Intelligent hazardous waste sorting method based on machine vision

Abstract

The invention relates to the technical field of image processing, and discloses a hazardous waste intelligent sorting method based on machine vision, which comprises the following steps: acquiring a hazardous waste image to be sorted and extracting a multi-scale feature map; calculating an asymmetric weight factor, generating a gradient adjustment coefficient, and weighting the four distances by using the asymmetric weight factor and the gradient adjustment coefficient to obtain a centrality score of the pixel point; the centrality score and the classification confidence coefficient are fused to obtain a comprehensive sorting index; determining the number K of high-resolution pixel points, and selecting K high-resolution pixel points by adopting an iterative screening strategy according to the comprehensive sorting index; performing category classification, bounding box regression and contour frequency domain regression based on the features of the selected high-resolution pixel point set; and jointly outputting the types, bounding box positions and refined contours of the dangerous waste products. According to the method, the regression precision of the bounding box is improved, and the refined contour extraction can be realized in a lightweight manner, so that the problem that the calculation amount is large and real-time application is difficult is solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing, in particular to a dangerous waste intelligent sorting method based on machine vision. BACKGROUND

[0002] Automatic and intelligent sorting is a link for improving processing efficiency and reducing manual operation risk in the field of hazardous waste (referred to as hazardous waste) treatment. The hazardous waste target detection technology based on computer vision provides an effective solution. The hazardous waste target detection technology based on computer vision can automatically identify the category and position of hazardous waste products on the conveying belt.

[0003] The hazardous waste target detection algorithm based on computer vision in the prior art is mainly divided into two stages and a single stage. The two-stage algorithm such as FasterR-CNN algorithm has the core idea that the first stage generates candidate regions, and the second stage classifies and regresses the boundary box of the candidate regions. In the first stage, a region candidate network is used to quickly generate high-quality candidate regions. RPN slides on the feature map, predicts which of the multiple anchor boxes is a foreground target for each position, and performs preliminary regression on the boundary box of the foreground target. In the second stage, the candidate regions generated by the RPN are sent to the RoI Pooling layer, and the candidate region features of different sizes are cropped and aligned to a fixed size. Then, these fixed-size feature maps are sent to the subsequent fully connected layer for final target classification and accurate boundary box regression. The detection accuracy of the FasterR-CNN algorithm is high, but the speed is slow, which is difficult to meet the real-time requirements of the flow line.

[0004] The single-stage algorithm such as YOLO, SSD, and FCOS has excellent detection speed in industrial scenarios and is suitable for industrial real-time applications. The FCOS algorithm is a single-stage and anchor-free target detection algorithm. It discards the traditional anchor mechanism and directly predicts the distance from each pixel point on the feature map to the four edges of the target boundary box. However, the FCOS algorithm has the problem of low boundary box detection accuracy, which leads to low accuracy of the detection result. SUMMARY

[0005] The present application provides a hazardous waste intelligent sorting method based on machine vision to solve the problems of slow detection speed in the prior art, which is difficult to meet the real-time requirements of the flow line, and low boundary box detection accuracy, which leads to low accuracy of the detection result.

[0006] The hazardous waste intelligent sorting method based on machine vision of the present application comprises the following steps: Obtain an image of hazardous waste to be sorted, and use the backbone network and feature pyramid network to extract a multi-scale feature map. For any pixel on the multi-scale feature map, calculate an asymmetric weight factor based on the predicted distance of the pixel to the four sides of the bounding box, combine the gradient information of the local neighborhood of the pixel to generate a gradient adjustment coefficient, and use the asymmetric weight factor and the gradient adjustment coefficient to weight the four distances to obtain the center score of the pixel. The centrality score and the classification confidence are combined to obtain a comprehensive ranking index; the number K of high-scoring pixels is determined based on prior information about the size and category of the hazardous waste, and an iterative screening strategy is used to select K spatially distributed high-scoring pixels based on the comprehensive ranking index; the iterative screening strategy is as follows: after each highest-scoring pixel is selected, Gaussian decay is applied to the comprehensive ranking indexes of other candidate pixels in the spatial neighborhood of the pixel; Based on the features of the selected high-resolution pixel point set, category classification, bounding box regression and contour frequency domain regression are performed; the contour frequency domain regression branch regresses the discrete Fourier transform coefficients of the hazardous waste contour from the features and reconstructs the refined contour through inverse transformation; the category, bounding box position and refined contour of the hazardous waste are jointly output.

[0007] Preferably, the asymmetric weight factor and the gradient adjustment coefficient are used to weight the four distances to obtain the center score of the pixel point, including: calculating the distance of the pixel point to the left, right, top and bottom four sides of the bounding box ; Calculate the horizontal distance ratio and vertical distance ratio ; The feature map of the pixel point The gradient amplitude is calculated by applying the Sobel operator in the neighborhood; the center score of the pixel point is calculated by the formula Calculated, where C is the centrality score, is the horizontal distance ratio, is the vertical distance ratio, and G is the gradient amplitude.

[0008] Preferably, the fusion of the centrality score and the classification confidence to obtain a comprehensive ranking index includes: obtaining the centrality score and the classification confidence corresponding to the pixel point; and using the geometric mean of the centrality score and the classification confidence as the comprehensive ranking index of the pixel point.

[0009] Preferably, the method of determining the number K of high-resolution pixels based on the size and category prior information of the hazardous waste includes: for each detected hazardous waste, obtaining the predicted category and predicted bounding box area of ​​the hazardous waste; according to the predicted category, searching for the corresponding basic number and size coefficient from the preset category configuration; the number of high-resolution pixels , where D represents the predicted category and A represents the predicted bounding box area. Represents the basic number corresponding to the predicted category, Represents the size coefficient corresponding to the predicted category, rounds the K value, and limits the value of K to the preset range corresponding to the predicted category Inside.

[0010] Preferably, the iterative screening strategy is used to select K spatially distributed high-score pixel points, including: arranging all candidate pixel points in descending order according to the comprehensive ranking index; iterating K times, selecting the pixel point with the highest current ranking in each iteration and adding the pixel point to the high-score pixel point set, and applying Gaussian attenuation to the comprehensive ranking index of all remaining candidate pixel points.

[0011] Preferably, the contour frequency domain regression branch regresses the discrete Fourier transform coefficients of the hazardous waste contour from the features, including: fusing the features of K high-resolution pixels into a fused feature vector through an average pooling operation; inputting the fused feature vector into a regression head network, and the regression head network outputs a 128-dimensional vector, which represents the real and imaginary parts of the 64 Fourier coefficients of the hazardous waste contour.

[0012] Preferably, the reconstructing the refined contour by inverse transformation includes: reorganizing the 128-dimensional vector into 64 complex Fourier coefficients; applying an inverse discrete Fourier transform to calculate the coordinates of 64 sampling points on the contour from the 64 complex Fourier coefficients; and connecting the 64 sampling point coordinates in sequence to form a closed refined contour.

[0013] Preferably, the backbone network is a ResNet-50 network.

[0014] Preferably, the Gaussian attenuation is calculated based on the Euclidean distance between the remaining candidate pixels and the newly selected high-resolution pixel.

[0015] Preferably, the standard deviation of the Gaussian function of the Gaussian decay Proportional to the size of the hazardous waste prediction bounding box.

[0016] The beneficial effects of the present application are: the present application combines the gradient adjustment coefficient of the pixel point and the asymmetric weight factor to calculate the center degree score, so that the center degree score can better reflect the positioning quality of the prediction box, and the regression accuracy of the bounding box is significantly improved for irregularly shaped hazardous waste. The center degree score and the classification confidence are fused into a comprehensive ranking index, and the high comprehensive ranking index pixel points are selected as high score pixel points according to the comprehensive ranking index using an iterative screening strategy, so that a high score pixel point set with reasonable spatial distribution and low information redundancy can be obtained, and the training efficiency and detection performance of the network are improved. In addition, the contour frequency domain regression branch in the present application predicts the Fourier coefficient of the contour and inversely transforms and reconstructs, so as to realize the extraction of the fine contour in a lightweight manner, thereby solving the problem of large calculation amount and difficult real-time application. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 A flowchart of a hazardous waste intelligent sorting method based on machine vision provided by an embodiment of the present application is shown. DETAILED DESCRIPTION

[0018] The embodiments of the present application are described in detail below, and examples of the embodiments are shown in the accompanying drawings. The embodiments described below by referring to the drawings are exemplary and are intended to explain the present application, and cannot be understood as a limitation of the present application.

[0019] As shown in Figure 1 The hazardous waste intelligent sorting method based on machine vision provided by the embodiment of the present application specifically includes the following steps: S1, obtaining a hazardous waste image to be sorted, and extracting multi-scale feature maps by using a backbone network and a feature pyramid network.

[0020] Specifically, a hazardous waste image on a conveying belt is collected, and the hazardous waste image is input into a detection model with a ResNet-50 network as a backbone; after the ResNet-50 network outputs feature maps of different levels, a top-down and horizontal connection operation is performed by a feature pyramid network to fuse strong semantic information of a high layer and rich spatial details of a bottom layer, and five multi-scale feature maps of P3 to P7 are generated, which are respectively used for detecting hazardous waste of different sizes.

[0021] S2, for any pixel point on the multi-scale feature map, an asymmetric weight factor is calculated based on the distance of the pixel point prediction to the four sides of the bounding box, a gradient adjustment coefficient is generated by combining the gradient information of the local neighborhood of the pixel point, and the four distances are weighted by using the asymmetric weight factor and the gradient adjustment coefficient to obtain the center degree score of the pixel point.

[0022] S3, the centrality score is integrated with the classification confidence to obtain a comprehensive ranking index; the number K of high-scoring pixels is determined based on the size and category prior information of the hazardous waste, and K spatially distributed high-scoring pixels are selected using an iterative screening strategy based on the comprehensive ranking index; the iterative screening strategy is: after each selection of a highest-scoring pixel, Gaussian attenuation is applied to the comprehensive ranking indexes of other candidate pixels in the spatial neighborhood of the pixel.

[0023] Specifically, in each iteration, the pixel with the highest comprehensive ranking index is selected as the first high-scoring pixel. With this pixel as the center, the comprehensive ranking indexes of other candidate pixels within a certain spatial radius are multiplied by a Gaussian function value. The Gaussian function value increases with distance, thereby reducing the probability of the neighboring points of the high-scoring pixel being selected in the next iteration. Repeat this process. times, select high-resolution pixels that are evenly distributed in space.

[0024] S4, based on the features of the selected high-resolution pixel point set, performs category classification, bounding box regression, and contour frequency domain regression; the contour frequency domain regression branch regresses the discrete Fourier transform coefficients of the hazardous waste contour from the features, and reconstructs the refined contour through inverse transformation; the category, bounding box position, and refined contour of the hazardous waste are jointly output.

[0025] Specifically, the selected The features of each high-resolution pixel are input into three parallel prediction heads, each of which consists of four The system consists of three parallel prediction heads: a category classification head, a bounding box regression head, and a contour frequency domain regression head. The category classification head is used to output the category of hazardous waste, such as waste batteries and waste light bulbs. The bounding box regression head is used to output the center point coordinates, width, and height of the target box. The contour frequency domain regression head is used to predict the real and imaginary parts of the first M low-frequency coefficients of the discrete Fourier transform (DFT) of the target contour. During inference, the predicted M Fourier coefficients are used to efficiently reconstruct the coordinates of the N contour vertices of the target through the inverse fast Fourier transform (IDFT) algorithm, thereby achieving refined contour extraction.

[0026] In an optional embodiment, the asymmetric weight factor and the gradient adjustment coefficient are used to weight the four distances to obtain the center score of the pixel point, including: calculating the distance of the pixel point to the left, right, top and bottom four sides of the bounding box ; Calculate the horizontal distance ratio and vertical distance ratio ; The horizontal distance ratio and the vertical distance ratio constitute the asymmetric weight factor, which is used to determine the feature map of the pixel point. The gradient amplitude G is calculated by applying the Sobel operator in the neighborhood, and the gradient adjustment coefficient is calculated by adding 1 to the gradient amplitude G; the center score of the pixel point is calculated by the formula Calculated, where C is the centrality score, is the horizontal distance ratio, is the vertical distance ratio, and G is the gradient amplitude.

[0027] Assume that for a bounding box predicted to be a battery, the size of the bounding box is Pixels, the distance from a candidate pixel inside to the left boundary 40 pixels, the distance to the right border 160 pixels, the distance to the top edge 30 pixels, the distance to the bottom border is 70 pixels. Then, the horizontal distance is =0.25, vertical distance ratio =0.43. These two ratios measure the degree of pixel centering, and the closer to 1, the more centered.

[0028] In order to determine whether a pixel point is in the area of ​​hazardous waste contour, the Sobel operator is used to calculate the pixel point on the feature map. Gradient magnitude within the neighborhood For example, if the pixel is located in the center of the hood of the vehicle, the gradient amplitude is Very low, such as gradient amplitude On the contrary, if the pixel point happens to fall on the edge of the window or the position of the headlight, the gradient amplitude is will be high, such as the gradient amplitude Substituting all factors into the formula to calculate the centrality score, we get a centrality score of approximately 0.36 for the points on the hood and 0.62 for the points on the headlights.

[0029] In an optional embodiment, the fusion of the centrality score and the classification confidence to obtain a comprehensive ranking index includes: obtaining the centrality score and the classification confidence corresponding to the pixel point; using the geometric mean of the centrality score and the classification confidence as the comprehensive ranking index of the pixel point, that is, .

[0030] In the target detection task, a pixel has two indicators: the center score C and the classification confidence S. If only one of the indicators is used as the basis for judgment, the result may not be the best choice. For example, a center score For example, a pixel point with a classification confidence S of 0.95 is not a good representative pixel point because it is located in the blurred background, and the classification confidence S of the pixel point is only 0.2. For another example, a pixel point with a classification confidence S of 0.99 may be located at the edge of the target, and the center degree score C of the pixel point is only 0.1, so the pixel point is not suitable as a representative pixel point.

[0031] To balance the center degree score C and the classification confidence S, the embodiment fuses the center degree score C and the classification confidence S in a geometric mean manner. When the center degree score C and the classification confidence S of a pixel point are both high, the pixel point can obtain a higher comprehensive ranking index. Compared with an arithmetic mean, a geometric mean is more sensitive to a lower score, can effectively suppress the case of one high and one low, and screen out better candidate pixel points.

[0032] In an optional embodiment, the number K of high-score pixel points is determined according to prior information of the size and category of the hazardous waste, including: for each detected hazardous waste, obtaining a predicted category and a predicted bounding box area of the hazardous waste; according to the predicted category, searching for a corresponding basic number and a size coefficient from a preset category configuration; the number K of high-score pixel points is determined according to the following formula: wherein D represents the predicted category, A represents the predicted bounding box area, represents the basic number corresponding to the predicted category, represents the size coefficient corresponding to the predicted category, K is rounded, and the value of K is limited in a preset range corresponding to the predicted category .

[0033] The number of high-score pixel points required when describing different targets is different, so a fixed value cannot adapt to all targets. For example, describing a waste oil drum with a complex structure and a huge size requires more high-score pixel points than describing a battery with a simple structure and a small size. The embodiment establishes a category-related configuration table. For example, for a target with a category of waste oil drum, the basic number is 20, the corresponding size coefficient is 0.15, and the preset range is 15 to 50. For a target with a category of battery, the basic number is 10, the size coefficient is 0.1, and the preset range is 10 to 20.

[0034] For example, when a target is detected, it determines the target category and bounding box area. Suppose a waste oil barrel is detected, and the bounding box area A of the waste oil barrel is 2500 square pixels. According to the configuration table, the basic number and size coefficient corresponding to the waste oil barrel are queried, and then the bounding box area, basic number and size coefficient are substituted into the calculation formula of the number of high-resolution pixels K. The value of K is equal to 27.5, which is rounded to 28. The value is within the preset range of 15 to 50, so 28 high-resolution pixels are determined to be selected.

[0035] In an optional embodiment, the iterative screening strategy is used to select K spatially distributed high-scoring pixels, including: arranging all candidate pixels in descending order according to the comprehensive ranking index; iterating K times, selecting the pixel with the highest current ranking in each iteration and adding the pixel to the high-scoring pixel set, and applying Gaussian decay to the comprehensive ranking index of all remaining candidate pixels; the Gaussian decay is calculated based on the Euclidean distance between the remaining candidate pixels and the newly selected pixel, and the standard deviation of the Gaussian function is It is proportional to the size of the hazardous waste prediction bounding box.

[0036] In order to avoid all the K high-scoring pixels being gathered in the same small area of ​​the target, and thus obtain a set of high-scoring pixels that can represent the entire target shape and are evenly distributed in space, this embodiment adopts an iterative suppression screening process. All candidate pixels within the target bounding box are sorted from high to low according to their comprehensive ranking index. In the first iteration, the top-ranked pixel point P1 is unconditionally selected into the high-scoring pixel point set. After pixel point P1 is selected, in order to prevent the next selected pixel point from being too close to P1, the comprehensive ranking indexes of all remaining candidate pixel points are updated. For any candidate point P i , calculate the Euclidean distance between it and P1 on the image .P i The comprehensive ranking index after Gaussian decay is given by P i The initial comprehensive ranking index is calculated by multiplying it by a Gaussian attenuation factor. The size of the standard deviation σ of the Gaussian function is proportional to the size of the target bounding box. For example, for a pixel point 5 pixels away from P1, the comprehensive ranking index may be attenuated to 30% of the original, while a pixel point 50 pixels away from P1 may only attenuate to 95% of the original. After updating all comprehensive ranking indices, the pixel point P2 with the highest current comprehensive ranking index is selected in the second iteration. P2 must be a pixel point with a higher comprehensive ranking index and a certain distance from P1. This process is repeated Each time, the current optimal pixel is selected and the neighboring points of the optimal pixel are suppressed.

[0037] In an optional embodiment, the contour frequency domain regression branch regresses discrete Fourier transform coefficients of the hazardous waste contour from the features, including: fusing the features of the K high-scored pixels into a fused feature vector through average pooling operation; inputting the fused feature vector into a regression head network, the regression head network outputting a 128-dimensional vector, the vector representing real parts and imaginary parts of 64 Fourier coefficients of the hazardous waste contour.

[0038] After selecting K representative high-scored pixels through the iterative screening strategy, the visual information contained in these high-scored pixels is used to describe the accurate contour of the target. Each high-scored pixel has a corresponding feature vector in the feature map of the deep neural network, for example, a 256-dimensional vector, which encodes the local image pattern around the high-scored pixel. In this embodiment, the 256-dimensional feature vectors of the K high-scored pixels are average-pooled, and the first dimension of all K vectors is added and divided by K, the second dimension is added and divided by K, and so on, to generate a 256-dimensional fused feature vector.

[0039] The fused feature vector is sent to the regression head network, which is composed of several fully connected layers and can map the input 256-dimensional fused feature vector to a 128-dimensional output vector. The 128-dimensional vector is a feature representation of the target contour in the frequency domain. Specifically, the first 64 elements of the 128-dimensional vector are interpreted as the real parts of 64 Fourier coefficients describing the contour shape, and the last 64 elements are interpreted as the imaginary parts of these coefficients.

[0040] In an optional embodiment, the reconstruction of the refined contour through inverse transformation includes: reorganizing the 128-dimensional vector into 64 complex Fourier coefficients; applying inverse discrete Fourier transform to calculate 64 sample point coordinates on the contour from the 64 complex Fourier coefficients; and connecting the 64 sample point coordinates in order to form a closed refined contour.

[0041] Specifically, the 0th element and the 64th element of the vector are combined into a first complex Fourier coefficient, where the former is the real part and the latter is the imaginary part. Similarly, the 1st element and the 65th element form a second complex Fourier coefficient, and so on until all 64 complex Fourier coefficients are generated. This series of complex Fourier coefficients completely describes the contour shape of the target in the frequency domain. Applying an inverse discrete Fourier transform to the 64 complex Fourier coefficients can convert the frequency domain signal back to the time domain or spatial domain signal. In this embodiment, the output of the inverse discrete Fourier transform is a set of 64 complex numbers, and the real and imaginary parts of these complex numbers are interpreted as the x and y coordinates of the 64 sampling points on the contour, respectively. For example, the inverse discrete Fourier transform may produce the sequence (125.3+89.1i), (128.0+92.4i),..., which corresponds to the contour points (125.3, 89.1), (128.0, 92.4), etc. Connecting the 64 coordinate points in order according to their sequence in the sequence with straight lines, and connecting the last point with the first point, forms a closed polygon. This polygon is the refined contour of the target, which can more accurately depict the true shape of the target than the bounding box.

[0042] The implementation principle of the hazardous waste intelligent sorting method based on machine vision in the embodiment of the present application is that: the center degree score is calculated by combining the gradient adjustment coefficient of the pixel point and the asymmetric weight factor, so that the center degree score can better reflect the positioning quality of the prediction box, and the regression accuracy of the bounding box is significantly improved. In addition, the center degree score and the classification confidence are fused into a comprehensive ranking index, and the high comprehensive ranking index pixel points are selected as high score pixel points according to the comprehensive ranking index by using an iterative screening strategy, so that a high score pixel point set with reasonable spatial distribution and low information redundancy can be obtained, thereby improving the training efficiency and detection performance of the network. Moreover, the contour frequency domain regression branch in the present application realizes the extraction of the refined contour in a lightweight manner by predicting the Fourier coefficients of the contour and inverse transforming and reconstructing, thereby solving the problem of large calculation amount and difficulty in real-time application.

[0043] Although the embodiments of the present application have been shown and described above, it should be understood that the above embodiments are exemplary and should not be construed as limiting the present application, and those skilled in the art can make changes, modifications, replacements and variations to the above embodiments within the scope of the present application.

Claims

1. The intelligent sorting method for hazardous waste based on machine vision is characterized by: The steps include: Obtain an image of hazardous waste to be sorted, and use the backbone network and feature pyramid network to extract a multi-scale feature map. For any pixel on the multi-scale feature map, calculate an asymmetric weight factor based on the predicted distance of the pixel to the four sides of the bounding box, combine the gradient information of the local neighborhood of the pixel to generate a gradient adjustment coefficient, and use the asymmetric weight factor and the gradient adjustment coefficient to weight the four distances to obtain the center score of the pixel. The centrality score and the classification confidence are combined to obtain a comprehensive ranking index; the number K of high-scoring pixels is determined based on prior information about the size and category of the hazardous waste, and an iterative screening strategy is used to select K spatially distributed high-scoring pixels based on the comprehensive ranking index; the iterative screening strategy is as follows: after each highest-scoring pixel is selected, Gaussian decay is applied to the comprehensive ranking indexes of other candidate pixels in the spatial neighborhood of the pixel; Based on the features of the selected high-resolution pixel point set, category classification, bounding box regression and contour frequency domain regression are performed; the contour frequency domain regression branch regresses the discrete Fourier transform coefficients of the hazardous waste contour from the features and reconstructs the refined contour through inverse transformation; the category, bounding box position and refined contour of the hazardous waste are jointly output.

2. The method for intelligent sorting of hazardous waste based on machine vision according to claim 1 is characterized in that: The asymmetric weight factor and the gradient adjustment coefficient are used to weight the four distances to obtain the center score of the pixel point, including: Calculate the distance of the pixel prediction to the left, right, top, and bottom edges of the bounding box ; Calculate the horizontal distance ratio and vertical distance ratio ; The feature map of the pixel point The gradient amplitude is calculated by applying the Sobel operator in the neighborhood; The center score of a pixel is given by the formula Calculated, where C is the centrality score, is the horizontal distance ratio, is the vertical distance ratio, and G is the gradient amplitude.

3. The method for intelligent sorting of hazardous waste based on machine vision according to claim 1 is characterized in that: The fusion of the centrality score and the classification confidence to obtain a comprehensive ranking index includes: Obtaining the center score and classification confidence corresponding to the pixel point; The geometric mean of the centrality score and the classification confidence is used as the comprehensive ranking index of the pixels.

4. The method for intelligent sorting of hazardous waste based on machine vision according to claim 1 is characterized in that: Determining the number K of high-resolution pixels based on the prior information of the size and category of the hazardous waste includes: For each detected hazardous waste, obtain the predicted category and predicted bounding box area of ​​the hazardous waste; According to the predicted category, find the corresponding basic quantity and size coefficient from the preset category configuration; The number of high-resolution pixels , where D represents the predicted category and A represents the predicted bounding box area. Represents the basic number corresponding to the predicted category, Represents the size coefficient corresponding to the predicted category, rounds the K value, and limits the value of K to the preset range corresponding to the predicted category Inside.

5. The method for intelligent sorting of hazardous waste based on machine vision according to claim 1 is characterized in that: The iterative screening strategy is used to select K spatially distributed high-resolution pixels, including: Arrange all candidate pixels in descending order according to the comprehensive ranking index; Iterate K times. In each iteration, the pixel with the highest current ranking is selected and added to the high-score pixel set. Gaussian attenuation is applied to the comprehensive ranking index of all other candidate pixels.

6. The method for intelligent sorting of hazardous waste based on machine vision according to claim 1 is characterized in that: The profile frequency domain regression branch regresses the discrete Fourier transform coefficients of the hazardous waste profile from the features, including: The features of K high-resolution pixels are fused into a fused feature vector through average pooling operation; The fused feature vector is input into a regression head network, and the regression head network outputs a 128-dimensional vector, which represents the real and imaginary parts of 64 Fourier coefficients of the hazardous waste profile.

7. The method for intelligent sorting of hazardous waste based on machine vision according to claim 6 is characterized in that: The reconstructing the refined contour by inverse transformation includes: Reorganize the 128-dimensional vector into 64 complex Fourier coefficients; Applying an inverse discrete Fourier transform to calculate the coordinates of 64 sampling points on the contour from the 64 complex Fourier coefficients; The coordinates of the 64 sampling points are connected in sequence to form a closed refined contour.

8. The method for intelligent sorting of hazardous waste based on machine vision according to claim 1, characterized in that: The backbone network is the ResNet-50 network.

9. The method for intelligent sorting of hazardous waste based on machine vision according to claim 5, characterized in that: The Gaussian attenuation is calculated based on the Euclidean distance between the remaining candidate pixels and the newly selected high-score pixel.

10. The method for intelligent sorting of hazardous waste based on machine vision according to claim 9, characterized in that: The standard deviation of the Gaussian function of the Gaussian decay Proportional to the size of the hazardous waste prediction bounding box.

Images