YOLO target detection system based on fuzzy strategy

By introducing global pooling module, fuzzy fusion module and fuzzy attention mechanism into the YOLO object detection model, the problem of uncertainty in the object detection model in complex scenarios is solved, the detection accuracy and robustness are improved, and suitable for industrial detection. field.

CN119091126AActive Publication Date: 2024-12-06TONGJI UNIV
View PDF 4 Cites 0 Cited by

Patent Information

Application Number
CN202411318299.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-09-20
Publication Date
2024-12-06
Estimated Expiration
2044-09-20

AI Technical Summary

Technical Problem

There is uncertainty in existing object detection models when dealing with complex scenarios, which makes it difficult to take into account both detection accuracy and efficiency. In particular, the complexity of the second-stage model has limited the application of the first-stage model in the field of industrial detection.

Method used

A YOLO object detection system based on fuzzy strategy was designed. The global semantic information of the image was extracted through the global pooling module. The fuzzy fusion module carried out multi-scale feature fusion, and a fuzzy attention mechanism was used in the detection head to alleviate feature uncertainty. , improve the robustness of the model in complex scenarios.

Benefits of technology

Without increasing the complexity of the model, the detection capability and robustness of the model are significantly improved, the processing capability of complex scenarios is improved, and the accuracy of target positioning and classification is enhanced.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119091126A_ABST
    Figure CN119091126A_ABST
Patent Text Reader

Abstract

The invention provides a YOLO target detection system based on a fuzzy strategy. The YOLO target detection system comprises a data preprocessing unit and a detection network designed based on a YOLO model. The detection network comprises a YOLO backbone network which is a feature extraction unit of the system and is used for extracting multi-scale feature information of an image; the global pooling module is used for extracting high-level semantic features of the images; the YOLO Neck part is used for carrying out simple adjacent scale feature fusion on shallow layer features with relatively small scales of the first few layers in the backbone network and high-level semantic features extracted by the global pooling module; the fuzzy fusion module takes the output features of the YOLO Neck as input and maps feature information into a fuzzy feature space for selective fusion; the fuzzy detection head alleviates the uncertainty of the features through fuzzy attention, thereby enhancing the features, and achieving the positioning and classification of a target. According to the method, on the premise that the complexity of the model is not increased, the detection capability of the model and the robustness in a complex scene are greatly improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the fields of machine vision and artificial intelligence, and in particular to a YOLO target detection system based on a fuzzy strategy. Background Art

[0002] Object detection aims to identify and locate target objects in visual scenes. As an important task in the field of machine vision, it can be widely used in subsequent visual tasks such as instance segmentation, image description generation, and target tracking.

[0003] Since target detection essentially has the ability to automatically identify and locate targets, it has been widely used in the field of industrial detection.

[0004] The target detection model based on deep learning can be simply divided into two categories: one-stage target detection model and two-stage target detection model. The two-stage model, such as the RCNN series model, divides the entire process of target detection into two stages (or two subtasks: positioning subtask and classification subtask). The main task of the model in the first stage is to find the approximate location of the target, and the main task in the second stage is to classify and accurately locate the target. The one-stage model combines the two stages of the above-mentioned two-stage model into one, and merges the two subtasks of positioning and classification into one stage, such as the typical YOLO series algorithm and SSD model. Generally speaking, the detection accuracy of the two-stage model algorithm is higher, but its complex network structure will lead to poor detection efficiency of the model. On the contrary, the network structure of the one-stage model is simple and the real-time performance is good, but its detection results are worse than those of the two-stage model. For the field of industrial detection, since the complexity of the two-stage model is an inherent difficulty that is difficult to overcome, it is obvious that the one-stage model is more suitable.

[0005] Currently, we have reached a bottleneck in improving the performance of detection models. Adding too many and overly complex modules to the model will cause the complexity of the entire network to become uncontrollable. One of the main reasons for the above results is that convolutional neural networks are not good at handling complex scenes with high uncertainty.

[0006] In the field of machine learning, fuzzy theory is an effective tool for dealing with uncertainty problems. It introduces concepts such as fuzzy sets, fuzzy logic, and fuzzy reasoning to deal with fuzzy, imprecise or uncertain information, enabling machine learning models to better cope with uncertainty. Summary of the invention

[0007] In order to effectively control the uncertainty of the target detection model and improve the detection ability of the model in complex scenes without increasing the complexity of the model, the present invention provides a YOLO target detection system based on a fuzzy strategy.

[0008] In order to achieve the above object, the technical solution adopted by the present invention is as follows:

[0009] The YOLO target detection system based on the fuzzy strategy includes a data preprocessing unit, which preprocesses the input image and normalizes the size of the image to ensure the consistency of the image;

[0010] Characterized in that the system also includes a detection network designed based on the YOLO model.

[0011] The detection network:

[0012] The backbone network of YOLO is used as the feature extraction unit of the system to realize feature extraction: the image preprocessed by the data preprocessing unit is input into the backbone network of YOLO, and the multi-scale feature information of the image is extracted by using the powerful multi-scale representation ability of the backbone network, including the shallow features with smaller scales in the first few layers and the deep features with the largest scale in the last layer;

[0013] Including the global pooling block (GPB): the input is the deep features with the largest scale in the last layer of the YOLO backbone network (Backbone), and through multi-layer cascade pooling operations, the high-level semantic features of the image are extracted to obtain the global semantic information of the target;

[0014] Including the YOLO Neck part: the input is the shallow features of the first few layers of the YOLO backbone network (Backbone) and the high-level semantic features output by the global pooling module (GPB), and the above input is simply fused with adjacent scale features;

[0015] Including Fuzzy Fusion Module (FFM): taking the output features of YOLO Neck as input, mapping the feature information into the fuzzy feature space for selective fusion, and obtaining more stable and more representative fusion features. This module realizes the embedding of information of different scales and improves the representation of objects of different scales.

[0016] It includes a fuzzy detection head (Fuzzy Head): connected to the output of the fuzzy fusion module (FFM), the fuzzy detection head is designed based on the YOLO Head, and uses fuzzy channel attention (CFA) and fuzzy spatial attention (SFA) to construct a fuzzy attention (FA) mechanism. Fuzzy attention (FA) is used to alleviate feature uncertainty, thereby enhancing features, achieving target positioning and classification, and outputting detection results.

[0017] Compared with the existing methods, the design principles and advantages of the YOLO target detection system technical solution based on fuzzy strategy in the present invention are as follows:

[0018] 1. To address the problem that the YOLO model is not capable of extracting global information, the invented global pooling module is used in the last layer of the backbone network, so that the entire model obtains more reliable global semantic information of the image than the YOLO model.

[0019] 2. Aiming at the uncertainty problem existing in the fusion of YOLO multi-scale features, a fuzzy fusion module (FFM) designed by the present invention is set at the output position of Neck, and the features with high membership are selected and fused by the fuzzy logic method, thereby improving the robustness of the model in processing complex scene information fusion.

[0020] 3. Aiming at the problem of weak detection discrimination caused by feature uncertainty in the detection head of the YOLO model, the fuzzy attention (FA, including fuzzy channel attention and fuzzy spatial attention) designed by the present invention is used to replace the traditional channel attention and spatial attention to enhance the detection head features.

[0021] 4. Compared with other YOLO-based detection methods, this application greatly improves the detection capability of the model and the robustness of the model in complex scenarios without increasing the complexity of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] FIG1( a ) is a schematic diagram of the structure of the detection network of the system of the present invention designed based on the YOLO model.

[0023] FIG1( b ) is a schematic diagram of the working process of the fuzzy detection head of the system of the present invention.

[0024] Figure 2 The figure is a schematic diagram of the workflow of the YOLO target detection system based on the fuzzy strategy of the present invention.

[0025] Figure 3 Schematic diagram of the operation flow of the global pooling module.

[0026] Figure 4 This is a schematic diagram of the operation flow of the fuzzy fusion module.

[0027] Figure 5 Schematic diagram of using fuzzy channel attention to enhance the detection head features.

[0028] Figure 6Schematic diagram of using fuzzy spatial attention to enhance the detection head features. DETAILED DESCRIPTION

[0029] The technical solution provided by the present application will be further described below in conjunction with specific embodiments and accompanying drawings. The advantages and features of the present application will become more apparent with the following description.

[0030] As shown in FIG1 , the YOLO target detection system based on the fuzzy strategy includes a data preprocessing unit, which preprocesses the input image and normalizes the size of the image to ensure the consistency of the image;

[0031] Characterized in that the system also includes a detection network designed based on the YOLO model.

[0032] The detection network (implementation, taking YOLOv6s as an example):

[0033] The backbone network (Backbone) of YOLOv6s is included as the feature extraction unit of the system to realize feature extraction: the image preprocessed by the data preprocessing unit is input into the backbone network of YOLOv6s, and the powerful multi-scale representation ability of the backbone network is used to extract the multi-scale feature information of the image, including the shallow features with smaller scales in the first few layers and the deep features with the largest scale in the last layer;

[0034] Including the global pooling block (GPB): the input is the deep features with the largest scale in the last layer of the backbone network of YOLOv6s. Through multi-layer cascade pooling operations, the high-level semantic features of the image are extracted to obtain the global semantic information of the target;

[0035] Including the YOLOv6s Neck part, which is an existing technology: the input is the shallow features with smaller scales of the first few layers (the third and fourth layers in the embodiment) of the backbone network (Backbone) of YOLOv6s and the high-level semantic features output by the global pooling module (GPB), and the above input is simply fused with adjacent scale features;

[0036] Including Fuzzy Fusion Module (FFM): Taking the output features of YOLOv6s Neck as input, the feature information is mapped into the fuzzy feature space for selective fusion to obtain more stable and more representative fusion features. This module realizes the mutual embedding of information of different scales and improves the representation of targets of different scales. Since the fuzzy reasoning method can effectively reduce the uncertainty of features, fuzzy fusion can obtain more stable and more representative fusion features.

[0037] It includes a fuzzy detection head (Fuzzy Head): connected to the output of the fuzzy fusion module (FFM), the fuzzy detection head is designed based on the YOLOv6s Head, and uses fuzzy channel attention (CFA) and fuzzy spatial attention (SFA) to construct a fuzzy attention (FA) mechanism. Fuzzy attention (FA) is used to alleviate feature uncertainty, thereby enhancing features, achieving target positioning and classification, and outputting detection results.

[0038] like Figure 2 As shown in the figure, after the input image is preprocessed by the data preprocessing unit, it is subjected to feature extraction by the feature extraction unit, and then the global context feature is extracted by the global pooling block (GPB); next, the feature information extracted by the feature extraction unit and the global pooling block (GPB) is fuzzily fused by the fuzzy fusion module (FFM) of the Neck part to obtain the fused feature; then, the fuzzy detection head performs target detection based on the fused feature, and finally outputs the detection result.

[0039] The technology of the present invention is further described below.

[0040] The above-mentioned data preprocessing unit performs standardization processing on the input natural image, including scaling, cropping and rotation, to adapt to the subsequent processing flow.

[0041] The above feature extraction unit: uses the backbone network of YOLOv6s as the backbone network to extract features. The features obtained in this part are expressed as: I∈R C×H×W , where C, H, and W represent the number of channels, length, and width of the feature, respectively.

[0042] The above-mentioned global pooling block (GPB) includes multiple GB (GhostBlock) modules, PE modules, skip-layer connections and cascade (Concat) operations to extract the global context information of the target. Figure 3 As shown, the implementation includes the following steps:

[0043] S21: The feature I∈R obtained by the feature extraction unit C×H×W Input into GPB as input feature of GPB.

[0044] S22: First, in the GB (GhostBlock) module, convolution is used to obtain a larger receptive field (i.e., a larger context information window) and extract global information. Specifically, it includes:

[0045] S221: Use Conv1×1 to compress the number of channels of input feature I from C to c to remove redundant features as much as possible. This operation can be expressed by the following formula:

[0046] F o =f 1×1 (I)

[0047] In the above formula, f 1×1 is Conv1×1.

[0048] S222: Use N DWConv3×3 (depth-separable convolution, convolution kernel size is 3×3, and the void rate is d) to extract F respectively o The features of each channel in , N features are obtained, which are expressed as: in, Represents the feature extracted by the nth depthwise separable convolution, n = 1, 2, ..., N. Preferably, N is 2.

[0049] S223: The features extracted in steps S221 and S222 are concatenated to expand the number of channels from c to (N+1)×c, as shown in the following formula:

[0050]

[0051] It should be noted that the operation of the GB module in S22 is a well-known operation in the field, so its specific implementation is not described in detail. Figure 3 Displayed in.

[0052] S23: Next, the features obtained after GB processing It is further implemented in two ways, one of which is input into the PE module to further expand the receptive field and capture more context information, and then enter step S24; the other is passed to the final Concat operation through the skip-layer connection and then enters step S27.

[0053] S24: PE structure is realized by two parallel branches:

[0054] 1) A piece of spatial context information used to maintain the original scale of the input features.

[0055] 2) The other one is used to capture the relative global features at the input scale.

[0056] The operations of the PE module are as follows:

[0057]

[0058] Where GAP is the global average pooling operation.

[0059] S25: The features processed by PE are passed to the next "GB (GhostBlock) + PE" unit.

[0060] S26: After multiple “GB+PE” combination processing, multiple levels of global context information are obtained.

[0061] S27: Finally, through a simple concatenation (Concat) operation and Conv1×1 processing, feature integration is achieved and the number of feature channels is restored to C. The specific operations are as follows:

[0062]

[0063] The above-mentioned fuzzy fusion module (FFM) first uses fuzzy logic methods to measure the uncertainty of multi-scale features, and then determines the importance of different scale features during feature fusion according to the feature membership, thereby achieving efficient fusion of multi-scale information.

[0064] like Figure 4 As shown, its operation is:

[0065] S31: The input of the fuzzy fusion module comes from the output of YOLOv6s Neck, which is still multi-scale features, namely and Among them, i, i-1 and i+1 correspond to the current layer, the previous layer and the next layer.

[0066] S32: Before inputting the fuzzy fusion module, in order to ensure the consistency of feature scale, the three features ( and ) are adjusted to be consistent. The adjusted features are expressed as: and

[0067] S33: For For each feature point in (as an embodiment, i=1, 2, 3), M Gaussian fuzzy membership functions are set to calculate its uncertainty, which is expressed by the following formula:

[0068]

[0069] C, H, and W represent the number of channels, length, and width of the feature, respectively, and (x, y, c) represents the coordinates of the feature point. represents the uncertainty score assigned to the feature point (x, y, c) by the mth membership function. m,c and σm,c They represent the mean and variance of the mth membership function respectively.

[0070] Preferably, the value of M is 3.

[0071] S34: Establishing Features The one-to-one correspondence between the feature points and the fuzzy features includes:

[0072] First, The splicing is performed along the dimension m of the membership function, which is expressed by the following formula:

[0073]

[0074] In the above formula Indicates that the mth membership function acts on the feature The channel-independent fuzzy features obtained by the c-th channel of The x and y axes of are aligned with the fuzzy features to achieve the mapping from feature space to fuzzy semantic space.

[0075] Afterwards, Perform channel splicing and use simple ordinary convolution operations and normalization operations to obtain the final blurred features Will The number of channels is compressed to The same. It is represented by the following formula:

[0076]

[0077] In the above formula, f 3×3 Represents a Conv 3×3 operation.

[0078] S35: From fuzzy features From the set composed of and Perform element-by-element multiplication. The formal description is as follows:

[0079]

[0080] In the above formula Indicates the superscript of the element with the maximum value.

[0081]

[0082] It is the feature after fuzzy fusion.

[0083] The above-mentioned fuzzy detection head uses fuzzy channel attention (CFA) and fuzzy spatial attention (SFA) in turn to enhance the target representation capability.

[0084] Its operation is:

[0085] The output of the fuzzy fusion module is sent to the fuzzy detection head, which first undergoes 1×1 convolution and 3×3 convolution (see Figure 1) to obtain feature I h , and then apply fuzzy attention to operate.

[0086] S41 fuzzy channel attention operation Figure 5 As shown, including:

[0087] S411:I h Represents the input of this part, which comes from the output of the second convolution of the blur detection head.

[0088] S412: For I h First, use N Gaussian fuzzy membership functions to fuzzify it and generate N fuzzy features (each element in the fuzzy feature represents the degree of the feature point in the channel under the membership function). Preferably, N is 2, 3, or 4. The formal description is as follows:

[0089]

[0090] In the above formula represents the nth membership function in the i-th channel. The formal definition of the membership function is:

[0091] Where C, H, and W represent the number of channels, length, and width of the feature, respectively. (x, y, c) represents the coordinates of the feature point. (x,y,c,n) represents the uncertainty score assigned to the feature point (x, y, c) by the nth membership function. n,c and σ n,c They represent the mean and variance of the nth membership function respectively.

[0092] S413: Next, global average pooling processing To obtain the membership assigned to the channel by each membership function. The process is formally described as:

[0093]

[0094] In the above formula is a (1×1×C) blur vector, serving as the attention weight.

[0095] S414: Next, the blurred vectors of all feature channels obtained through the above operations are concatenated, and the channel dimension is restored to (1×1×C) using two Conv1×1 operations.

[0096] S415: Finally, the importance of the channels is allocated using the following formula.

[0097]

[0098] S42 fuzzy spatial attention operation Figure 6 As shown, including:

[0099] S421: I s ∈R C×H×W As input, I s =I Ch .

[0100] S422: Input feature I s Divide into (H×W) groups, named group I si (i=1,2,…,H×W). Each group of C feature points in the channel direction corresponds to I s The coordinates in are {(x,y,1),…,(x,y,C)}.

[0101] S423: For elements in a group P Gaussian membership functions are used to spatially fuzzify it. Preferably, P is 2, 3, or 4. The process is formalized as:

[0102]

[0103] In the above formula represents the p-th spatial Gaussian fuzzy membership function. The formal definition of the membership function is:

[0104] C, H, and W represent the number of channels, length, and width of the feature, respectively. (x, y, c) represents the coordinates of the feature point. (x,y,c,p) represents the uncertainty score assigned to the feature point (x, y, c) by the pth membership function. p,c and σ p,c They represent the mean and variance of the p-th membership function respectively.

[0105] S424: After executing the fuzzy transformation of all spatial positions, all the fuzzified features are concatenated according to the two-dimensional coordinate positions to obtain a spatial fuzzy feature F with a dimension of (C×P, h, w) S The formula is as follows:

[0106]

[0107] S425: Two Conv1×1 operations are applied to F S Compress it to (1,h,w). The formula is as follows:

[0108]

[0109] S426: Embed the fuzzy spatial attention into the input feature I through element-by-element multiplication s The formula is as follows:

[0110]

[0111] The above description is only a description of the preferred embodiments of the present application, and is not intended to limit the scope of the present application. Any changes or modifications made by any person skilled in the art based on the above disclosed technical contents shall be deemed as equivalent effective embodiments and shall fall within the scope of protection of the technical solution of the present application.

Claims

1. A YOLO target detection system based on a fuzzy strategy, including a data preprocessing unit, which preprocesses the input image and normalizes the size of the image to ensure image consistency; It is characterized in that The system also includes a detection network designed based on the YOLO model; The detection network: The backbone network of YOLO is used as the feature extraction unit of the system to realize feature extraction: the image preprocessed by the data preprocessing unit is input into the backbone network of YOLO, and the multi-scale feature information of the image is extracted by using the powerful multi-scale representation ability of the backbone network, including the shallow features with smaller scales in the first few layers and the deep features with the largest scale in the last layer; Including the global pooling block (GPB): the input is the deep features with the largest scale in the last layer of the YOLO backbone network (Backbone), and through multi-layer cascade pooling operations, the high-level semantic features of the image are extracted to obtain the global semantic information of the target; Including the YOLO Neck part: the input is the shallow features of the first few layers of the YOLO backbone network (Backbone) and the high-level semantic features output by the global pooling module (GPB), and the above input is simply fused with adjacent scale features; Including Fuzzy Fusion Module (FFM): taking the output features of YOLO Neck as input, mapping the feature information into the fuzzy feature space for selective fusion, and obtaining more stable and more representative fusion features. This module realizes the embedding of information of different scales and improves the representation of objects of different scales. It includes a fuzzy detection head (Fuzzy Head): connected to the output of the fuzzy fusion module (FFM), the fuzzy detection head is designed based on the YOLO Head, and uses fuzzy channel attention (CFA) and fuzzy spatial attention (SFA) to construct a fuzzy attention (FA) mechanism. Fuzzy attention (FA) is used to alleviate feature uncertainty, thereby enhancing features, achieving target positioning and classification, and outputting detection results.

2. The YOLO target detection system based on fuzzy strategy as claimed in claim 1, characterized in that The feature extraction unit uses the backbone network of YOLO as the backbone network to extract features, and the obtained features are expressed as: I∈R C×H×W , where C, H, and W represent the number of channels, length, and width of the feature, respectively.

3. The YOLO target detection system based on fuzzy strategy as claimed in claim 1, characterized in that The global pooling module includes multiple GB (GhostBlock) modules, PE modules, skip-layer connections and cascade (Concat) operations to extract the global context information of the target, and the implementation includes the following steps: S21: The feature I∈R obtained by the feature extraction unit C×H×W Input into GPB as input features of GPB; S22: First, in the GB (GhostBlock) module, convolution is used to obtain a larger receptive field, that is, a larger context information window, and extract global information, including: S221: Use Conv1×1 to compress the number of channels of the input feature I from C to c to remove redundant features. This operation is expressed by the following formula: F o =f 1×1 (I) In the above formula, f 1×1 is Conv1×1; S222: Use N depth-separable convolutions DWConv3×3 to extract F o The features of each channel in , N features are obtained, which are expressed as: in, Represents the features extracted by the nth depth-wise separable convolution, n=1,2,…,N; S223: The features extracted in steps S221 and S222 are concatenated to expand the number of channels from c to (N+1)×c, as shown in the following formula: S23: Next, the features obtained after GB processing It is further implemented in two ways, one of which is input into the PE module to further expand the receptive field and capture more context information, and then enter step S24; the other is passed to the final Concat operation through the skip layer connection and then enters step S27; S24: PE contains two parallel branches: one is used to maintain the spatial context information of the original scale of the input feature, and the other is used to capture the relative global features under the input scale; The operations of the PE module are as follows: Among them, GAP is the global average pooling operation; S25: The features processed by PE are passed to the next "GB (GhostBlock) + PE" unit; S26: After multiple "GB+PE" combination processing, multiple levels of global context information are obtained; S27: Finally, feature integration is achieved through a simple concatenation (Concat) operation and Conv1×1 processing, and the number of feature channels is restored to C at the same time; the specific operations are as follows:

4. The YOLO target detection system based on fuzzy strategy as claimed in claim 1, characterized in that , The fuzzy fusion module first uses fuzzy logic to determine the uncertainty of multi-scale features, and then determines the importance of different scale features during feature fusion according to the feature membership, thereby achieving efficient fusion of multi-scale information; Its operation is: S31: The input of the fuzzy fusion module comes from the output of YOLO Neck, which is still multi-scale features, namely and Among them, i, i-1 and i+1 correspond to the current layer, the previous layer and the next layer; S32: Before entering the fuzzy fusion module, in order to ensure the consistency of feature scale, the three features are first and The sizes are adjusted to be consistent, and the adjusted features are expressed as follows: and S33: For For each feature point in , M Gaussian fuzzy membership functions are set to calculate its uncertainty, which can be expressed by the following formula: C, H, and W represent the number of channels, length, and width of the feature, respectively. (x, y, c) represents the coordinates of the feature point. represents the uncertainty score assigned to the feature point (x, y, c) by the mth membership function; u m,c and σ m,c Respectively represent the mean and variance of the mth membership function; S34: Establishing Features The one-to-one correspondence between the feature points and the fuzzy features includes: First, The splicing is performed along the dimension m of the membership function, which is expressed by the following formula: In the above formula Indicates that the mth membership function acts on the feature The channel-independent fuzzy features obtained by the c-th channel of The x and y axes of are aligned with the fuzzy features to achieve the mapping from feature space to fuzzy semantic space; Afterwards, Perform channel splicing and use simple ordinary convolution operations and normalization operations to obtain the final blurred features Will The number of channels is compressed to The same; expressed by the following formula: In the above formula, f 3×3 Represents Conv 3×3 operation; S35: From fuzzy features From the set composed of and Perform element-by-element multiplication; the formal description is as follows: In the above formula Indicates the superscript of the element with the maximum value; It is the feature after fuzzy fusion.

5. The YOLO target detection system based on fuzzy strategy as claimed in claim 1, characterized in that The fuzzy detection head uses fuzzy channel attention and fuzzy spatial attention in turn to enhance the target representation capability, and its operation is as follows: The output of the fuzzy fusion module is sent to the fuzzy detection head, which first undergoes 1×1 convolution and 3×3 convolution to obtain feature I h , and then apply fuzzy attention to operate, specifically: S41 fuzzy channel attention operation, including: S411:I h Indicates the input of this part; S412: For I h First, use N Gaussian fuzzy membership functions to fuzzify it and generate N fuzzy features. Each element in the fuzzy feature represents the degree of belonging of the feature point in the channel under the membership function; the formal description is as follows: In the above formula represents the nth membership function in the ith channel. The formal definition of the membership function is: Among them, C, H, W represent the number of channels, length and width of the feature respectively, (x, y, c) represents the coordinates of the feature point, and F (x,y,c,n) represents the uncertainty score assigned to the feature point (x, y, c) by the nth membership function, u n,c and σ n,c Respectively represent the mean and variance of the nth membership function; S413: Next, global average pooling processing To obtain the membership assigned to the channel by each membership function, the process is formally described as: In the above formula is a (1×1×C) blur vector, serving as the attention weight; S414: Next, the fuzzy vectors of all feature channels obtained through the above operations are concatenated, and two Conv1×1 operations are performed to restore the channel dimension to (1×1×C); S415: Finally, the importance of the channel is allocated by the following formula: S42 fuzzy spatial attention operation, including: S421: I s ∈R C×H×W As input, I s =I Ch ; S422: Input feature I s Divide into (H×W) groups, named I si (i=1,2,…,H×W); each group of C feature points in the channel direction corresponds to I s The coordinates in are {(x,y,1),…,(x,y,C)}; S423: For elements in a group P Gaussian membership functions are used to spatially fuzzify it. The process is formalized as follows: In the above formula represents the p-th spatial Gaussian fuzzy membership function. The formal definition of the membership function is: C, H, and W represent the number of channels, length, and width of the feature, respectively. (x, y, c) represents the coordinates of the feature point. F (x,y,c,p) represents the uncertainty score assigned to the feature point (x, y, c) by the pth membership function, u p,c and σ p,c Respectively represent the mean and variance of the p-th membership function; S424: After executing the fuzzy transformation of all spatial positions, all the fuzzy features are concatenated according to the two-dimensional coordinate position to obtain a spatial fuzzy feature F with a dimension of (C×O,h,w) S , the formula is as follows: S425: Two Conv1×1 operations are applied to F S Compress it to (1,h,w), the formula is as follows: S426: Embed the fuzzy spatial attention into the input feature I through element-by-element multiplication s In the formula, the following is true:

Citation Information

Patent Citations

  • Weak and small target detection method

    CN112990317A

  • Target detection algorithm for tiny object

    CN117392375A

  • Small target detection system and method based on improved YOLOv5

    CN117523267A

  • Arithmetic question marking system based on mixnet-yolov3 and convolutional recurrent neural network (CRNN)

    WO2022147965A1