DACM) enhances the boundary perception ability by modeling global and local features. Subsequently, the multi-path feature aggregation (MPFA) mechanism adopts a weighted strategy to perform bidirectional cross-scale fusion to maximize the retention of key information. The fused features are processed by the global context fusion module (GCFM) to improve the small target detection capability in complex backgrounds. Finally, the prediction head refines the target query and generates accurate class labels and bounding boxes.

[0051] Step 1, construct a dual-branch attention convolution module that integrates convolution and attention mechanisms, and through a global-local feature collaborative enhancement mechanism to enhance the boundary perception ability.

[0052] Due to the discontinuity and complexity of edge features in map data, it has become a core challenge in target detection to accurately extract boundary features while ensuring the integrity of the boundary. To this end, the present application proposes a feature enhancement module that integrates convolution and attention mechanisms, namely the dual-branch attention convolution module, as shown in Figure 3 By constructing a local-global feature collaborative enhancement mechanism, this module aims to model the fine expression of complex map edge patterns, thereby improving the integrity and robustness of boundary features. Specifically, after channel compression, the input features will be sent to independent branches for modeling and fusion.

[0053] In the embodiment of the present application, step 1 includes:

[0054] The global correlation modeling path: firstly, 1x1 and 3x3 convolution operations are adopted on the input feature to generate query Q, key K and value V tensors; subsequently, the attention map is calculated through the weighted filtering strategy to realize the interaction between local features, thereby enhancing the selectivity and position sensitivity of the features; then, the weighted product operation is performed on the attention map and the value V to obtain the final output feature map , combined with dynamic convolution, to optimize the modeling effect of the edge; this process not only strengthens the edge features of the target, but also establishes the potential connection between the broken edges, thereby improving the continuity of the boundary. Among them, The expression is:

[0055] (1);

[0056] Wherein, X∈ represents the input feature map, ∈ are 3x3 depth separable convolution kernels of query Q, key K and value V respectively; is a scaling factor, and is taken as =C; wherein, H represents the height of the feature map, W represents the width of the feature map, and C represents the number of channels of the feature map; the final output feature map contains local texture information and cross-region feature correlation information.

[0057] The local channel interaction path: firstly, the input feature is compressed and reconstructed through 1x1 convolution operation to enhance the compactness of the feature representation; subsequently, the input feature is divided into N subgroups along the channel dimension, and each subgroup extracts fine-grained local information through depth separable convolution, thereby strengthening the modeling ability of local feature patterns; then, in order to enhance the information interaction between channels, channel rearrangement is introduced to rearrange the channel index, and dynamic convolution and random dropout are combined;

[0058] Finally, the results of the global correlation modeling path and the local channel interaction path are fused after element-wise addition and summation to obtain the output feature.

[0059] The attention mechanism plays a key role in enhancing edge representation, but traditional methods are difficult to adapt to complex boundary shapes and suppress background noise, often leading to incomplete or unstable edge extraction. The invention constructs a dual-branch parallel module that integrates convolution and attention mechanism, also known as a dual-branch attention convolution module. The module can enhance boundary perception ability through a global-local feature collaborative enhancement mechanism, where the global branch is used to model long-range dependencies to enhance overall boundary perception ability; the local branch optimizes local features through channel rearrangement, dynamic convolution and random dropout operations to improve detail representation ability and model robustness. The synergistic effect of global-local dual branches enables the model to stably extract complete edge features in complex boundary and high noise scenes, ensuring accurate identification of the boundary region.

[0060] Step 2: According to step 1, a multi-path feature aggregation mechanism is designed to dynamically reconstruct multi-scale feature representations through bidirectional path interaction and adaptive weighted fusion.

[0061] Map data usually has multiple resolutions, resulting in complex and diverse feature distributions at different scales. Existing feature fusion methods have certain limitations in dealing with complex backgrounds and scale changes in maps, which in turn affects the positioning accuracy of the model. To this end, the invention proposes a multi-path feature aggregation mechanism, as shown in Figure 2 The module dynamically reconstructs multi-scale feature representations through bidirectional path interaction and adaptive weighted fusion (AWF), achieving more in-depth and accurate cross-scale information integration, thereby effectively improving the model's ability to express features at different scales and preserving detailed information.

[0062] In the embodiment of the invention, step 2 includes:

[0063] The multi-path feature aggregation mechanism passes deep semantic information to shallow features through expression (2) to enhance semantic perception ability, and the form of expression (2) is:

[0064] (2);

[0065] wherein, represents the output feature map after fusion of the i-th layer; represents the feature of the i-th layer; and represents the weight for adjusting the contribution proportion of the up-sampled feature and the current layer feature, respectively; the function represents up-sampling the feature of the i+1-th layer to match the resolution of the i-th layer feature;

[0066] Meanwhile, the fine-grained information in the shallow features is fed back to the deep features through the bottom-up path to enhance the expression ability of the deep features to the detailed information, and the expression of the process is:

[0067] (3);

[0068] wherein the function represents down-sampling the features of the i-1th layer to match the resolution of the features of the ith layer;

[0069] In the part of the bidirectional path interaction, the multi-path feature aggregation mechanism introduces adaptive weighted fusion; the weights from the two different scale features are dynamically adjusted through expression (4) to realize feature fusion, and finally obtain the fusion output O, and the form of expression (4) is:

[0070] (4);

[0071] wherein, represents the feature map of the first scale path; represents the feature map of the second scale path; represents a constant, and the value is 1x10 -4 ;

[0072] In the embodiment of the application, the features are extracted through the three-stage backbone network to obtain a plurality of different scale features, and the weight of two different scale features from the plurality of different scale features is dynamically adjusted each time through expression (4).

[0073] In the embodiment of the application, to prevent information loss, the multi-path feature aggregation mechanism adopts a cross-scale feature connection strategy, so that the features of the same level can be directly transmitted to the final output layer. This design not only ensures the semantic integrity in the multi-scale feature fusion process, but also retains the fine local details, especially when dealing with complex boundaries (such as broken national boundary lines), which shows a significant advantage.

[0074] The multi-path feature aggregation mechanism proposed in the application not only breaks through the limitations of traditional methods in multi-scale target detection, but also effectively suppresses the information loss and redundancy problems in complex scenes, and significantly enhances the understanding ability and positioning accuracy of the model to different scale targets.

[0075] Single-scale feature extraction is difficult to handle targets of different sizes and shapes at the same time, which limits the performance of the model in target recognition tasks. In addition, single-scale methods also have shortcomings in capturing the overall structure of large targets, leading to a decline in positioning accuracy. The present application designs a multi-path feature aggregation mechanism to enhance the recursive connection of cross-scale features through bidirectional path interaction and adaptive weighted fusion. This module uses bidirectional path interaction to achieve more efficient feature fusion and introduces a cascade interaction design combining adaptive weighted fusion and Cross Stage Partial Block (CSP) to ensure the sufficiency and effectiveness of feature aggregation. With this design, the multi-path feature aggregation mechanism not only effectively improves the model's perception of targets of different scales, but also significantly improves the positioning accuracy of targets.

[0076] Step 3, according to step 2, a global context fusion module is constructed to balance global information and local information.

[0077] Existing models still have shortcomings in small target feature extraction, making it difficult to effectively distinguish islands from background textures in low-contrast sea scenes. At the same time, the model lacks the ability to model the overall outline of the island, resulting in incomplete representation of its shape and boundary. To solve the above problems, the present application proposes a global context fusion module, as shown in Figure 4 which models features of different levels through a multi-branch structure to achieve comprehensive extraction and fusion of global information and local details.

[0078] In the embodiment of the present application, step 3 comprises:

[0079] The global context fusion module models features of different levels through a multi-branch structure. First, the global context fusion module combines global average pooling GAP and global maximum pooling GMP to extract global information of input features through expression (5), which has the form of expression (5):

[0080] (5);

[0081] wherein, , ; H represents the height of the feature map, W represents the width of the feature map, represents the feature value at spatial position (i, j); and represent the weights of linear transformation of the features extracted by global average pooling and global maximum pooling, respectively; represents the global distribution feature extracted by global average pooling, represents the salient region feature extracted by global maximum pooling; after weighted summation by the attention weights generated by the activation function Softmax, the global feature output is obtained ;

[0082] Secondly, the global context fusion module introduces a self-attention mechanism, generates query Q and key K through 1x1 convolution, and performs nonlinear mapping through an activation function to improve the feature expression ability, and at the same time, the value V is directly obtained from the input feature; then, the query Q, the key K and the value V are multiplied element by element to generate enhanced global feature representation, and the expression of the process is:

[0083] (6);

[0084] Among them, represents the global enhanced feature output generated by the self-attention mechanism; , are the 1x1 convolution kernels of the query Q and the key K, respectively;

[0085] Finally, the 1x1 convolution processed is fused with for prediction.

[0086] Through feature complementation and dynamic focusing of the self-attention mechanism, this module effectively enhances the recognition and representation ability of small targets, thereby improving the detection accuracy of the model in complex scenes.

[0087] The existing target detection model usually uses multiple downsampling operations in the backbone network to expand the receptive field. However, this method may compress or lose the position and detail information of small targets in the low-level feature map, especially when the small target only occupies a few pixels, which is prone to miss detection. To solve this problem, the present application constructs a global context fusion module to enhance the feature representation of small targets and reduce the miss detection rate. The global context fusion module uses a multi-branch attention mechanism, combines global average pooling and global maximum pooling, and extracts global and local statistical information, thereby improving the integrity of the features. In addition, the global context fusion module introduces a self-attention mechanism to optimize the spatial distribution of the features, guide the model to focus on the key areas around the small target, improve the distinguishability of the target, and effectively suppress the background interference. By fusing the outputs of different branches, the global context fusion module realizes the balance of global and local information, preserves the details and textures of small targets in complex backgrounds, thereby improving the recognition accuracy and reducing the detection error.

[0088] Step 4, according to step 3, the target query is refined through the prediction head, and a class label and a bounding box are generated.

[0089] The technical scheme provided by the application comprises the following steps: constructing a double-branch attention convolution module fusing convolution and attention mechanism, and enhancing the boundary perception ability through a global-local feature collaborative enhancement mechanism; designing a multi-path feature aggregation mechanism, dynamically reconstructing multi-scale feature representation through bidirectional path interaction and adaptive weighted fusion; constructing a global context fusion module to balance global information and local information; and refining the target query through a prediction head and generating a category label and a bounding box, so that the overall boundary perception ability is enhanced, the sufficiency and effectiveness of feature aggregation are ensured, and the positioning accuracy of the target and the accuracy of the boundary region identification are improved.

[0090] The various steps of the embodiment of the application can be executed by an electronic device.

[0091] The embodiment of the application provides a computer readable storage medium, which comprises a stored program, wherein when the program is executed, the electronic device where the computer readable storage medium is located executes the embodiment of the detection method for boundary enhancement and small target perception of a problem map.

[0092] Figure 5 A schematic diagram of an electronic device provided by the embodiment of the application is shown in FIG. 1. Figure 5 As shown in FIG. 1, the electronic device 21 comprises a processor 211, a memory 212, and a computer program 213 stored in the memory 212 and executable on the processor 211, wherein the computer program 213 is executed by the processor 211 to implement the detection method for boundary enhancement and small target perception of a problem map in the embodiment, and details are not described herein to avoid repetition.

[0093] The electronic device 21 comprises, but is not limited to, the processor 211 and the memory 212. Figure 5 The electronic device 21 is only an example and does not constitute a limitation on the electronic device 21, and can comprise more or fewer components than those shown, or combine certain components, or different components, for example, the electronic device can also comprise an input / output device, a network access device, a bus, etc.

[0094] The processor 211 can be a central processing unit (CPU), and can also be other general-purpose processors, a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic device, discrete gate or transistor logic, discrete hardware components, or the like. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor.

[0095] The memory 212 can be an internal storage unit of the electronic device 21, such as a hard disk or a memory of the electronic device 21. The memory 212 can also be an external storage device of the electronic device 21, such as a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, or the like. Further, the memory 212 can include both the internal storage unit and the external storage device of the electronic device 21. The memory 212 is used to store computer programs and other programs and data required by the network device. The memory 212 can also be used to temporarily store data that has been output or will be output.

[0096] Those skilled in the art can clearly understand that, for the convenience and brevity of the description, the specific working processes of the above-described system, device and unit can refer to the corresponding processes in the foregoing method embodiments, which will not be described herein.

[0097] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A detection method for boundary enhancement and small target perception of problem map, characterized in that, The method comprises: Step 1, constructing a double-branch attention convolution module fusing convolution and attention mechanism, and enhancing the boundary perception ability through global-local feature collaborative enhancement mechanism; Step 2, according to step 1, designing a multi-path feature aggregation mechanism, dynamically reconstructing multi-scale feature representation through bidirectional path interaction and adaptive weighted fusion; Step 3, according to step 2, constructing a global context fusion module to balance global information and local information; Step 4, according to step 3, refining the target query through the prediction head, and generating class labels and bounding boxes; The step 2 comprises: The multi-path feature aggregation mechanism transmits deep semantic information to shallow features through expression (2) to enhance semantic perception ability, and the expression (2) is in the form of: （2）； wherein, represents the output feature map after fusion of the i-th layer; represents the feature of the i-th layer; and represents the weight for adjusting the contribution ratio of the up-sampling feature and the feature of the current layer, respectively; the function represents up-sampling the feature of the i+1-th layer to match the resolution of the i-th layer feature; At the same time, the feature information in the shallow feature is fed back to the deep feature through bottom-up path back propagation, and the process is expressed as: （3）； wherein the function denotes down-sampling the features of the i-1th layer to match the resolution of the features of the i-th layer; In the part of bidirectional path interaction, the multi-path feature aggregation mechanism introduces adaptive weighted fusion; through expression (4), the weights of features from two different scales are dynamically adjusted to realize feature fusion, and finally the fusion output O is obtained, and the expression (4) is in the form of: （4）； wherein, represents a feature map of the first scale path; represents a feature map of the second scale path; represents a constant, taking a value of 1 x 10 -4 .

2. The method of claim 1, wherein, The step 1 comprises: The global correlation modeling path comprises the following steps: firstly, 1*1 and 3*3 convolution operations are adopted on input features to generate query Q, key K and value V tensors; subsequently, an attention map is calculated through a weighted filtering strategy to realize interaction between local features; and then, weighted product operation is performed on the attention map and the value V to obtain a final output feature map , and dynamic convolution is combined to optimize the modeling effect of the edge; wherein, The expression is: （1）； wherein X∈ represents an input feature map, ∈ are 3x3 depthwise separable convolution kernels for query Q, key K and value V respectively; is a scaling factor, and takes =C; wherein H represents the height of the feature map, W represents the width of the feature map, and C represents the number of channels of the feature map. The local channel interaction path: first, the input feature is compressed and reconstructed through 1×1 convolution operation; then, the input feature is divided into N subgroups along the channel dimension, and each subgroup extracts fine-grained local information through depth separable convolution; then, channel rearrangement is introduced, and channel index is rearranged, and dynamic convolution and random dropout are used; Finally, the results of the global correlation modeling path and the local channel interaction path are summed element by element and then fused to obtain the output feature.

3. The method of claim 1, wherein, The step 3 comprises: The global context fusion module models features of different levels through a multi-branch structure, first, the global context fusion module combines global average pooling GAP and global maximum pooling GMP, extracts global information of the input feature through expression (5), and the expression (5) is in the form of: （5）； wherein, , ; H represents the feature map height, W represents the feature map width, represents the feature value at the spatial position (i, j); and respectively represent the weights of linear transformation of the features extracted by the global average pooling and the global maximum pooling; represents the global distribution feature extracted by the global average pooling, represents the salient region feature extracted by the global maximum pooling; both are weighted and summed after being weighted by the attention weight generated by the activation function Softmax to obtain the global feature output ; Secondly, the global context fusion module introduces a self-attention mechanism, generates query Q, key K through 1×1 convolution, and performs nonlinear mapping through an activation function to improve feature expression ability, and at the same time, the value V is directly obtained from the input feature; then, the query Q, the key K and the value V are multiplied element by element to generate enhanced global feature representation, and the process is expressed as: （6）； wherein, represents a global enhanced feature output generated by a self-attention mechanism; , are 1x1 convolution kernels for the query Q and the key K, respectively. Finally, the 1x1 convolution processed with fusion for prediction.

4. A computer-readable storage medium, characterized in that, The computer readable storage medium comprises a stored program, wherein when the program runs, the computer readable storage medium controls the device where the computer readable storage medium is located to execute the detection method for boundary enhancement and small target perception of problem map in any one of claims 1-3.

5. An electronic device, comprising: Comprise: One or more processors; Memory; And one or more computer programs, wherein the one or more computer programs are stored in the memory, and the one or more computer programs comprise instructions, which when executed by the device, cause the device to execute the detection method for boundary enhancement and small target perception of problem map in any one of claims 1-3.

Images