A multi-scale target detection method based on environment perception and channel gating mechanism

Abstract

The application relates to the fields of computer vision and intelligent traffic technology, and discloses a multi-scale target detection method based on an environment perception and a channel gating mechanism, which comprises the following steps: using a backbone network and an aggregation factor technology to output a multi-scale feature map corresponding to an enhanced traffic sign image and an aging state of the traffic sign; performing fusion processing on the multi-scale features to generate a feature pyramid, and inputting the feature pyramid into a detection model to obtain a traffic sign prediction frame and traffic sign category information; judging the accuracy of the traffic sign prediction frame and the traffic sign category information, correcting the detection model based on the accuracy result, and performing export tamper-proof processing on the detection model after the correction is completed, so that the detection of the traffic sign in a high-speed scene is realized by using the monitoring model and the aging state of the traffic sign to output results. The channel gating mechanism is introduced to realize adaptive fusion of deep and shallow features, so that the small target detection performance is significantly improved.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and intelligent transportation technology, and more specifically, to a multi-scale target detection method based on environmental perception and channel gating mechanisms. Background Technology

[0002] In intelligent transportation systems, traffic sign detection and recognition are crucial for vehicle environmental perception, playing a vital role in autonomous driving, road monitoring, and traffic safety early warning. Existing traffic sign detection methods largely rely on convolutional neural network (CNN) structures, such as Faster R-CNN, Cascade R-CNN, and YOLO. However, these methods still suffer from the following problems in detecting small objects in high-speed driving scenarios:

[0003] Because small traffic signs at long distances occupy a very small proportion of an image, their feature representation is insufficient, making it difficult to effectively distinguish between the background and the target. Furthermore, existing feature pyramid structures (FPN) use simple weighting or summation methods for multi-scale feature fusion, lacking adaptive selection capabilities between channels, leading to information redundancy and oversmoothing issues between features at different levels. In addition, during multi-scale fusion, the weight allocation of shallow and deep features cannot be dynamically adjusted according to the content of the input image, thus limiting the sensitivity of the detection model to small targets.

[0004] There are currently no effective solutions to the problems in the relevant technologies. Summary of the Invention

[0005] To address the problems in related technologies, this invention proposes a multi-scale target detection method based on environmental perception and channel gating mechanisms to overcome the aforementioned technical problems in existing related technologies.

[0006] Therefore, the specific technical solution adopted by the present invention is as follows:

[0007] In a first aspect, the present invention provides a multi-scale target detection method based on environmental perception and channel gating mechanisms, the method comprising:

[0008] The traffic sign image to be detected is segmented based on brightness and contrast. Based on the segmentation results, the traffic sign image to be detected is adaptively optimized to generate an enhanced traffic sign image. The enhanced traffic sign image is then processed using a backbone network and aggregation factor technology to output the multi-scale feature map corresponding to the enhanced traffic sign image and the aging status of the traffic sign.

[0009] Multi-scale features are fused using upsampling and layer-by-layer fusion methods to generate a feature pyramid. The feature pyramid is then input into the detection model to obtain traffic sign prediction boxes and traffic sign category information.

[0010] The accuracy of the traffic sign prediction bounding box and traffic sign category information is judged, the detection model is corrected based on the accuracy results, and the detection model is exported and tamper-proofed after correction. The detection of traffic signs in high-speed scenarios is realized by using the monitoring model and the aging state output of traffic signs.

[0011] Preferably, the traffic sign image to be detected is segmented based on brightness and contrast. Based on the segmentation results, adaptive optimization processing is performed on the traffic sign image to be detected to generate an enhanced traffic sign image. The enhanced traffic sign image is then processed using a backbone network and aggregation factor techniques, outputting a multi-scale feature map corresponding to the enhanced traffic sign image and the aging state of the traffic sign, including:

[0012] Calculate the average brightness and variance of the grayscale image of the traffic sign to be detected, and compare the average brightness with the brightness threshold and the variance with the contrast threshold.

[0013] If the average brightness is less than the brightness threshold, it indicates that the traffic sign image to be detected is in a low-light condition. In this case, the image enhancement theory and technology are used to perform low-light enhancement processing on the traffic sign image to be detected.

[0014] If the variance is less than the contrast threshold and the average brightness is greater than or equal to the brightness threshold, then the traffic sign image to be detected is determined to be in a foggy situation, and the dark channel prior algorithm is used to perform defogging and noise reduction processing on the traffic sign image to be detected.

[0015] The consistency verification mechanism is used to calculate the consistency between the low-light enhancement and defogging and denoising processing of the traffic sign image to be detected. The enhanced traffic sign image is output after the consistency result meets the requirements. At the same time, the aging state of the traffic sign in the enhanced traffic sign image is analyzed.

[0016] A self-distillation mechanism is used to train a self-supervised learning network to generate a backbone network. Enhanced traffic sign images are then input into the backbone network, and hierarchical convolution extraction techniques are applied to the backbone network to output multi-scale feature maps.

[0017] Preferably, the low-light enhancement processing of the traffic sign image to be detected using image enhancement theory techniques includes:

[0018] By convolving the traffic sign image under low light conditions with Gaussian kernels of different scales, the ambient brightness estimation components under different receptive fields are extracted, and the reflection components are calculated in the number domain based on the ambient brightness estimation components. The uneven illumination problem of the traffic sign image under the detection is eliminated based on the reflection components.

[0019] The dark channel prior algorithm is used to perform dehazing and noise reduction processing on the traffic sign image to be detected, including:

[0020] The minimum values ​​of the three primary color channels are searched within a local window of the traffic sign image to be detected in foggy conditions. The initial transmittance is estimated by combining the atmospheric light value. The four-directional gradient operator is used as the weight constraint of the guided filter to refine the initial transmittance, so as to complete the defogging process of the traffic sign image to be detected in foggy conditions.

[0021] Preferably, the consistency between the low-light enhancement and the dehazing and denoising processed traffic sign image is calculated based on the consistency verification mechanism, and the enhanced traffic sign image is output after the consistency result meets the requirements. Simultaneously, the aging state of the traffic signs in the enhanced traffic sign image is analyzed, including:

[0022] The structural similarity of the traffic sign image to be detected before and after low light enhancement and before and after defogging and noise reduction is calculated based on pixel mean, covariance and stability coefficient, and the difference between the structural similarity and the similarity threshold is calculated.

[0023] If the structural similarity is less than the similarity threshold, it indicates that the traffic sign images to be detected are inconsistent before and after low-light enhancement and before and after defogging and denoising. In this case, the traffic sign images to be detected before and after low-light enhancement and before and after defogging and denoising are weighted and fused to output an enhanced traffic sign image.

[0024] Select several pixels in the enhanced traffic sign image as points of interest, and use the standard deviation of the points of interest in the enhanced traffic sign image to calculate the aggregation factor to identify the crack texture of the traffic sign.

[0025] The fuzzy maximum entropy method is used to calculate the separation points between the target class and the background class in the enhanced traffic sign image, and the optimal classification surface is constructed based on the separation points to determine the degree of color attenuation in the enhanced traffic sign image.

[0026] The aging status of traffic signs is determined by examining the crack texture of the traffic signs and the degree of color decay in the enhanced traffic sign image.

[0027] Preferably, selecting several pixels in the enhanced traffic sign image as points of interest, and using the standard deviation of these points in the enhanced traffic sign image to calculate the aggregation factor to identify the crack texture of the traffic sign includes:

[0028] The enhanced traffic sign image is subjected to grayscale difference enhancement processing. After the processing is completed, the grayscale change rate of each pixel in the enhanced traffic sign image is calculated and the grayscale change rates are sorted in descending order.

[0029] Based on the sorting results, select several pixels with the largest gray-scale change rate as points of interest, establish a coordinate system based on the enhanced traffic sign image, and calculate the standard deviation of the points of interest in the horizontal and vertical directions of the coordinate system to reflect the dispersion of the points of interest.

[0030] The standard deviation is normalized to define the aggregation factors in the horizontal and vertical directions. A scatter plot of the aggregation factor distribution is constructed by combining traffic sign sample images with and without cracks. The elliptical arc boundary is determined by statistical analysis based on the scatter plot.

[0031] Crack warning regions are defined based on the area enclosed by the elliptical arc boundary and the coordinate system, and the crack texture of traffic signs in enhanced traffic sign images is identified based on the aggregation factor falling into the crack warning regions.

[0032] Preferably, the fuzzy maximum entropy method is used to calculate the separation points between the target class and the background class in the enhanced traffic sign image, and the optimal classification surface is constructed based on the separation points to determine the degree of color attenuation of the enhanced traffic sign image, including:

[0033] Some enhanced traffic sign images are converted into grayscale images, and the grayscale matrix of the grayscale image is obtained. The grayscale matrix is ​​then mapped into a fuzzy number matrix through a membership function to construct a fuzzy set of target class and background class.

[0034] Calculate the average gray value of the fuzzy sets of the target class and the background class respectively, calculate the fuzzy entropy using the principle of maximum fuzzy entropy, select the average gray value that maximizes the fuzzy entropy, and take the corresponding average gray value as the best separation point between the target class and the background class.

[0035] The maximum ratio of the three primary colors of the target class is extracted using the optimal separation point as the boundary, and the maximum ratio is used as the sample feature. Combined with the standard color benchmark value, a training sample set is constructed to build the optimal classification surface.

[0036] Input the maximum ratios of the three primary colors of the remaining enhanced traffic sign images into the optimal classification surface, determine the changes in the maximum ratios of the three primary colors in the enhanced traffic sign images, and output the quantification results of the color attenuation degree.

[0037] Preferably, multi-scale features are fused using upsampling and layer-by-layer fusion methods to generate a feature pyramid. This feature pyramid is then input into the detection model to obtain traffic sign prediction boxes and traffic sign category information, including:

[0038] The number of channels of multi-scale features is unified by using target convolution, and the multi-scale features are fused layer by layer from high to low according to the stage layer of the multi-scale features. The deep features corresponding to the multi-scale features are calculated based on the fusion results.

[0039] Based on deep features and channel number, a channel-level gating mechanism is used for weighted fusion. The fused feature map is output as a feature pyramid, and the feature pyramid is input into a detection model consisting of three consecutive bounding box regression and classification sub-networks.

[0040] The detection model is used to initially generate candidate boxes and classification results corresponding to the feature pyramid, and the candidate boxes and classification results are used as inputs to optimize the bounding boxes and class confidence within the candidate box region.

[0041] The bounding box and category confidence optimization results are used as inputs, and the candidate boxes and classification results are fused for real-time correction. Based on the correction results, the traffic sign prediction box and traffic sign category information are obtained.

[0042] Preferably, the accuracy of the traffic sign prediction bounding box and traffic sign category information is determined, the detection model is corrected based on the accuracy results, and after correction, the detection model is exported with anti-tampering processing. This allows for the detection of traffic signs in high-speed scenarios using the monitoring model and the aging state output of the traffic signs.

[0043] Calculate the actual aspect ratio of the traffic sign prediction frame, and calculate the relative error between the actual aspect ratio and the standard aspect ratio. Compare the relative error with the error threshold.

[0044] If the relative error is less than or equal to the error threshold, it means that there is no anomaly in the traffic sign prediction box. If the relative error is greater than the error threshold, it means that there is anomaly in the traffic sign prediction box. The original confidence of the detection model is then corrected, and the traffic sign prediction box and traffic sign category information are re-output.

[0045] The corrected detection model is exported as a neural network exchange format. During the export process, the key layer weights of the corrected detection model are sampled and calculated to generate a unique hash fingerprint.

[0046] The hash fingerprint is embedded into the encrypted header file as a security benchmark for the operation of the corrected detection model. The corrected detection model and the aging state output of traffic signs are used to realize the detection of traffic signs in high-speed scenarios.

[0047] Secondly, the present invention also provides a multi-scale target detection system based on environmental perception and channel gating mechanisms, the system comprising:

[0048] The feature map and aging state output module is used to segment the traffic sign image to be detected based on brightness and contrast. Based on the segmentation result, the traffic sign image to be detected is adaptively optimized to generate an enhanced traffic sign image. The enhanced traffic sign image is then processed using a backbone network and aggregation factor technology to output the multi-scale feature map and aging state of the traffic sign corresponding to the enhanced traffic sign image.

[0049] The traffic sign prediction box and category information output module is used to fuse multi-scale features using upsampling and layer-by-layer fusion methods to generate a feature pyramid, and input the feature pyramid into the detection model to obtain traffic sign prediction boxes and traffic sign category information.

[0050] The detection model correction and anti-tampering processing module is used to determine the accuracy of the traffic sign prediction box and traffic sign category information, correct the detection model based on the accuracy results, and perform anti-tampering processing on the detection model after correction, so as to realize the detection of traffic signs in high-speed scenarios by using the monitoring model and the aging state output of traffic signs.

[0051] The beneficial effects of this invention are as follows:

[0052] This invention achieves adaptive fusion of deep and shallow features by introducing a channel gating mechanism, thereby significantly improving the performance of small target detection. It can adapt to various complex environments and has strong robustness to images in complex environments. The backbone network enhances the advantage of image feature extraction, and the combination with CG-FPN provides stronger feature fusion capabilities. It can autonomously select the fusion ratio of deep and shallow features, making it interpretable and having strong generalization performance. It is not only applicable to highway scenarios but can also be extended to other small target detection fields. Attached Figure Description

[0053] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0054] Figure 1 This is a flowchart of a multi-scale target detection method based on environmental perception and channel gating mechanism according to an embodiment of the present invention;

[0055] Figure 2 This is an architecture diagram of a multi-scale target detection method based on environmental perception and channel gating mechanism according to an embodiment of the present invention;

[0056] Figure 3 This is one of the result illustrations in a multi-scale target detection method based on environmental perception and channel gating mechanism according to an embodiment of the present invention;

[0057] Figure 4 This is the second result illustration in a multi-scale target detection method based on environmental perception and channel gating mechanism according to an embodiment of the present invention;

[0058] Figure 5 This is a comparison diagram of the feature responses of FPN and CG-FPN in a multi-scale target detection method based on environmental perception and channel gating mechanism according to an embodiment of the present invention. Detailed Implementation

[0059] To further illustrate the various embodiments, the present invention provides accompanying drawings, which are part of the disclosure of the present invention. These drawings are mainly used to illustrate the embodiments and can be used in conjunction with the relevant descriptions in the specification to explain the operating principles of the embodiments. With reference to these drawings, those skilled in the art should be able to understand other possible implementation methods and the advantages of the present invention.

[0060] According to an embodiment of the present invention, a multi-scale target detection method based on environmental perception and channel gating mechanism is provided.

[0061] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments, such as... Figure 1 As shown, according to an embodiment of the present invention, a multi-scale target detection method based on environmental perception and channel gating mechanism includes:

[0062] Step S1: Segment the traffic sign image to be detected based on brightness and contrast. Perform adaptive optimization processing on the traffic sign image to be detected according to the segmentation results to generate an enhanced traffic sign image. Then, use the backbone network and aggregation factor technology to process the enhanced traffic sign image and output the multi-scale feature map corresponding to the enhanced traffic sign image and the aging state of the traffic sign.

[0063] In one embodiment, the traffic sign image to be detected is segmented based on brightness and contrast. Adaptive optimization processing is then performed on the segmentation results to generate an enhanced traffic sign image. The enhanced traffic sign image is further processed using a backbone network and aggregation factor techniques. The output includes a multi-scale feature map corresponding to the enhanced traffic sign image and the aging state of the traffic sign, including:

[0064] Calculate the average brightness and variance of the grayscale image of the traffic sign to be detected, and compare the average brightness with the brightness threshold and the variance with the contrast threshold.

[0065] If the average brightness is less than the brightness threshold, it indicates that the traffic sign image to be detected is in a low-light condition. In this case, the image enhancement theory and technology are used to perform low-light enhancement processing on the traffic sign image to be detected.

[0066] If the variance is less than the contrast threshold and the average brightness is greater than or equal to the brightness threshold, then the traffic sign image to be detected is determined to be in a foggy situation, and the dark channel prior algorithm is used to perform defogging and noise reduction processing on the traffic sign image to be detected.

[0067] The consistency verification mechanism is used to calculate the consistency between the low-light enhancement and defogging and denoising processing of the traffic sign image to be detected. The enhanced traffic sign image is output after the consistency result meets the requirements. At the same time, the aging state of the traffic sign in the enhanced traffic sign image is analyzed.

[0068] A self-distillation mechanism is used to train a self-supervised learning network to generate a backbone network. Enhanced traffic sign images are then input into the backbone network, and hierarchical convolution extraction techniques are applied to the backbone network to output multi-scale feature maps.

[0069] In one embodiment, the low-light enhancement processing of the traffic sign image to be detected using image enhancement theory techniques includes:

[0070] By convolving the traffic sign image under low light conditions with Gaussian kernels of different scales, the ambient brightness estimation components under different receptive fields are extracted, and the reflection components are calculated in the number domain based on the ambient brightness estimation components. The uneven illumination problem of the traffic sign image under the detection is eliminated based on the reflection components.

[0071] The dark channel prior algorithm is used to perform dehazing and noise reduction processing on the traffic sign image to be detected, including:

[0072] The minimum values ​​of the three primary color channels are searched within a local window of the traffic sign image to be detected in foggy conditions. The initial transmittance is estimated by combining the atmospheric light value. The four-directional gradient operator is used as the weight constraint of the guided filter to refine the initial transmittance, so as to complete the defogging process of the traffic sign image to be detected in foggy conditions.

[0073] In one embodiment, the consistency between the low-light enhancement and the dehazing and denoising processed traffic sign image is calculated based on a consistency verification mechanism. Once the consistency result meets the requirements, the enhanced traffic sign image is output. Simultaneously, the aging state of the traffic signs in the enhanced traffic sign image is analyzed, including:

[0074] The structural similarity of the traffic sign image to be detected before and after low light enhancement and before and after defogging and noise reduction is calculated based on pixel mean, covariance and stability coefficient, and the difference between the structural similarity and the similarity threshold is calculated.

[0075] If the structural similarity is less than the similarity threshold, it indicates that the traffic sign images to be detected are inconsistent before and after low-light enhancement and before and after defogging and denoising. In this case, the traffic sign images to be detected before and after low-light enhancement and before and after defogging and denoising are weighted and fused to output an enhanced traffic sign image.

[0076] Select several pixels in the enhanced traffic sign image as points of interest, and use the standard deviation of the points of interest in the enhanced traffic sign image to calculate the aggregation factor to identify the crack texture of the traffic sign.

[0077] The fuzzy maximum entropy method is used to calculate the separation points between the target class and the background class in the enhanced traffic sign image, and the optimal classification surface is constructed based on the separation points to determine the degree of color attenuation in the enhanced traffic sign image.

[0078] The aging status of traffic signs is determined by examining the crack texture of the traffic signs and the degree of color decay in the enhanced traffic sign image.

[0079] In one embodiment, selecting several pixels in the enhanced traffic sign image as points of interest, and using the standard deviation of the points of interest in the enhanced traffic sign image to calculate the aggregation factor to identify the crack texture of the traffic sign includes:

[0080] The enhanced traffic sign image is subjected to grayscale difference enhancement processing. After the processing is completed, the grayscale change rate of each pixel in the enhanced traffic sign image is calculated and the grayscale change rates are sorted in descending order.

[0081] Based on the sorting results, select several pixels with the largest gray-scale change rate as points of interest, establish a coordinate system based on the enhanced traffic sign image, and calculate the standard deviation of the points of interest in the horizontal and vertical directions of the coordinate system to reflect the dispersion of the points of interest.

[0082] The standard deviation is normalized to define the aggregation factors in the horizontal and vertical directions. A scatter plot of the aggregation factor distribution is constructed by combining traffic sign sample images with and without cracks. The elliptical arc boundary is determined by statistical analysis based on the scatter plot.

[0083] Crack warning regions are defined based on the area enclosed by the elliptical arc boundary and the coordinate system, and the crack texture of traffic signs in enhanced traffic sign images is identified based on the aggregation factor falling into the crack warning regions.

[0084] It needs to be explained that in the process of recognizing traffic sign crack texture based on interest point standard deviation and aggregation factor, the enhanced traffic sign image is processed by calculating the gray-level difference between a pixel and its neighboring pixels and amplifying it (e.g., using a gray-level difference operator or adaptive enhancement algorithm). Then, each pixel in the image is traversed, and a 3×3 or 5×5 neighborhood window is selected with that pixel as the center. The sum of the absolute gray-level differences between all pixels in the window and the center pixel is calculated. Then, the difference is amplified by a preset enhancement coefficient and superimposed back onto the original pixel gray-level value, finally obtaining a gray-level difference enhanced image. This can amplify the gray-level contrast between the crack area and the normal area in the traffic sign image. Since the crack texture is essentially a dense area of ​​gray-level abrupt changes, while the gray-level changes in the normal area are gradual, the enhancement process can make the gray-level change features at the crack more prominent. This solves the problem that the gray-level difference of the crack texture in the original image is not obvious and is easily masked by background noise, laying a feature foundation for subsequent interest point extraction.

[0085] Meanwhile, the calculation of grayscale change rate needs to be based on the image after grayscale difference enhancement. For each pixel, first calculate its grayscale gradient in the four directions of horizontal, vertical and diagonal (the gradient magnitude can be calculated using the Prewitt operator). The average of the gradient magnitudes in the four directions is taken as the grayscale change rate of the pixel. After completing the calculation of the grayscale change rate of all pixels, sort the grayscale change rate values ​​of all pixels in descending order from large to small, while retaining the pixel coordinates corresponding to each value. This allows for the quantification of the degree of grayscale change of each pixel and the selection of the candidate set of pixels with the most dramatic grayscale changes.

[0086] Based on the descending sort results, pixels with the largest grayscale change rate are selected as points of interest according to a preset ratio (such as the top 5% or top 10%). At the same time, a two-dimensional coordinate system is established with the top left corner of the enhanced image as the origin, the horizontal axis to the right as the x-axis, and the vertical axis downward as the y-axis. The (a,b) coordinates of all points of interest are recorded. The standard deviations of the x-coordinates and y-coordinates of all points of interest are calculated respectively. Then, the dispersion of the points of interest in the horizontal and vertical directions is quantified by the standard deviation. The points of interest in the crack area are distributed along the crack direction, and their horizontal / vertical standard deviations will show specific discrete characteristics. The points of interest in the crack-free area are mostly random noise, and the dispersion is irregular.

[0087] We collected a large number of samples of traffic signs with and without cracks, and the maximum and minimum standard deviations of the x-coordinates and y-coordinates of the points of interest. We then used the min-max normalization formula to convert the standard deviations of the x-coordinates and y-coordinates into values ​​between 0 and 1, and defined a lateral aggregation factor C. a =1-σ a '(x-coordinate of the point of interest σ)' a The smaller the value, the higher the lateral aggregation factor C. a The larger the value, the more concentrated the points of interest are along the horizontal axis; the vertical clustering factor C... b =1-σ b ',σ b The standard deviation of the ordinate of the point of interest is denoted as 0. The clustering factor ranges from 0 to 1. The larger the value, the higher the degree of clustering of the point of interest in the corresponding direction. This can eliminate the influence of traffic sign images of different sizes and shooting angles on the standard deviation value and transform the standard deviation into a clustering factor of a uniform scale.

[0088] Collect a sufficient number of traffic sign samples with cracks (positive samples) and without cracks (negative samples) (C a C b Data pairs, in C a For the horizontal axis, C bA two-dimensional scatter plot is drawn with the vertical axis. The aggregation factor of positive samples will be concentrated in a specific area of ​​the scatter plot due to the spatial distribution characteristics of the cracks, while the negative samples are randomly distributed. An ellipse is fitted to the scatter distribution of positive samples using statistical analysis (such as kernel density estimation and ellipse fitting algorithms) to determine the elliptical arc boundary that covers the vast majority of positive samples. The equation of the ellipse is [(C a -c) 2 / d 2 ]+[(C b -e) 2 / f 2 ]=1 (where c and e are the coordinates of the center of the ellipse, and d and f are the major and minor axes of the ellipse), and then the distribution range of the crack sample aggregation factor is defined by statistical methods.

[0089] The region enclosed by the elliptical arc boundary and the coordinate system is defined as the crack early warning region, i.e., the aggregation factor (C). a C b The area falling within the boundary of the elliptical arc is identified as a crack warning area, and the area outside is identified as a non-warning area; for the enhanced traffic sign image to be identified, the aggregation factor (C) of its points of interest is calculated. a C b If the aggregation factor falls into the crack warning area, it is determined that there is a crack texture in the image. The closer the aggregation factor is to the center of the ellipse, the more obvious the crack feature is. If it does not fall into the warning area, it is determined that there is no obvious crack texture.

[0090] In one embodiment, the fuzzy maximum entropy method is used to calculate the separation points between the target class and the background class in the enhanced traffic sign image, and the optimal classification surface is constructed based on the separation points to determine the degree of color attenuation of the enhanced traffic sign image, including:

[0091] Some enhanced traffic sign images are converted into grayscale images, and the grayscale matrix of the grayscale image is obtained. The grayscale matrix is ​​then mapped into a fuzzy number matrix through a membership function to construct a fuzzy set of target class and background class.

[0092] Calculate the average gray value of the fuzzy sets of the target class and the background class respectively, calculate the fuzzy entropy using the principle of maximum fuzzy entropy, select the average gray value that maximizes the fuzzy entropy, and take the corresponding average gray value as the best separation point between the target class and the background class.

[0093] The maximum ratio of the three primary colors of the target class is extracted using the optimal separation point as the boundary, and the maximum ratio is used as the sample feature. Combined with the standard color benchmark value, a training sample set is constructed to build the optimal classification surface.

[0094] Input the maximum ratios of the three primary colors of the remaining enhanced traffic sign images into the optimal classification surface, determine the changes in the maximum ratios of the three primary colors in the enhanced traffic sign images, and output the quantification results of the color attenuation degree.

[0095] It needs to be explained that in the process of determining the separation point between the target and the background based on the fuzzy maximum entropy method and identifying the degree of color decay of traffic signs by combining the best classification surface, some enhanced traffic sign images are converted into grayscale images and a fuzzy set is constructed. A certain number of enhanced traffic sign sample images (as training set) are selected, and the color enhanced images are converted into single-channel grayscale images by using the grayscale formula (grayscale value = 0.299×R + 0.587×G + 0.114×B). The two-dimensional grayscale matrix of the grayscale image is extracted (each element in the matrix corresponds to the grayscale value of the pixel from 0 to 255). At the same time, a membership function is selected to map the grayscale value of each pixel in the grayscale matrix to a membership value in the range of 0-1, forming a fuzzy number matrix. According to the membership threshold (such as 0.5), the fuzzy number matrix is ​​divided into a target class fuzzy set (traffic sign main area, membership ≥ 0.5) and a background class fuzzy set (background area, membership < 0.5). Thus, the hard segmentation of the grayscale image can be transformed into fuzzy segmentation, weakening the absolute boundary between the target and the background.

[0096] Simultaneously, the average gray value of the target class fuzzy set (i.e., the weighted sum of all pixel gray values ​​of the target class × the corresponding membership degree / the total number of pixels in the target class) and the average gray value of the background class fuzzy set are calculated separately. Based on the principle of maximum fuzzy entropy, the fuzzy entropy is calculated. The larger the entropy value, the lower the uncertainty of the fuzzy division and the better the classification effect. By traversing all gray values ​​from 0 to 255 as candidate thresholds, the membership function is adjusted and the average gray value of the target class fuzzy set, the average gray value of the background class fuzzy set, and the fuzzy entropy are recalculated. The average gray value of the target class fuzzy set or the average gray value of the background class fuzzy set that maximizes the fuzzy entropy is selected as the optimal separation point (gray value threshold) between the target class and the background class. Thus, the optimal gray value segmentation threshold between the target and the background can be found from the perspective of information theory.

[0097] Using the optimal separation point as the grayscale threshold, the color-enhanced image is segmented to accurately extract the target region (traffic sign subject). The red (R), green (G), and blue (B) component values ​​of all pixels in this region are statistically analyzed, and the ratios of R / G, R / B, and G / B are calculated respectively. The ratio with the largest value is selected as the maximum ratio of the three primary colors for this sample (e.g., if a sample has R / G=2.5, R / B=3.0, and G / B=1.2, then the maximum ratio is 3.0). At the same time, the standard color reference value of the traffic sign without attenuation is obtained (e.g., the RGB standard values ​​of red, yellow, and blue signs as specified in national standards). The maximum ratio of the three primary colors corresponding to the standard reference value is calculated as a reference value. The difference / ratio between the actual maximum ratio of the three primary colors and the standard maximum ratio of all samples in the training set is used as the sample feature. The feature values ​​of all training samples and the corresponding color attenuation labels (e.g., no attenuation, slight attenuation, moderate attenuation, severe attenuation) are integrated to construct a complete training sample set, which can then be transformed into quantifiable numerical features.

[0098] Support Vector Machines (SVMs) are used to train the training sample set. The maximum ratios of the three primary colors are used as input, and color attenuation labels are used as output. A kernel function (such as a radial basis function, RBF) maps the low-dimensional features to a high-dimensional space. The hyperplane that maximizes the margin between samples of different attenuation categories, i.e., the optimal classification surface, is then solved. The remaining enhanced traffic sign images (test set) that were not used in training are processed in the same way to extract the maximum ratios of the target class's three primary colors, and these are input into the trained optimal classification surface. The classification surface calculates the position of the maximum ratio in the feature space based on the input maximum ratio, determining the color attenuation category of the sample. Simultaneously, by calculating the deviation between the actual maximum ratio and the standard maximum ratio, the degree of attenuation is quantified into a specific numerical value (e.g., deviation <5% indicates no attenuation, 5%-15% indicates mild attenuation, 15%-30% indicates moderate attenuation, and >30% indicates severe attenuation). Finally, the quantified result is output, achieving automated and quantitative determination of the degree of color attenuation.

[0099] like Figure 2 As shown, the image to be processed (i.e., the traffic sign image to be detected) is first input into the environment perception and image enhancement. Let the input image be I, and calculate its grayscale image's average brightness (Lavg) and variance (σ). 2 Set the brightness threshold T light and contrast threshold T contrast If T contrast <T light If the light level is low, such as at night, then the improved multi-scale Retinex algorithm is used for low-light enhancement. If σ 2 <T contrast And Lavg≥T lighIf the image is foggy, the image contrast is low and the features are blurry. The system uses an improved dark channel prior algorithm for defogging and noise reduction.

[0100] For the improved multi-scale Retinex enhancement algorithm, the input image I(x,y) is first convolved with three Gaussian kernels of different scales to extract the ambient brightness estimation component L under different receptive fields. i Then, the reflection component R is calculated in the logarithmic domain. k (x,y)=logI(x,y)-logL i (x,y) is used to eliminate the influence of uneven illumination, where I(x,y) represents the pixel value of the input image at coordinates (x,y). To overcome the color distortion problem caused by the traditional method in processing the target, this embodiment introduces a dynamic color restoration factor based on the standard deviation of local pixel distribution as the weighting function W(x,y)=f(std(I local This function can dynamically adjust the balance coefficients of the three channels according to the color richness of local areas of the image. While enhancing the brightness of dark areas, it forces the saturation of the sign's characteristic colors to be consistent with their physical properties, effectively avoiding post-classification errors caused by color shift.

[0101] For low-contrast foggy scenes, this embodiment uses an improved dark channel prior algorithm for sharpening. First, it searches for the minimum value of the R, G, and B channels within a local window to obtain a preliminary dark channel image. Then, it estimates the initial transmittance by combining the atmospheric light value A. To solve the halo phenomenon caused by the traditional DCP algorithm at the edges of small targets, this embodiment innovatively adopts gradient weighted guided filtering technology. By introducing a four-directional Sobel gradient operator as the weight constraint of the guided filter, it can automatically identify and protect the edge features of the target object when refining the image. This ensures that the traffic sign can still maintain the absolute sharpness of the outline after defogging, preventing the missed detection of small targets caused by edge blurring.

[0102] To avoid over-enhancing the image and affecting the accuracy of subsequent detection, this embodiment adds an image consistency verification mechanism after enhancement. This mechanism evaluates the integrity of the processing effect by calculating the structural similarity between the images before and after enhancement in real time. The specific formula is as follows:

[0103] ;

[0104] in and Original image and enhanced image, respectively. and These are the pixel averages, For variance, For covariance, , As a stability constant, SSIM comprehensively considers brightness, contrast, and structural information. A value closer to 1 indicates higher image quality. If the structural similarity is below 0.75, the system will determine it as overprocessed, and then perform a weighted fusion of the processed image and the original image to improve structural similarity. The final fused image is I. final The calculation formula is:

[0105] ;

[0106] in, This is a penalty factor that varies with SSIM; when SSIM is less than 0.75, A value of 0.8 is used to mitigate over-enhancement by incorporating components from the original image. When SSIM is greater than or equal to 0.75, A value of 1 indicates that the enhanced image is used entirely.

[0107] Subsequently, the enhanced image is input into a backbone network based on the Dinov3-ConvNeXt architecture for feature extraction. This architecture combines the advantages of self-supervised representation learning and hierarchical convolutional extraction, and utilizes a self-distillation mechanism for pre-training on massive amounts of unlabeled data, enabling the model to have stronger generalization ability. In the self-distillation pre-training stage, this embodiment employs a self-supervised contrastive learning strategy based on a teacher-student dual-branch architecture. For the input unlabeled traffic scene image, multi-scale random cropping and data augmentation techniques are used to generate multi-dimensional view combinations. The teacher network only receives the global view as input to extract macroscopic semantic features, while the student network simultaneously processes all view combinations, including both global and local features. During training, the student network learns to infer the global semantic structure from local fragmented information by minimizing the centered cross-entropy loss between its output probability distribution and that of the teacher network. This allows the model to automatically capture traffic features in the image without manual annotation. To further enhance the stability of pre-trained features, the inherent hierarchical relationship between the markers and the environmental background is explored. Instead of directly participating in gradient backpropagation updates, the teacher network parameters evolve smoothly through the exponential moving average mechanism of the student network parameters. This momentum update method ensures that the teacher network can serve as a stable knowledge base, guiding the student network to extract general representations that are highly robust to illumination fluctuations, viewpoint shifts, and object occlusion. Through this self-distillation mechanism, the backbone network obtains a more generalized and universal visual representation capability. In this embodiment, the backbone network extracts four layers of multi-scale features (C2, C3, C4, C5) from the input 640×640 image, providing input features for subsequent multi-scale fusion.

[0108] Step S2: Use upsampling and layer-by-layer fusion to fuse multi-scale features, generate a feature pyramid, and input the feature pyramid into the detection model to obtain traffic sign prediction boxes and traffic sign category information.

[0109] In one embodiment, multi-scale features are fused using upsampling and layer-by-layer fusion to generate a feature pyramid. The feature pyramid is then input into a detection model to obtain traffic sign prediction boxes and traffic sign category information, including:

[0110] The number of channels of multi-scale features is unified by using target convolution, and the multi-scale features are fused layer by layer from high to low according to the stage layer of the multi-scale features. The deep features corresponding to the multi-scale features are calculated based on the fusion results.

[0111] Based on deep features and channel number, a channel-level gating mechanism is used for weighted fusion. The fused feature map is output as a feature pyramid, and the feature pyramid is input into a detection model consisting of three consecutive bounding box regression and classification sub-networks.

[0112] The detection model is used to initially generate candidate boxes and classification results corresponding to the feature pyramid, and the candidate boxes and classification results are used as inputs to optimize the bounding boxes and class confidence within the candidate box region.

[0113] The bounding box and category confidence optimization results are used as inputs, and the candidate boxes and classification results are fused for real-time correction. Based on the correction results, the traffic sign prediction box and traffic sign category information are obtained.

[0114] It should be explained that, in the feature fusion stage, this embodiment designs a channel-gated feature pyramid network (CG-FPN). The classic feature pyramid network FPN connects from top to bottom, fusing deep and shallow features to achieve multi-scale fusion. Specifically, FPN first extracts feature maps C from different levels output by the backbone network. i A multi-scale feature pyramid P is generated through upsampling and layer-by-layer fusion. i These features are then input into the detection head. However, standard FPN has certain limitations in small object detection tasks. Because the spatial resolution of deep feature maps decreases significantly after multiple samplings, their detailed information is significantly reduced. Furthermore, the simple layer-by-layer addition fusion strategy cannot distinguish between useful and useless features in different channels, leading to a further reduction in small object features during the fusion process. Based on this, this embodiment proposes CG-FPN. This network maintains the multi-scale feature fusion structure of FPN while introducing a channel-level gating mechanism (CGM) to achieve the filtering and enhancement of channel features. Given the i-th layer fusion node, the input is the upsampled deep features. and shallow features with lateral connections First, the two features are concatenated, and then global information is extracted using global average pooling (GAP).

[0115] ;

[0116] ;

[0117] Where g is a vector after global average pooling, and channel-gated weights are generated through two fully connected layers:

[0118] ;

[0119] ;

[0120] Where r is the dimensionality reduction ratio, used to reduce the number of parameters, W is the global transformation weight matrix of the fusion layer, and the two fully connected layers are used to increase the nonlinear representation capability of the model, while avoiding overfitting through dimensionality reduction.

[0121] Divide W into W deep and W shallow Then, normalization is performed using Softmax:

[0122] ;

[0123] Among them W deep and W shallow These are the weight matrices for deep features and shallow features, respectively. and Belongs to R C And satisfy Using Softmax creates competitive constraints, making the model more inclined to select channels with larger weights and maintain stable feature amplitudes. Finally, gated weights are used for weighted fusion.

[0124] ;

[0125] Where P i This represents the i-th layer feature map generated after fusion. This indicates element-wise multiplication. This can be interpreted as determining the proportion of deep features that should be used in this channel. After adding CGM to FPN, the complete and detailed process is as follows:

[0126] Features C of each stage of the backbone network output in lateral connections i For i∈{2, 3, 4, 5}, the number of channels is unified through 1×1 convolution:

[0127] ;

[0128] Integration: Starting from the highest level, integrate layer by layer downwards:

[0129] ;

[0130] For i∈{2, 3, 4, 5}:

[0131] ;

[0132] ;

[0133] The final feature pyramid is {P2, P3, P4, P5}.

[0134] After feature fusion is completed, this embodiment uses a Cascade R-CNN three-stage detection head for target recognition and localization. This detection head consists of three consecutive bounding box regression and classification subnetworks. The detection results of each stage are used as inputs for the next stage to achieve progressive optimization. The first stage is responsible for initially generating candidate boxes and approximate classifications. The second stage further optimizes the bounding boxes and class confidence in the candidate box regions. The third stage corrects the results of the first two stages to obtain the final predicted boxes and class information.

[0135] To verify the effectiveness of the channel gating mechanism, a comparison chart of the feature responses to the same input with the standard FPN method was plotted, as shown below. Figure 5 As can be clearly seen, the high response area of ​​the standard FPN is only concentrated near the center of the target, and its ability to capture the overall contour is weak. In contrast, the CG-FPN proposed in this embodiment not only has a more complete response at the same position, but can also clearly delineate the contour of the traffic sign. It also has a more reasonable suppression of the surrounding non-target areas, which further verifies the effectiveness of the channel gating mechanism in the feature fusion stage. This enables the network to pay more attention to the channels that have the ability to discriminate small targets, thereby significantly improving the detection performance.

[0136] Step S3: Determine the accuracy of the traffic sign prediction box and traffic sign category information, correct the detection model based on the accuracy results, and perform anti-tampering processing on the detection model after correction, so as to realize the detection of traffic signs in high-speed scenarios by using the monitoring model and the aging state output results of traffic signs.

[0137] In one embodiment, the accuracy of the traffic sign prediction bounding box and traffic sign category information is determined, the detection model is corrected based on the accuracy result, and after correction, the detection model is subjected to tamper-proof processing to achieve traffic sign detection in high-speed scenarios using the monitoring model and the aging state output of traffic signs.

[0138] Calculate the actual aspect ratio of the traffic sign prediction frame, and calculate the relative error between the actual aspect ratio and the standard aspect ratio. Compare the relative error with the error threshold.

[0139] If the relative error is less than or equal to the error threshold, it means that there is no anomaly in the traffic sign prediction box. If the relative error is greater than the error threshold, it means that there is anomaly in the traffic sign prediction box. The original confidence of the detection model is then corrected, and the traffic sign prediction box and traffic sign category information are re-output.

[0140] The corrected detection model is exported as a neural network exchange format. During the export process, the key layer weights of the corrected detection model are sampled and calculated to generate a unique hash fingerprint.

[0141] The hash fingerprint is embedded into the encrypted header file as a security benchmark for the operation of the corrected detection model. The corrected detection model and the aging state output of traffic signs are used to realize the detection of traffic signs in high-speed scenarios.

[0142] It should be noted that, in order to further enhance the target detection performance, this embodiment adds a post-processing verification function based on physical priors after this section to calculate the actual aspect ratio of the predicted bounding box. And calculate its aspect ratio R compared to the standard aspect ratio. std relative error Set penalty factors ,like If the value is greater than the threshold of 0.3, it indicates that the predicted bounding box is abnormal, and a penalty weight is applied in this case. It should be 0.8, otherwise The original confidence S of the Cascade R-CNN output is 1.0. origin Make corrections to obtain the final confidence level. Finally, the trained model is exported as ONNX.

[0143] During the model export and integration phase, this embodiment introduces a security protection mechanism based on dynamic verification of weight fingerprints. When exporting the model to ONNX format, the weights of key layers are sampled and calculated using the SHA-256 algorithm to generate a unique hash fingerprint. This fingerprint is then embedded in the software's encrypted header file as a runtime security benchmark. During software operation, the security monitoring module performs real-time integrity measurement, periodically extracting weight snapshots of the first 10 layers and the last 3 layers of the model in memory. By comparing the real-time weight features with the pre-stored hash fingerprint, it can accurately identify whether the model has been subjected to unauthorized pruning, parameter fine-tuning, or architectural tampering. If consistency verification fails, self-protection logic is immediately triggered, forcibly blocking the inference link and locking the core computing library. This effectively avoids performance degradation or misleading interference caused by architectural tampering, ensuring the security and reliability of the target detection output. Figure 3 and Figure 4 As shown.

[0144] According to another embodiment of the present invention, a multi-scale target detection system based on environmental perception and channel gating mechanism is also provided, the system comprising:

[0145] The feature map and aging state output module is used to segment the traffic sign image to be detected based on brightness and contrast. Based on the segmentation result, the traffic sign image to be detected is adaptively optimized to generate an enhanced traffic sign image. The enhanced traffic sign image is then processed using a backbone network and aggregation factor technology to output the multi-scale feature map and aging state of the traffic sign corresponding to the enhanced traffic sign image.

[0146] The traffic sign prediction box and category information output module is used to fuse multi-scale features using upsampling and layer-by-layer fusion methods to generate a feature pyramid, and input the feature pyramid into the detection model to obtain traffic sign prediction boxes and traffic sign category information.

[0147] The detection model correction and anti-tampering processing module is used to determine the accuracy of the traffic sign prediction box and traffic sign category information, correct the detection model based on the accuracy results, and perform anti-tampering processing on the detection model after correction, so as to realize the detection of traffic signs in high-speed scenarios by using the monitoring model and the aging state output of traffic signs.

[0148] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A multi-scale object detection method based on environmental perception and channel gating mechanism, characterized in that, The method includes: The traffic sign image to be detected is segmented based on brightness and contrast. Based on the segmentation results, the traffic sign image to be detected is adaptively optimized to generate an enhanced traffic sign image. The enhanced traffic sign image is then processed using a backbone network and aggregation factor technology to output the multi-scale feature map corresponding to the enhanced traffic sign image and the aging status of the traffic sign. Multi-scale features are fused using upsampling and layer-by-layer fusion methods to generate a feature pyramid. The feature pyramid is then input into the detection model to obtain traffic sign prediction boxes and traffic sign category information. The accuracy of the traffic sign prediction bounding box and traffic sign category information is judged, the detection model is corrected based on the accuracy results, and the detection model is exported and tamper-proofed after correction. The detection of traffic signs in high-speed scenarios is realized by using the monitoring model and the aging state output of traffic signs.

2. The multi-scale target detection method based on environmental perception and channel gating mechanism according to claim 1, characterized in that, The process involves segmenting the traffic sign image to be detected based on brightness and contrast, adaptively optimizing the segmentation results to generate an enhanced traffic sign image, and then processing the enhanced traffic sign image using a backbone network and aggregation factor techniques. The output includes a multi-scale feature map corresponding to the enhanced traffic sign image and the aging state of the traffic sign. Calculate the average brightness and variance of the grayscale image of the traffic sign to be detected, and compare the average brightness with the brightness threshold and the variance with the contrast threshold. If the average brightness is less than the brightness threshold, it indicates that the traffic sign image to be detected is in a low-light condition. In this case, the image enhancement theory and technology are used to perform low-light enhancement processing on the traffic sign image to be detected. If the variance is less than the contrast threshold and the average brightness is greater than or equal to the brightness threshold, then the traffic sign image to be detected is determined to be in a foggy situation, and the dark channel prior algorithm is used to perform defogging and noise reduction processing on the traffic sign image to be detected. The consistency verification mechanism is used to calculate the consistency between the low-light enhancement and defogging and denoising processing of the traffic sign image to be detected. The enhanced traffic sign image is output after the consistency result meets the requirements. At the same time, the aging state of the traffic sign in the enhanced traffic sign image is analyzed. A self-distillation mechanism is used to train a self-supervised learning network to generate a backbone network. Enhanced traffic sign images are then input into the backbone network, and hierarchical convolution extraction techniques are applied to the backbone network to output multi-scale feature maps.

3. The multi-scale target detection method based on environmental perception and channel gating mechanism according to claim 2, characterized in that, The low-light enhancement processing of the traffic sign image to be detected using image enhancement theory and technology includes: By convolving the traffic sign image under low light conditions with Gaussian kernels of different scales, the ambient brightness estimation components under different receptive fields are extracted, and the reflection components are calculated in the number domain based on the ambient brightness estimation components. The uneven illumination problem of the traffic sign image under the detection is eliminated based on the reflection components. The process of performing dehazing and noise reduction on the traffic sign image to be detected using the dark channel prior algorithm includes: The minimum values ​​of the three primary color channels are searched within a local window of the traffic sign image to be detected in foggy conditions. The initial transmittance is estimated by combining the atmospheric light value. The four-directional gradient operator is used as the weight constraint of the guided filter to refine the initial transmittance, so as to complete the defogging process of the traffic sign image to be detected in foggy conditions.

4. The multi-scale target detection method based on environmental perception and channel gating mechanism according to claim 3, characterized in that, The consistency verification mechanism calculates the consistency between the low-light enhancement and dehazing / denoising processed traffic sign images, and outputs the enhanced traffic sign image after the consistency result meets the requirements. Simultaneously, it analyzes the aging state of the traffic signs in the enhanced traffic sign image, including: The structural similarity of the traffic sign image to be detected before and after low light enhancement and before and after defogging and noise reduction is calculated based on pixel mean, covariance and stability coefficient, and the difference between the structural similarity and the similarity threshold is calculated. If the structural similarity is less than the similarity threshold, it indicates that the traffic sign images to be detected are inconsistent before and after low-light enhancement and before and after defogging and denoising. In this case, the traffic sign images to be detected before and after low-light enhancement and before and after defogging and denoising are weighted and fused to output an enhanced traffic sign image. Select several pixels in the enhanced traffic sign image as points of interest, and use the standard deviation of the points of interest in the enhanced traffic sign image to calculate the aggregation factor to identify the crack texture of the traffic sign. The fuzzy maximum entropy method is used to calculate the separation points between the target class and the background class in the enhanced traffic sign image, and the optimal classification surface is constructed based on the separation points to determine the degree of color attenuation in the enhanced traffic sign image. The aging status of traffic signs is determined by examining the crack texture of the traffic signs and the degree of color decay in the enhanced traffic sign image.

5. The multi-scale target detection method based on environmental perception and channel gating mechanism according to claim 4, characterized in that, The step of selecting several pixels in the enhanced traffic sign image as points of interest, and using the standard deviation of these points in the enhanced traffic sign image to calculate the aggregation factor to identify the crack texture of the traffic sign includes: The enhanced traffic sign image is subjected to grayscale difference enhancement processing. After the processing is completed, the grayscale change rate of each pixel in the enhanced traffic sign image is calculated and the grayscale change rates are sorted in descending order. Based on the sorting results, select several pixels with the largest gray-scale change rate as points of interest, establish a coordinate system based on the enhanced traffic sign image, and calculate the standard deviation of the points of interest in the horizontal and vertical directions of the coordinate system to reflect the dispersion of the points of interest. The standard deviation is normalized to define the aggregation factors in the horizontal and vertical directions. A scatter plot of the aggregation factor distribution is constructed by combining traffic sign sample images with and without cracks. The elliptical arc boundary is determined by statistical analysis based on the scatter plot. Crack warning regions are defined based on the area enclosed by the elliptical arc boundary and the coordinate system, and the crack texture of traffic signs in enhanced traffic sign images is identified based on the aggregation factor falling into the crack warning regions.

6. The multi-scale target detection method based on environmental perception and channel gating mechanism according to claim 5, characterized in that, The step of using the fuzzy maximum entropy method to calculate the separation points between the target class and the background class in the enhanced traffic sign image, and constructing the optimal classification surface based on the separation points to determine the degree of color attenuation in the enhanced traffic sign image includes: Some enhanced traffic sign images are converted into grayscale images, and the grayscale matrix of the grayscale image is obtained. The grayscale matrix is ​​then mapped into a fuzzy number matrix through a membership function to construct a fuzzy set of target class and background class. Calculate the average gray value of the fuzzy sets of the target class and the background class respectively, calculate the fuzzy entropy using the principle of maximum fuzzy entropy, select the average gray value that maximizes the fuzzy entropy, and take the corresponding average gray value as the best separation point between the target class and the background class. The maximum ratio of the three primary colors of the target class is extracted using the optimal separation point as the boundary, and the maximum ratio is used as the sample feature. Combined with the standard color benchmark value, a training sample set is constructed to build the optimal classification surface. Input the maximum ratios of the three primary colors of the remaining enhanced traffic sign images into the optimal classification surface, determine the changes in the maximum ratios of the three primary colors in the enhanced traffic sign images, and output the quantification results of the color attenuation degree.

7. The multi-scale target detection method based on environmental perception and channel gating mechanism according to claim 6, characterized in that, The self-distillation mechanism is a self-supervised contrastive learning strategy based on a dual-branch architecture of teacher network and student network; The teacher network is used to extract macroscopic semantic features, and the student network is used to process all view combinations, including global and local ones, at the same time. The student network learns to infer the global semantic structure from local fragment information by minimizing the centralized cross-entropy loss between its output probability distribution and that of the teacher network.

8. The multi-scale target detection method based on environmental perception and channel gating mechanism according to claim 1, characterized in that, The process of fusing multi-scale features using upsampling and layer-by-layer fusion to generate a feature pyramid, and then inputting the feature pyramid into the detection model to obtain traffic sign prediction boxes and traffic sign category information, includes: The number of channels of multi-scale features is unified by using target convolution, and the multi-scale features are fused layer by layer from high to low according to the stage layer of the multi-scale features. The deep features corresponding to the multi-scale features are calculated based on the fusion results. Based on deep features and channel number, a channel-level gating mechanism is used for weighted fusion. The fused feature map is output as a feature pyramid, and the feature pyramid is input into a detection model consisting of three consecutive bounding box regression and classification sub-networks. The detection model is used to initially generate candidate boxes and classification results corresponding to the feature pyramid, and the candidate boxes and classification results are used as inputs to optimize the bounding boxes and class confidence within the candidate box region. The bounding box and category confidence optimization results are used as inputs, and the candidate boxes and classification results are fused for real-time correction. Based on the correction results, the traffic sign prediction box and traffic sign category information are obtained.

9. The multi-scale target detection method based on environmental perception and channel gating mechanism according to claim 1, characterized in that, The process of determining the accuracy of the traffic sign prediction bounding box and traffic sign category information, correcting the detection model based on the accuracy results, and performing tamper-proof processing on the detection model after correction to achieve traffic sign detection in high-speed scenarios using the monitoring model and the aging state output of traffic signs includes: Calculate the actual aspect ratio of the traffic sign prediction frame, and calculate the relative error between the actual aspect ratio and the standard aspect ratio. Compare the relative error with the error threshold. If the relative error is less than or equal to the error threshold, it means that there is no anomaly in the traffic sign prediction box. If the relative error is greater than the error threshold, it means that there is anomaly in the traffic sign prediction box. The original confidence of the detection model is then corrected, and the traffic sign prediction box and traffic sign category information are re-output. The corrected detection model is exported as a neural network exchange format. During the export process, the key layer weights of the corrected detection model are sampled and calculated to generate a unique hash fingerprint. The hash fingerprint is embedded into the encrypted header file as a security benchmark for the operation of the corrected detection model. The corrected detection model and the aging state output of traffic signs are used to realize the detection of traffic signs in high-speed scenarios.

10. A multi-scale target detection system based on environmental perception and channel gating mechanism, used to implement the multi-scale target detection method based on environmental perception and channel gating mechanism as described in any one of claims 1-9, characterized in that, The system includes: The feature map and aging state output module is used to segment the traffic sign image to be detected based on brightness and contrast. Based on the segmentation result, the traffic sign image to be detected is adaptively optimized to generate an enhanced traffic sign image. The enhanced traffic sign image is then processed using a backbone network and aggregation factor technology to output the multi-scale feature map and aging state of the traffic sign corresponding to the enhanced traffic sign image. The traffic sign prediction box and category information output module is used to fuse multi-scale features using upsampling and layer-by-layer fusion methods to generate a feature pyramid, and input the feature pyramid into the detection model to obtain traffic sign prediction boxes and traffic sign category information. The detection model correction and anti-tampering processing module is used to determine the accuracy of the traffic sign prediction box and traffic sign category information, correct the detection model based on the accuracy results, and perform anti-tampering processing on the detection model after correction, so as to realize the detection of traffic signs in high-speed scenarios by using the monitoring model and the aging state output of traffic signs.

Images