YOLO Object Detection System Based on Fuzzy Strategy

Abstract

This invention proposes a YOLO target detection system based on a fuzzy strategy, comprising a data preprocessing unit and a detection network designed based on the YOLO model. The detection network includes: a YOLO backbone network, serving as the system's feature extraction unit, extracting multi-scale feature information from the image; a global pooling module, extracting high-level semantic features from the image; a YOLO Neck part, which performs simple adjacent-scale feature fusion by combining the shallow features of the first few layers of the backbone network with the high-level semantic features extracted by the global pooling module; a fuzzy fusion module, which uses the output features of the YOLO Neck as input to map feature information to a fuzzy feature space for selective fusion; and a fuzzy detection head, which alleviates feature uncertainty through fuzzy attention, thereby enhancing features and achieving target localization and classification. This invention significantly improves the model's detection capability and robustness in complex scenes without increasing model complexity.

Description

Technical Field

[0001] This invention relates to the fields of machine vision and artificial intelligence, specifically to a YOLO target detection system based on a fuzzy strategy. Background Technology

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

[0003] Because target detection inherently possesses the ability to automatically identify and locate targets, it has been widely applied in the field of industrial inspection.

[0004] Deep learning-based object detection models can be broadly categorized into two types: one-stage and two-stage object detection models. Two-stage models, such as the R-CNN series, divide the entire object detection process into two stages (or two sub-tasks: localization and classification). The primary task of the first stage is to determine the approximate location of the target, while the main task of the second stage is to classify and precisely locate the target. One-stage models combine the two stages of the two-stage model into one, integrating the localization and classification sub-tasks into a single stage, such as the typical YOLO series algorithms and the SSD model. Generally, two-stage models offer higher detection accuracy, but their complex network structure leads to lower detection efficiency. Conversely, one-stage models have simpler network structures and better real-time performance, but their detection results are generally inferior to those of two-stage models. For industrial inspection, the inherent complexity of two-stage models presents a significant challenge, making one-stage models clearly more suitable.

[0005] Currently, improvements in detection model performance have reached a bottleneck. Adding too many or overly complex modules to the model leads to uncontrollable complexity of the entire network. A major reason for this is the poor ability of convolutional neural networks to handle complex scenarios with high uncertainty.

[0006] In the field of machine learning, fuzzy theory is an effective tool for dealing with uncertainty. It introduces concepts such as fuzzy sets, fuzzy logic, and fuzzy reasoning to handle 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 capability of the model in complex scenarios without increasing the model complexity, this invention provides a YOLO target detection system based on a fuzzy strategy.

[0008] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0009] The YOLO object detection system based on fuzzy strategy includes a data preprocessing unit that preprocesses the input image and normalizes the image size to ensure image consistency.

[0010] The system is characterized by further including a detection network designed based on the YOLO model.

[0011] The detection network:

[0012] The YOLO backbone network serves as the system's feature extraction unit, enabling feature extraction: the image, after being preprocessed by the data preprocessing unit, is input into the YOLO backbone network, which leverages the backbone network's powerful multi-scale representation capabilities to extract multi-scale feature information from the image, including shallow features with smaller scales in the first few layers and deep features with the largest scales in the last layer.

[0013] It includes the Global Pooling Block (GPB): The input is the largest deep feature in the last layer of the YOLO backbone network. Through multi-layer cascaded pooling operations, it extracts high-level semantic features of the image, thereby obtaining 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 with small scale and the high-level semantic features output by the global pooling module (GPB). The above inputs are fused by simple adjacent scale features.

[0015] It includes a fuzzy fusion module (FFM): taking the output features of YOLO Neck as input, mapping the feature information to a fuzzy feature space for selective fusion, and obtaining more stable and more representative fused features. This module realizes the mutual embedding of information at different scales and improves the representation of targets at different scales.

[0016] It includes a fuzzy head connected to the output of the fuzzy fusion module (FFM). This fuzzy head is based on the YOLO Head design and uses Channel Fuzzy Attention (CFA) and Spatial Fuzzy Attention (SFA) to construct a fuzzy attention (FA) mechanism. Fuzzy attention (FA) alleviates feature uncertainty, thereby enhancing features and achieving target localization and classification to output detection results.

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

[0018] 1. To address the issue of insufficient global information extraction capability of the YOLO model, a global pooling module is used in the last layer of the backbone network, thereby enabling the entire model to obtain more reliable global semantic information of the image than the YOLO model.

[0019] 2. To address the uncertainty issue in YOLO's multi-scale feature fusion, a fuzzy fusion module (FFM) designed in this invention is set at the output position of the Neck. This module uses fuzzy logic to select features with high membership for fusion, thereby improving the robustness of the model in handling information fusion in complex scenarios.

[0020] 3. To address the issue of weak detection and discrimination due to feature uncertainty in the YOLO model's detection head, this invention utilizes fuzzy attention (FA, including fuzzy channel attention and fuzzy spatial attention) to replace traditional channel attention and spatial attention, thereby enhancing the features of the detection head.

[0021] 4. Compared with other YOLO-based detection methods, this application greatly improves the detection capability and robustness of the model in complex scenarios without increasing model complexity. Attached Figure Description

[0022] Figure 1(a) is a schematic diagram of the detection network structure of the system of the present invention after being designed based on the YOLO model.

[0023] Figure 1(b) is a schematic diagram of the working process of the fuzzy detection head of the system of the present invention.

[0024] Figure 2 This is a schematic diagram illustrating the workflow of the YOLO target detection system based on a fuzzy strategy according to the present invention.

[0025] Figure 3 This is a schematic diagram of the global pooling module operation process.

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

[0027] Figure 5 This diagram illustrates the use of fuzzy channel attention to enhance the features of a fuzzy detection head.

[0028] Figure 6This is a schematic diagram illustrating the use of fuzzy spatial attention to enhance the features of a fuzzy detection head. Detailed Implementation

[0029] The technical solutions provided in this application will be further described below with reference to specific embodiments and accompanying drawings. The advantages and features of this application will become clearer from the following description.

[0030] As shown in Figure 1, the YOLO target detection system based on the fuzzy strategy includes a data preprocessing unit, which preprocesses the input image and normalizes the image size to ensure image consistency.

[0031] The system is characterized by further including a detection network designed based on the YOLO model.

[0032] The detection network (example, YOLOv6s):

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

[0034] It includes the Global Pooling Block (GPB): The input is the largest deep feature in the last layer of the YOLOv6s backbone network. Through multi-layer cascaded pooling operations, it extracts high-level semantic features of the image, thereby obtaining 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 small scale in the first few layers (the third and fourth layers in this example) of the YOLOv6s backbone network and the high-level semantic features output by the global pooling module (GPB), and the above input is fused with features of adjacent scales.

[0036] The module includes a fuzzy fusion module (FFM): it takes the output features of the YOLOv6s Neck as input, maps the feature information to a fuzzy feature space for selective fusion, and obtains more stable and more representative fused features. This module realizes the mutual embedding of information at different scales, improving the representation of targets at different scales. Since fuzzy inference can effectively reduce the uncertainty of features, fuzzy fusion can obtain more stable and more representative fused features.

[0037] It includes a fuzzy detection head: connected to the output of the fuzzy fusion module (FFM). This 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) alleviates feature uncertainty, thereby enhancing features and realizing the localization and classification of targets to output detection results.

[0038] like Figure 2 As shown, after the input image is preprocessed by the data preprocessing unit, it undergoes feature extraction by the feature extraction unit, followed by global context feature extraction 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) in the Neck section to obtain fused features. Then, the fuzzy detection head performs target detection based on the fused features, and finally outputs the detection results.

[0039] The technology of this invention will be further described in detail below.

[0040] The aforementioned data preprocessing unit performs standardization processing on the input natural image, including scaling, cropping, and rotation, to adapt it to subsequent processing flows.

[0041] The feature extraction unit described above uses a YOLOv6s backbone network to extract features. The features obtained from this part are represented 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 aforementioned Global Pooling Block (GPB) includes multiple Ghost Block (GB) modules, a PE module, skip connections, and concat operations to extract the target's global context information. For example... Figure 3 As shown, the implementation includes the following steps:

[0043] S21: Extract the features I∈R from the feature extraction unit. C×H×W It is input into GPB as its input feature.

[0044] S22: First, in the GB (GhostBlock) module, a larger receptive field (i.e., a larger contextual information window) is obtained through convolution to extract global information. Specifically, this includes:

[0045] S221: Use Conv1×1 to compress the number of channels of input feature I from C to c in order 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 It is Conv1×1.

[0048] S222: Use N DWConv3×3 (depthwise separable convolutions, kernel size 3×3, dilation rate d) to extract F respectively. o The features of each channel are used to obtain N features, which are represented as follows: in, This represents the features extracted by the nth depthwise separable convolution, where 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 operations of the GB module in S22 are well-known operations within the domain, therefore their specific implementation is not discussed here. Figure 3 It is displayed in the middle.

[0052] S23: The features obtained after GB processing. The process is further implemented in two paths. One path is input into the PE module to further expand the receptive field and capture more contextual information, proceeding to step S24; the other path is passed to the final Concat operation through a jump-layer connection, proceeding to step S27.

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

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

[0055] 2) Another approach is used to capture the relative global features at this input scale.

[0056] The operation of the PE module is as follows:

[0057]

[0058] GAP stands for Global Average Pooling.

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

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

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

[0062]

[0063] The aforementioned Fuzzy Fusion Module (FFM) first uses fuzzy logic to determine the uncertainty of multi-scale features, and then determines the importance of features at different scales during feature fusion based on the membership degree of the features, thereby achieving efficient fusion of multi-scale information.

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

[0065] S31: The input to the fuzzy fusion module comes from the output of YOLOv6s Neck, and is still a multi-scale feature, namely... and Where 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 The sizes of the features are adjusted to be consistent. The adjusted features are represented as follows: and

[0067] S33: For For each feature point in (i = 1, 2, 3) as an example, 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. This represents the uncertainty score assigned to the feature point (x, y, c) by the m-th membership function. m,c and σm,c Let represent the mean and variance of the m-th membership function, respectively.

[0070] Preferably, M is 3.

[0071] S34: Establish Features The process of establishing a one-to-one correspondence between feature points in fuzzy features includes:

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

[0073]

[0074] In the above formula This indicates that the m-th membership function acts on the feature. The channel-independent blur features obtained from the c-th channel are assumed to be... The x and y axes are aligned with the fuzzy features, thus achieving a mapping from the feature space to the fuzzy semantic space.

[0075] After that, Channel stitching is performed, and the final blurred features are obtained using simple ordinary convolution operations and normalization operations. Will The number of channels is compressed to the same level as The same. This can be expressed by the following formula:

[0076]

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

[0078] S35: From fuzzy features From the set, C features with channels are selected to form a new attention feature, and finally, these features are applied to... and Perform element-wise multiplication. The formal description is as follows:

[0079]

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

[0081]

[0082] These are the features after fuzzy fusion.

[0083] The aforementioned fuzzy detection head employs fuzzy channel attention (CFA) and fuzzy spatial attention (SFA) sequentially to enhance its ability to represent targets.

[0084] Its operation is as follows:

[0085] The output of the fuzzy fusion module is fed into the fuzzy detection head, where it undergoes 1×1 and 3×3 convolutions (see Figure 1) to obtain feature I. h Then, fuzzy attention is applied to perform the operation.

[0086] S41 fuzzy channel attention operation as follows Figure 5 As shown, it includes:

[0087] S411: I h This indicates that the input for this part comes from the output of the second convolution of the blur detection head.

[0088] S412: For I h First, N Gaussian blur membership functions are used to blur the feature, generating N blurry features (each element in the blurry feature represents the degree to which a feature point in the channel belongs under that membership function). Preferably, N is 2, 3, or 4. The formal description is as follows:

[0089]

[0090] In the above formula This 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. F (x,y,c,n) This represents the uncertainty score assigned to the feature point (x, y, c) by the nth membership function. n,c and σ n,c Let represent the mean and variance of the nth membership function, respectively.

[0092] S413: Next, global average pooling is performed. This is used to obtain the membership degree assigned to each channel by each membership function. The process is formally described as follows:

[0093]

[0094] In the above formula It is a (1×1×C) fuzzy vector, used as attention weights.

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

[0096] S415: Finally, the importance of the channel is assigned using the following formula.

[0097]

[0098] S42 fuzzy spatial attention operation as follows Figure 6 As shown, it includes:

[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, and name the groups I. si (i = 1, 2, ..., H × W). Each group contains C feature points along the channel direction, corresponding to I. s The coordinates in the equation are {(x,y,1),…,(x,y,C)}.

[0101] S423: For elements in a group Spatial fuzzification is performed using P Gaussian membership functions. Preferably, P is 2, 3, or 4. This process is formalized as follows:

[0102]

[0103] In the above formula Let p be the 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. F (x,y,c,p) This represents the uncertainty score assigned to the feature point (x, y, c) by the p-th membership function. p,c and σ p,c Let represent the mean and variance of the p-th membership function, respectively.

[0105] S424: After performing fuzzy transformations on all spatial locations, concatenate all the fuzzified features according to their two-dimensional coordinate positions to obtain a spatial fuzzy feature F with dimension (C×P,h,w). S The formula is as follows:

[0106]

[0107] S425: Two Conv1×1 operations act on F SCompress it to (1, h, w). The formula is as follows:

[0108]

[0109] S426: Embedding fuzzy spatial attention into input feature I through element-wise multiplication. s The formula is as follows:

[0110]

[0111] The above description is merely a description of preferred embodiments of this application and is not intended to limit the scope of this application in any way. Any changes or modifications made by those skilled in the art based on the above-disclosed technical content should be considered as equivalent and valid embodiments and fall within the scope of protection of the technical solution of this application.

Claims

1. A YOLO object detection system based on a fuzzy strategy, including a data preprocessing unit, which preprocesses the input image and normalizes the image size to ensure image consistency; Its features are, The system also includes a detection network designed based on the YOLO model; The detection network: The YOLO backbone network serves as the system's feature extraction unit, enabling feature extraction: the image, after being preprocessed by the data preprocessing unit, is input into the YOLO backbone network, which leverages the backbone network's powerful multi-scale representation capabilities to extract multi-scale feature information from the image, including shallow features with smaller scales in the first few layers and deep features with the largest scales in the last layer. It includes the Global Pooling Block (GPB): The input is the largest deep feature in the last layer of the YOLO backbone network. Through multi-layer cascaded pooling operations, it extracts high-level semantic features of the image, thereby obtaining the global semantic information of the target. The YOLO Neck component includes: inputs to the shallow, small-scale features of the first few layers of the YOLO backbone and high-level semantic features output from the Global Pooling Module (GPB), which are then fused using features of adjacent scales. It includes a fuzzy fusion module (FFM): taking the output features of YOLO Neck as input, mapping the feature information to a fuzzy feature space for selective fusion, and obtaining more stable and more representative fused features. This module realizes the mutual embedding of information at different scales and improves the representation of targets at different scales. It includes a fuzzy head connected to the output of the fuzzy fusion module (FFM). This fuzzy head is based on the YOLO Head design and uses Channel Fuzzy Attention (CFA) and Spatial Fuzzy Attention (SFA) to construct a fuzzy attention (FA) mechanism. Fuzzy attention (FA) alleviates feature uncertainty, thereby enhancing features and achieving target localization and classification to output detection results.

2. The YOLO target detection system based on a fuzzy strategy as described in claim 1, characterized in that, The feature extraction unit uses the YOLO backbone network as the backbone network to extract features, and the resulting feature representation is as follows: 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 a fuzzy strategy as described in claim 1, characterized in that, The global pooling module includes multiple GB (GhostBlock) modules, PE modules, skip connections, and concat operations to extract the target's global context information, implementing the following steps: S21: The features obtained from the feature extraction unit... The input is fed into the GPB as its input features; S22: First, in the GB (GhostBlock) module, a larger receptive field, i.e., a larger contextual information window, is obtained through convolution to extract global information, specifically including: S221: Use Conv1×1 to process the input features The number of channels from C Compress to c To remove redundant features, this operation is expressed by the following formula: In the above formula, For Conv1×1; S222: Extract using N depthwise separable convolutions DWConv3×3 respectively. The features of each channel are used to obtain N features, which are represented as follows: ,in, The nth depthwise separable convolutional feature represents the feature extracted, where n = 1, 2, ..., N; S223: The features extracted in steps S221 and S222 are concatenated to increase the number of channels from... c Expand to (N+1)× c One, as shown in the following formula: S23: The features obtained after GB processing. The process is further implemented in two paths. One path is input into the PE module to further expand the receptive field and capture more contextual information, proceeding to step S24; the other path is passed to the final Concat operation through a jump-layer connection, proceeding to step S27. S24: PE contains two parallel branches: one for preserving the spatial context information of the original scale of the input features, and the other for capturing the relative global features at that input scale; The operation of the PE module is as follows: in This is a 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 steps, multiple levels of global context information are obtained; S27: Finally, feature integration is achieved through a simple concatenation operation and Conv1×1 processing, while simultaneously restoring the number of feature channels to [previous state]. C The specific steps are as follows: 。 4. The YOLO target detection system based on a fuzzy strategy as described 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 features at different scales during feature fusion based on the membership degree of the features, thereby achieving efficient fusion of multi-scale information. Its operation is as follows: S31: The input to the fuzzy fusion module comes from the output of YOLO Neck, and is still a multi-scale feature, namely... , and , where i, i-1 and i+1 correspond to the current layer, the previous layer and the next layer; S32: Before inputting the fuzzy fusion module, in order to ensure the consistency of feature scale, the three features are first processed. , and The sizes are adjusted to be consistent, and the adjusted features are represented as follows: , and ; S33: Regarding For each feature point in the model, M Gaussian fuzzy membership functions are set to calculate its uncertainty, which is expressed by the following formula: ； C, H, and W represent the number of channels, length, and width of the feature, respectively, and (x, y, c) represent the coordinates of the feature point. This represents the uncertainty score assigned to the feature point (x, y, c) by the m-th membership function; and Let represent the mean and variance of the m-th membership function, respectively. S34: Establish Features The process of establishing a one-to-one correspondence between feature points in fuzzy features includes: First, The concatenation along the dimension m of the membership function is expressed by the following formula: In the above formula This indicates that the m-th membership function acts on the feature. The c Each channel yields independent fuzzy features, by default. The x and y axes are aligned with the fuzzy features, thus achieving a mapping from the feature space to the fuzzy semantic space; After that, Channel concatenation is performed, and the final blurred features are obtained using ordinary convolution and normalization operations. ,Will The number of channels is compressed to the same level as Same; expressed by the following formula: In the above formula This indicates a Conv 3×3 operation; S35: From fuzzy features , , From the set, select C Each channel-based feature is combined to form a new attention feature, which is then applied to... , , and Perform element-wise multiplication; formally described as follows: In the above formula This indicates the superscript of the element with the maximum value. These are the features after fuzzy fusion.

5. The YOLO target detection system based on a fuzzy strategy as described in claim 1, characterized in that... The fuzzy detection head sequentially employs fuzzy channel attention and fuzzy spatial attention to enhance its ability to represent targets. Its operation is as follows: The output of the fuzzy fusion module is fed into the fuzzy detection head, where it undergoes 1×1 and 3×3 convolutions to obtain the features. Then, fuzzy attention is applied for the operation, specifically as follows: S41 fuzzy channel attention operations include: S411: This indicates the input for this section; S412: For First, N Gaussian blur membership functions are used to blur the feature, generating N blur features. Each element in the blur feature represents the degree to which a feature point in the channel belongs under the membership function. The formal description is as follows: In the above formula This represents the nth membership function in the i-th channel. The formal definition of the membership function is: Where C, H, and W represent the number of channels, length, and width of the feature, respectively, and (x, y, c) represent the coordinates of the feature point. This represents the uncertainty score assigned to the feature point (x, y, c) by the nth membership function. and Let represent the mean and variance of the nth membership function, respectively. S413: Next, global average pooling is performed. The process of obtaining the membership degree assigned to each channel by each membership function is formally described as follows: In the above formula It is a (1×1×) C ) fuzzy vectors, used as attention weights; S414: Next, the blurred vectors of all feature channels obtained through the above operations are concatenated, and the channel dimensions are restored to (1×1×) using two Conv1×1 operations. C ); S415: Finally, the importance of allocating channels is determined by the following formula: S42 fuzzy spatial attention operations include: S421: Will As input, ; S422: Input features Divided into Groups, group name Each group includes the channel direction. C Each feature point corresponds to The coordinates in are respectively }; S423: For elements in a group We use P Gaussian membership functions to spatially fuzzify it, and the process is formalized as follows: In the above formula Let p be the 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, and (x, y, c) represent the coordinates of the feature point. Let represent the uncertainty score assigned to the feature point (x, y, c) by the p-th membership function. and Let represent the mean and variance of the p-th membership function, respectively. S424: After performing fuzzy transformations on all spatial locations, concatenate all the fuzzified features according to their two-dimensional coordinates to obtain a dimension... Spatial fuzzy features The formula is as follows: ； S425: Two Conv1×1 operations act on... Compress it to The formula is as follows: S426: Embedding fuzzy spatial attention into input features through element-wise multiplication. In Chinese, the formula is as follows: 。

Images