Feature extraction and detection method and system for targets with missing contours in severe weather

Abstract

The present invention discloses a method and system for extracting and detecting features of targets with missing contours in bad weather. First, low-quality videos in bad weather are obtained, and high-dimensional features of video images are extracted; then robust information of the missing contour area is searched; significant information of targets with missing contours of different scales is adaptively searched, and multi-scale features are extracted; the extracted multi-scale features are processed, and cross-level information interaction is performed on features of different scales; then the correlation between the contour area, the missing contour area, and the background area is extracted; then the category and positioning information of the missing contour target are predicted; finally, redundant detection frames are eliminated to obtain the final missing contour target detection result. The present invention can quickly and accurately detect targets with missing contours in bad weather, realize intelligent monitoring in complex environments such as haze, heavy rain, and low light, or be used to enhance the robustness of unmanned driving systems.

Description

Technical Field

[0001] The present invention belongs to the field of computer vision technology, and relates to a method and system for extracting and detecting features of targets with missing contours, and specifically to a method and system for extracting and detecting robust information of targets with missing contours caused by severe weather. Background Art

[0002] Object detection technology has a wide range of applications in various fields, such as autonomous driving, security monitoring, and intelligent manufacturing. Its key lies in its ability to automatically identify and locate objects in images, greatly improving efficiency and safety. In autonomous driving, object detection can identify pedestrians, vehicles, etc., improving driving safety; in security, it can detect abnormal behavior in real time to ensure public safety.

[0003] Bad weather conditions, such as haze, heavy rain, or low light, often cause the target outline to be blurred or partially missing, which seriously affects the accuracy of target detection methods. In these environments, traditional target detection methods have difficulty in accurately identifying and locating targets because bad weather can significantly reduce the quality and clarity of the target. Therefore, the detector must have a stronger feature search capability to extract key information from the target image with blurred or missing outlines. At the same time, multi-scale expression capabilities become particularly important because the appearance and size of the target may change under bad weather conditions, making the target more diverse in appearance. This requires the detection system to effectively extract and fuse features at different scales to achieve cross-level interaction of multi-scale features to ensure detection performance and accuracy under various weather conditions. In bad weather, the outline of the target may be partially occluded or blurred, resulting in the integrity of the target being affected. Therefore, the correlation extraction between the missing outline area and the valid outline area is also crucial. Therefore, in response to the challenges of bad weather, how to perform more advanced feature extraction, cross-scale feature interaction, and correlation extraction has become a difficult problem to improve the robustness of the target detection system.

[0004] The feature extraction and detection technology of missing contour targets in bad weather aims to locate and classify key targets such as cars, pedestrians, ships, riders, etc. in bad scenes. In recent years, many advanced detectors have emerged in the field of target detection, such as the YOLO series, SSD, and DINO. However, these existing detectors still have insufficient feature extraction capabilities when facing targets with missing or blurred contours, and it is difficult to achieve cross-scale interaction. Under bad weather conditions, the contours of the target may be partially missing or blurred, which makes these detectors face challenges in extracting the correlation between contours. Although image enhancement algorithms such as Zero-DCE and RIDCP can enhance low-light images or haze images, their generalization performance is poor and it is difficult to apply them to actual scenes. Therefore, the detection problem of detectors in bad weather still exists. Summary of the invention

[0005] In order to solve the above technical problems, the present invention provides a method and system for extracting and detecting features of targets with missing contours in severe weather conditions, in combination with a deep learning-based target detection algorithm.

[0006] The technical solution adopted by the method of the present invention is: a method for extracting and detecting features of targets with missing contours in bad weather, comprising the following steps:

[0007] Step 1: Extract high-dimensional features of the target video image;

[0008] Step 2: Search for robust information of the target with missing contours and extract high-dimensional features containing robust information ;

[0009] Step 3: Target high-dimensional features that contain robust information , adaptively search for information at different scales, and obtain multi-scale contour missing target features with cross-scale information interaction ;

[0010] Step 4: Extract the correlation between contour areas, contour missing areas, and background areas to generate a unified and efficient feature representation Fed ;

[0011] Step 5: Use the classifier and bounding box regressor to predict the category and location information of the target with missing contours;

[0012] Step 6: Combine the category and positioning information obtained in step 5, eliminate redundant detection boxes, and obtain the final key target positioning information and category information.

[0013] Preferably, in step 1, a parallel atrous convolution structure with multiple receptive fields is constructed to extract high-dimensional features with strong expressive power;

[0014] The parallel dilated convolution structure is composed of dilated convolutions with dilated rates of 1, 2, and 3 respectively. The input image is defined as ,in R represents the feature domain, H and W represent the height and width of the feature respectively, which are input into the parallel dilated convolution structure to obtain high-dimensional features:

[0015] ;

[0016] In the formula, and They represent dilated convolutions with dilation rates of 1, 2, and 3 respectively, and C represents the dimension of high-dimensional features.

[0017] Preferably, in step 2, a feature depth search network is used to search for robust information of the target with missing contours; the feature depth search network is composed of a five-stage feature extraction structure, each stage of which is composed of a convolutional layer and a feature depth search module, and is used to extract high-dimensional features containing robust information. ;

[0018] Input high-dimensional features ,in R represents the feature domain, H and W represent the height and width of the feature respectively, and C represents the dimension of the high-dimensional feature; the feature depth search module first uses conventional convolution operations to predict the possible locations of valid features:

[0019] ;

[0020] In the formula, is the output of the convolution, represents a convolution operation with a kernel size of 3×3, k It represents the kernel size used for convolution within the extraction area; right Perform a split operation to divide it into bias vectors of deformable convolutions used for regional feature extraction , and the position vector used to search for the valid area and ;

[0021] Afterwards, As the bias tensor input into the deformable convolution, the local information of the current position is extracted, and the area search process is performed. The search results are:

[0022] ;

[0023] In the formula, Represents the average operation, the vector and the cut out By summing the vectors, we can get more accurate position information;

[0024] Then, the searched location information is passed to the next deformable convolution to perform feature extraction on the proposed valid area:

[0025] ;

[0026] In the formula, S It is set as a learnable parameter to control the search radius of this area search, and its initialization value represents the initial search radius;

[0027] Finally, continue searching for the next area containing valid information and determine the search direction:

[0028] ;

[0029] ;

[0030] Thus, robust information of the target with missing contour is extracted;

[0031] The feature deep search network combines the convolution and feature search modules to construct a search network of five stages from stage 0 to stage 4, and the outputs of stages 2 to 4 are As input to step 3.

[0032] Preferably, in step 3, a multi-scale feature pyramid and a cross-scale information interaction structure are constructed to adaptively search for information of different scales;

[0033] The multi-scale feature pyramid first converts the high-dimensional features containing robust information into As input, and As Layer features; then use 3×3 convolution to Processing is performed to implement downsampling operation, and we get ; Then use the nearest neighbor upsampling method to process , and Add together and get Layer Features:

[0034] ;

[0035] In the formula, represents the nearest neighbor upsampling method; then the same method is used to obtain Layer Features:

[0036] ;

[0037] but Compared with the original image, the scaled size is ;

[0038] The cross-scale information interaction structure will As multi-scale input features, firstly perform scale normalization operation on features of different scales, and unify the scales of all features through upsampling or downsampling operations:

[0039] ;

[0040] In the formula, Represents the feature scale unification operation, including upsampling and downsampling operations, represents the feature concatenation operation, represents all the concatenated features; next, the feature depth search module constructed in step 2 is used again to perform feature search on the multi-scale features;

[0041] Secondly, the multi-scale features after feature search processing and fusion are re-divided into corresponding scales:

[0042] ;

[0043] In the formula, Represents the division result of the fused features, corresponding to the original scale information, Then, the feature depth search module is used to search for features of different scales, process the divided features, and obtain features for enhancing the original scale features. ;

[0044] Finally, and Perform vector addition operations to obtain multi-scale contour missing target features with cross-scale information interaction .

[0045] Preferably, in step 4, a correlation extraction module is used to extract the correlation between the contour area, the contour missing area and the background area, so as to enhance the local detail and global semantic information perception ability of the feature;

[0046] The relevance extraction module is composed of conventional convolution, feature deep search module and Vision Mamba module. First, the contour area is processed by conventional convolution to extract key features and make up for the missing target contour. Then, the feature deep search module performs feature matching and association analysis between the contour and the contour missing area to accurately capture the target information. Finally, through the global semantic extraction function of the Vision Mamba module, the feature information of different areas is comprehensively processed to generate a unified and efficient feature representation:

[0047] ;

[0048] In the formula, represents the conventional convolution processing, represents the feature deep search module, It means Vision Mamba module;

[0049] Finally, the convolutional layer is used to build a classifier and a bounding box regressor to obtain the category and bounding box information of the target with missing contours:

[0050] ;

[0051] ;

[0052] In the formula, p Indicates the probability that the sample is predicted to be a certain type of foreground target, represents the predicted bounding box, and They represent the convolutional layers for classification and bounding box regression respectively.

[0053] Preferably, in step 5, a loss function is designed, a classifier and a bounding box regressor are trained, and the category and positioning information of the target with missing contour are predicted;

[0054] During training, the loss function of the bounding box regressor is for:

[0055] ;

[0056] In the formula, and Represent the real bounding box and the predicted bounding box respectively;

[0057] Loss function of the classifier for:

[0058] ;

[0059] In the formula, Used to adjust the importance of positive and negative samples in the sub-classifier learning process. Used to improve the sub-classifier's ability to mine difficult samples; when When , it means that the sample is a positive sample. When , it means that the sample is a negative sample. p Indicates the probability that the sample is predicted to be a certain type of foreground target;

[0060] The total loss during training is:

[0061] .

[0062] Preferably, in step 6, the frame information of the target to be detected is obtained. ,in, H and W Respectively represent the length and width of the positioning information, and their product is the number of prediction boxes;

[0063] The prediction result of each feature pixel is:

[0064] ;

[0065] in, They represent the distance from the pixel point in the feature map to the boundary of the target detection box, and finally the rectangular frame regression result is:

[0066] ;

[0067] ;

[0068] ;

[0069] ;

[0070] in, is the coordinate of the feature pixel point, is the coordinate of the upper left corner of the rectangular border, The coordinates of the lower right corner of the rectangle.

[0071] Preferably, in step 6, a non-maximum suppression method is used to eliminate redundant detection frames; the specific implementation includes the following sub-steps:

[0072] Step 6.1: Group all detection boxes by category and exclude background categories;

[0073] Step 6.2: Sort by the classification scores of the detection boxes within each category;

[0074] Step 6.3: Select the detection box with the highest classification score from the ranking as the final output;

[0075] Step 6.4: Calculate the area of ​​all detection boxes of the output category and evaluate the intersection-over-union ratio of the detection box with the highest classification score with other detection boxes;

[0076] Step 6.5: Delete all detection boxes whose intersection-over-union ratio is lower than the preset threshold;

[0077] Step 6.6: Repeat steps 6.3 to 6.5 until all detection boxes of the output category have been processed (i.e. selected or deleted);

[0078] Step 6.7: Repeat steps 6.2 to 6.6 for each category until the non-maximum suppression process is completed for all categories.

[0079] The technical solution adopted by the system of the present invention is: a system for extracting and detecting features of targets with missing contours in bad weather, comprising:

[0080] one or more processors;

[0081] A storage device is used to store one or more programs. When the one or more programs are executed by the one or more processors, the one or more processors implement the method for extracting and detecting features of targets with missing contours under severe weather conditions.

[0082] Compared with the existing detection methods, the present invention has the following advantages and positive effects:

[0083] (1) The present invention proposes a feature deep search network that can more efficiently extract robust information of objects with missing contours.

[0084] (2) The present invention proposes to construct a multi-scale feature pyramid and a cross-scale information interaction structure to adaptively search for information of different scales, thereby achieving cross-level interaction between multi-scale features and reducing redundant features.

[0085] (3) The present invention constructs a correlation extraction module that can extract the correlation between contour areas, contour-missing areas, and background areas, thereby enhancing the ability to perceive local details and global semantic information of features. BRIEF DESCRIPTION OF THE DRAWINGS

[0086] The technical solution of the present invention is further described below using embodiments and specific implementation methods. In addition, in the process of describing the technical solution, some drawings are also used. For those skilled in the art, other drawings and the intention of the present invention can also be obtained based on these drawings without paying creative work.

[0087] Figure 1 : Schematic diagram of the method of an embodiment of the present invention;

[0088] Figure 2 : Feature depth search network structure diagram of an embodiment of the present invention;

[0089] Figure 3 : A multi-scale feature pyramid and cross-scale information interaction structure diagram of an embodiment of the present invention;

[0090] Figure 4 : A structural diagram of a correlation extraction module according to an embodiment of the present invention;

[0091] Figure 5 : Schematic diagram of experimental results of an embodiment of the present invention. DETAILED DESCRIPTION

[0092] In order to facilitate ordinary technicians in this field to understand and implement the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and implementation cases. It should be understood that the implementation examples described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0093] Please see Figure 1 , this embodiment provides a method for extracting and detecting features of targets with missing contours in bad weather, comprising the following steps:

[0094] Step 1: Obtain surveillance or traffic video and extract high-dimensional features of the target video image;

[0095] In one implementation, in step 1, a parallel atrous convolution structure with multiple receptive fields is constructed to extract high-dimensional features with strong expressive power.

[0096] The parallel dilated convolution structure consists of dilated convolutions with dilated ratios of 1, 2, and 3. The input image is defined as ,in R represents the feature domain, H and W represent the height and width of the feature respectively, which are input into the parallel dilated convolution structure to obtain high-dimensional features:

[0097] ;

[0098] In the formula, and They represent dilated convolutions with dilation rates of 1, 2, and 3 respectively, and C represents the dimension of high-dimensional features.

[0099] Step 2: Search for robust information of the target with missing contours and extract high-dimensional features containing robust information ;

[0100] Please see Figure 2 In one embodiment, a feature deep search network is used to search for robust information of contour-missing targets; the feature deep search network is composed of a five-stage feature extraction structure, each stage is composed of a feature deep search module for extracting high-dimensional features containing robust information ;

[0101] The feature depth search module of this embodiment is used to progressively search for robust features of contour-missing targets. The input feature is ,in R Represents the feature domain, H and W represent the height and width of the feature respectively, and C represents the dimension of the high-dimensional feature. First, use the conventional convolution operation to predict the possible location of the effective feature:

[0102] ;

[0103] In the formula, is the output of the convolution, represents a convolution operation with a kernel size of 3×3, k It represents the kernel size used for convolution within the extraction area; right Perform a split operation to divide it into bias vectors of deformable convolutions used for regional feature extraction In addition, the position vector used to search the effective area is divided, namely and ;

[0104] Afterwards, As the bias tensor input into the deformable convolution, the local information of the current position is extracted, and the area search process is performed. The search results are:

[0105] ;

[0106] In the formula, represents the averaging operation, is the bias vector of the deformable convolution, and its average vector can provide a reference for the search direction of the effective area. and the cut out By summing the vectors, we can get more accurate position information;

[0107] Then, the searched location information is passed to the next deformable convolution to perform feature extraction on the proposed valid area:

[0108] ;

[0109] In the formula S It is set as a learnable parameter to control the search radius of this area search, and its initialization value represents the initial search radius;

[0110] Finally, continue searching for the next area containing valid information and determine the search direction:

[0111] ;

[0112] ;

[0113] Thus, robust information of the target with missing contour is extracted;

[0114] Step 3: Target high-dimensional features that contain robust information , adaptively search for information at different scales, and obtain multi-scale contour missing target features with cross-scale information interaction ;

[0115] Please see Figure 3 ,In one implementation method, a multi-scale feature pyramid and a cross-scale information interaction structure are constructed to adaptively search information at different scales;

[0116] The multi-scale feature pyramid of this embodiment first converts the high-dimensional features containing robust information into As input, and As Layer features; then use 3×3 convolution to Processing is performed to implement downsampling operation, and we get ; Then use the nearest neighbor upsampling method to process , and Add together and get Layer Features:

[0117] ;

[0118] In the formula, represents the nearest neighbor upsampling method; then the same method is used to obtain Layer Features:

[0119] ;

[0120] but Compared with the original image, the scaled size is ;

[0121] The cross-scale information interaction structure of this embodiment will As multi-scale input features, firstly perform scale normalization operation on features of different scales, and unify the scales of all features through upsampling or downsampling operations:

[0122] ;

[0123] In the formula, Represents the feature scale unification operation, including upsampling and downsampling operations, represents the feature concatenation operation, represents all the concatenated features; next, the feature depth search module constructed in step 2 is used again to perform feature search on the multi-scale features;

[0124] Secondly, the multi-scale features after feature search processing and fusion are re-divided into corresponding scales:

[0125] ;

[0126] In the formula, Represents the division result of the fused features, corresponding to the original scale information, Then, the feature depth search module is used to search for features of different scales, process the divided features, and obtain features for enhancing the original scale features. ;

[0127] Finally, and Perform vector addition operations to obtain multi-scale contour missing target features with cross-scale information interaction .

[0128] Step 4: Extract the correlation between contour areas, contour missing areas, and background areas, enhance the local details of features and the ability to perceive global semantic information, and generate a unified and efficient feature representation ;

[0129] Please see Figure 4 ,In one implementation , a correlation extraction module is used to extract the correlation between the contour area, the contour missing area and the background area to enhance the local details and global semantic information perception ability of the features;

[0130] The relevance extraction module is composed of conventional convolution, feature deep search module and Vision Mamba module. First, the contour area is processed by conventional convolution to extract key features and make up for the missing target contour. Then, the feature deep search module performs feature matching and association analysis between the contour and the contour missing area to accurately capture the target information. Finally, through the global semantic extraction function of the Vision Mamba module, the feature information of different areas is comprehensively processed to generate a unified and efficient feature representation:

[0131] ;

[0132] In the formula, represents the conventional convolution processing, represents the feature deep search module, It means Vision Mamba module;

[0133] Finally, the convolutional layer is used to build a classifier and a bounding box regressor to obtain the category and bounding box information of the target with missing contours:

[0134] ;

[0135] ;

[0136] In the formula, p Indicates the probability that the sample is predicted to be a certain type of foreground target, represents the predicted bounding box, and They represent the convolutional layers for classification and bounding box regression respectively.

[0137] Step 5: Use the classifier and bounding box regressor to predict the category and location information of the target with missing contours;

[0138] In one embodiment, a loss function is designed to train a classifier and a bounding box regressor to predict the category and location information of the object with missing contour;

[0139] During the training process of this embodiment, the loss function of the bounding box regressor is for:

[0140] ;

[0141] In the formula, and Represent the real bounding box and the predicted bounding box respectively;

[0142] Loss function of the classifier for:

[0143] ;

[0144] In the formula, Used to adjust the importance of positive and negative samples in the sub-classifier learning process. Used to improve the sub-classifier's ability to mine difficult samples; when When , it means that the sample is a positive sample. When , it means that the sample is a negative sample. p Indicates the probability that the sample is predicted to be a certain type of foreground target;

[0145] Therefore, the total loss during training is:

[0146] .

[0147] Step 6: Combine the category and positioning information obtained in step 5, eliminate redundant detection boxes, and obtain the final key target positioning information and category information.

[0148] In one embodiment, the frame information of the target to be detected is obtained. ,in, H and W Respectively represent the length and width of the positioning information, and their product is the number of prediction boxes;

[0149] The prediction result of each feature pixel is:

[0150] ;

[0151] in, They represent the distance from the pixel point in the feature map to the boundary of the target detection box, and finally the rectangular frame regression result is:

[0152] ;

[0153] ;

[0154] ;

[0155] ;

[0156] in, is the coordinate of the feature pixel point, is the coordinate of the upper left corner of the rectangular border, The coordinates of the lower right corner of the rectangle.

[0157] In one implementation, a non-maximum suppression method is used to eliminate redundant detection frames; the specific implementation includes the following sub-steps:

[0158] Step 6.1: Group all detection boxes by category and exclude background categories;

[0159] Step 6.2: Sort by the classification scores of the detection boxes within each category;

[0160] Step 6.3: Select the detection box with the highest classification score from the ranking as the final output;

[0161] Step 6.4: Calculate the area of ​​all detection boxes of the output category and evaluate the intersection-over-union ratio of the detection box with the highest classification score with other detection boxes;

[0162] Step 6.5: Delete all detection boxes whose intersection-over-union ratio is lower than the preset threshold;

[0163] Step 6.6: Repeat steps 6.3 to 6.5 until all detection boxes of the output category have been processed (i.e. selected or deleted);

[0164] Step 6.7: Repeat steps 6.2 to 6.6 for each category until the non-maximum suppression process is completed for all categories.

[0165] This embodiment also provides a system for extracting and detecting features of targets with missing contours in severe weather, including:

[0166] one or more processors;

[0167] A storage device is used to store one or more programs. When the one or more programs are executed by the one or more processors, the one or more processors implement the method for extracting and detecting features of targets with missing contours under severe weather conditions.

[0168] Based on the effect of the algorithm implemented by the present invention on key target detection in traffic scenes, Figure 5Some examples are given. A series of experiments verify the detection ability of the contour missing target detection model in complex environments, covering scenes such as haze, heavy rain and low light. The experiment uses a specially processed test set to simulate the visual effects under different severe weather and lighting conditions to evaluate the target detection effect of the model. During the experiment, we focused on analyzing the detection accuracy of the model under conditions of blurred contours, reduced contrast and severe occlusion. The results show that the model can still effectively identify the target contour in a haze environment with high accuracy; in a rainstorm scene, although the raindrop interference is obvious, the model can maintain a high detection consistency; in low light conditions, the model can detect targets with blurred edges through enhanced feature extraction capabilities. The overall experimental results prove that the present invention can quickly and accurately detect targets with missing contours in severe weather, realize intelligent monitoring in complex environments such as haze, heavy rain and low light, or be used to enhance the robustness of unmanned driving systems.

[0169] It should be understood that the embodiments described above are part of the embodiments of the present invention, rather than all of the embodiments. In addition, the technical features in the various embodiments or single embodiments provided by the present invention can be combined with each other arbitrarily to form a feasible technical solution. Such combination is not restricted by the sequence of steps and / or the structural composition mode, but must be based on the ability of ordinary technicians in the field to implement. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the protection scope required by the present invention.

[0170] It should be understood that the above description of the preferred embodiment is relatively detailed and cannot be regarded as limiting the scope of patent protection of the present invention. Under the enlightenment of the present invention, ordinary technicians in this field can also make substitutions or modifications without departing from the scope of protection of the claims of the present invention, which all fall within the scope of protection of the present invention. The scope of protection requested for the present invention shall be based on the attached claims.

Claims

1. A method for extracting and detecting features of targets with missing contours in bad weather, characterized in that: The following steps are involved: Step 1: Extract high-dimensional features of the target video image; Step 2: Search for robust information of the target with missing contours and extract high-dimensional features containing robust information ; Among them, a feature deep search network is used to search for robust information of contour-missing targets; the feature deep search network consists of a five-stage feature extraction structure, each stage consists of a convolutional layer and a feature deep search module, which is used to extract high-dimensional features containing robust information. ; Input high-dimensional features ,in R represents the feature domain, H and W represent the height and width of the feature respectively, and C represents the dimension of the high-dimensional feature; the feature depth search module first uses conventional convolution operations to predict the possible locations of valid features: ； In the formula, is the output of the convolution, represents a convolution operation with a kernel size of 3×3, k represents the kernel size used for convolution within the extraction region; right Perform a split operation to divide it into bias vectors of deformable convolutions used for regional feature extraction , and the position vector used to search for the valid area , ; Afterwards, As the bias tensor input into the deformable convolution, the local information of the current position is extracted, and the area search process is performed. The search results are: ； In the formula, Represents the average operation, the vector and the cut out By summing the vectors, we can get more accurate position information; Then, the searched location information is passed to the next deformable convolution to perform feature extraction on the proposed valid area: ； In the formula, S It is set as a learnable parameter to control the search radius of this area search, and its initialization value represents the initial search radius; Finally, continue searching for the next area containing valid information and determine the search direction: ； ； Thus, robust information of the target with missing contour is extracted; The feature deep search network combines the convolution and feature search modules to construct a search network of five stages from stage 0 to stage 4, and the outputs of stages 2 to 4 are As input to step 3; Step 3: Target high-dimensional features that contain robust information , adaptively search for information at different scales, and obtain multi-scale contour missing target features with cross-scale information interaction ; Step 4: Extract the correlation between contour area, contour missing area and background area, and generate feature representation Fed ; Step 5: Use the classifier and bounding box regressor to predict the category and location information of the target with missing contours; Step 6: Combine the category and positioning information obtained in step 5, eliminate redundant detection boxes, and obtain the final key target positioning information and category information.

2. The method for extracting and detecting features of targets with missing contours in bad weather according to claim 1, characterized in that: In step 1, a parallel atrous convolution structure with multiple receptive fields is constructed to extract high-dimensional features with strong expressive power; The parallel dilated convolution structure is composed of dilated convolutions with dilated rates of 1, 2, and 3 respectively. The input image is defined as ,in R represents the feature domain, H and W represent the height and width of the feature respectively, which are input into the parallel dilated convolution structure to obtain high-dimensional features: ； In the formula, and They represent dilated convolutions with dilation rates of 1, 2, and 3 respectively, and C represents the dimension of high-dimensional features.

3. The method for extracting and detecting target features with missing contours in bad weather according to claim 1, characterized in that: In step 3, a multi-scale feature pyramid and a cross-scale information interaction structure are constructed to adaptively search for information at different scales; The multi-scale feature pyramid first converts the high-dimensional features containing robust information into As input, and As Layer features; then Using 3×3 convolution Processing is performed to implement downsampling operation, and we get ; Then use the nearest neighbor upsampling method to process , and with Add together and get Layer Features: ； In the formula, represents the nearest neighbor upsampling method; then the same method is used to obtain Layer Features: ； but Compared with the original image, the scaled size is ; The cross-scale information interaction structure will As multi-scale input features, firstly perform scale normalization operation on features of different scales, and unify the scales of all features through upsampling or downsampling operations: ； In the formula, Represents the feature scale unification operation, including upsampling and downsampling operations, represents the feature concatenation operation, represents all the features after splicing; next, the feature depth search module constructed in step 2 is used again to perform feature search on multi-scale features; Secondly, the multi-scale features after feature search processing and fusion are re-divided into corresponding scales: ； In the formula, Represents the division result of the fused features, corresponding to the original scale information, Represents a partition operation; Then, the feature depth search module is used to search for features of different scales, process the divided features, and obtain features for enhancing the original scale features. ; Finally, and Perform vector addition operations to obtain multi-scale contour missing target features with cross-scale information interaction .

4. The method for extracting and detecting features of targets with missing contours in bad weather according to claim 1, characterized in that: In step 4, a correlation extraction module is used to extract the correlation between the contour area, the contour missing area and the background area, so as to enhance the local details of the features and the global semantic information perception ability; The association extraction module is composed of conventional convolution, feature depth search module and Vision Mamba module. First, the contour area is processed by conventional convolution to extract key features and make up for the missing target contour. Then, the feature depth search module performs feature matching and association analysis between the contour and the contour missing area to accurately capture the target information. Finally, the feature information of different areas is comprehensively processed through the global semantic extraction function of the Vision Mamba module to generate a feature representation. Fed : ； In the formula, represents the conventional convolution processing, represents the feature deep search module, It indicates the VisionMamba module; Finally, the convolutional layer is used to build a classifier and a bounding box regressor to obtain the category and bounding box information of the target with missing contours: ； ； In the formula, p Indicates the probability that a sample is predicted to be a certain type of foreground target, represents the predicted bounding box, and Represent the convolutional layers for classification and bounding box regression respectively.

5. The method for extracting and detecting features of targets with missing contours in bad weather according to claim 1, characterized in that: In step 5, the loss function is designed, and the classifier and bounding box regressor are trained to predict the category and location information of the target with missing contours. During training, the loss function of the bounding box regressor is for: ； In the formula, and Represent the real bounding box and the predicted bounding box respectively; Loss function of the classifier for: ； In the formula, Used to adjust the importance of positive and negative samples in the sub-classifier learning process. Used to improve the sub-classifier's ability to mine difficult samples; when When , it means that the sample is a positive sample. When , it means that the sample is a negative sample. p Indicates the probability that the sample is predicted to be a certain type of foreground target; The total loss during training is: 。 6. The method for extracting and detecting features of targets with missing contours in bad weather according to claim 1, characterized in that: In step 6, the bounding box information of the target to be detected is obtained ,in, H and W Respectively represent the length and width of the positioning information, and their product is the number of prediction boxes; The prediction result of each feature pixel is: ； in, They represent the distance from the pixel point in the feature map to the boundary of the target detection box, and finally the rectangular frame regression result is: ； ； ； ； in, is the coordinate of the feature pixel point, is the coordinate of the upper left corner of the rectangular border, The coordinates of the lower right corner of the rectangle.

7. The method for extracting and detecting features of targets with missing contours in bad weather according to any one of claims 1 to 6, characterized in that: In step 6, the non-maximum suppression method is used to eliminate redundant detection frames; the specific implementation includes the following sub-steps: Step 6.1: Group all detection boxes by category and exclude background categories; Step 6.2: Sort by the classification scores of the detection boxes within each category; Step 6.3: Select the detection box with the highest classification score from the ranking as the final output; Step 6.4: Calculate the area of ​​all detection boxes of the output category and evaluate the intersection-over-union ratio of the detection box with the highest classification score with other detection boxes; Step 6.5: Delete all detection boxes whose intersection-over-union ratio is lower than the preset threshold; Step 6.6: Repeat steps 6.3 to 6.5 until all detection boxes of the output category have been processed; Step 6.7: Repeat steps 6.2 to 6.6 for each category until the non-maximum suppression process is completed for all categories.

8. A system for extracting and detecting features of targets with missing contours in bad weather, characterized in that: include: one or more processors; A storage device for storing one or more programs, when the one or more programs are executed by the one or more processors, enables the one or more processors to implement the method for extracting and detecting features of targets with missing contours in severe weather as described in any one of claims 1 to 7.

Images