Longitudinal tear detection method for belt-type conveyor belt

Through the longitudinal tear detection method of belt conveyor belt, deep learning training of SVM classifier and yolov5 model, the problem of laser detection being susceptible to the environment is solved, and more accurate longitudinal tear detection is achieved, reducing false alarms and leaks Rate of return.

WO2024229990A1PCT designated stage expired Publication Date: 2024-11-14WUXI BOTON IOT TECH CO LTD

Patent Information

Application Number
PCT/CN2023/110920
Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
Priority Date
2023-05-09
Filing Date
2023-08-03
Publication Date
2024-11-14

AI Technical Summary

Technical Problem

In the prior art, longitudinal tear detection of belt conveyor belts through laser detection is easily affected by the environment, resulting in false alarms and missed alarms, and inaccurate detection.

Method used

A longitudinal tear detection method of belt conveyor belt is adopted, including image acquisition, screening, image preprocessing, manual calibration, yolov5 model training and detection result confirmation, abnormal images are screened through SVM classifier, and deep learning training is used for use of yolov5 model. Conveyor belt tear detection is carried out based on the trained model to reduce environmental impact and reduce false alarms and missed alarm rates.

Benefits of technology

By acquiring abnormal conveyor belt images in advance and deep learning training based on the yolov5 model, the environmental impact can be effectively reduced, false alarms and missed alarms can be reduced, and the accuracy of longitudinal tear detection can be improved.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN2023110920_14112024_PF_FP_ABST
    Figure CN2023110920_14112024_PF_FP_ABST
Patent Text Reader

Abstract

The present invention relates to the technical field of conveyor belt detection. Disclosed is a longitudinal tear detection method for a belt-type conveyor belt. According to the present invention, anomalous conveyor belt images are acquired in advance, defect calibration is carried out on the anomalous conveyor belt images to obtain calibrated images, deep learning training of a yolov5 model is carried out on the basis of a defect type on the calibrated images, and conveyor belt tear detection is carried out on the basis of the trained yolov5 model, so that the influence of an environment can be reduced, and the false alarm rate and the missing alarm rate are reduced.
Need to check novelty before this filing date? Find Prior Art

Description

A method for detecting longitudinal tearing of a conveyor belt Technical Field

[0001] The present invention relates to the technical field of conveyor belt detection, and in particular to a method for detecting longitudinal tearing of a belt conveyor belt. Background Art

[0002] Wire rope conveyor belts are often used in industries such as coal mining, mining, ports, power generation, metallurgy, and building materials, where tear resistance is critical. To improve the belt's resistance to transverse tears, steel ropes are embedded longitudinally within the belt's rubber. However, this approach doesn't address the belt's weak resistance to longitudinal tears. In actual use, conveyor belts are easily punctured by sharp objects. Combined with high-speed operation, this can cause long-distance longitudinal tears, resulting in significant economic losses. Therefore, longitudinal tear testing of conveyor belts is essential. Technical issues

[0003] Currently, most conveyor belts are inspected longitudinally using laser detection. This involves capturing conveyor belt images using a laser, then processing the images based on image algorithms to obtain target information from the images. This information is then processed to determine whether the conveyor belt has experienced a longitudinal tear. For example, patent publication CN105692122B discloses a method for detecting longitudinal tears in conveyor belts based on laser lines. This method uses a line connection to detect the center line of the laser area in the conveyor belt laser line image, and then uses multiple frames for joint judgment to identify longitudinal tears in the conveyor belt. However, laser detection is prone to false positives and omissions because laser fluctuations are significantly affected by the environment, leading to inaccurate longitudinal tear detection. Technical Solutions

[0004] In view of the shortcomings of the background technology, the present invention provides a method for detecting longitudinal tears of a belt conveyor belt. The technical problem to be solved is that the longitudinal tear detection of the existing conveyor belt is realized by laser, which is easily affected by the environment and may cause false alarms or missed alarms, resulting in inaccurate longitudinal tear detection.

[0005] To solve the above technical problems, the present invention provides the following technical solution: a method for detecting longitudinal tearing of a belt conveyor belt, comprising the following steps:

[0006] S1: Collect conveyor belt images;

[0007] S2: screening the conveyor belt images to filter out abnormal conveyor belt images in the conveyor belt images;

[0008] S3: performing image filtering, image enhancement, and image normalization processing on the abnormal conveyor belt image in sequence to obtain a preprocessed image, wherein the image normalization processing is used to scale the abnormal conveyor belt image to the same size;

[0009] S4: manually calibrating the preprocessed image to identify defects in the preprocessed image and generate a calibrated image;

[0010] S5: Divide all calibrated images into a training set, a validation set, and a test set; train the Yolov5 model based on the training set; and use the validation set to confirm the convergence effect of the Yolov5 model.

[0011] S6: After the training of the yolov5 model is completed, the test set is tested using the trained yolov5 model. If the test success rate of the test set is higher than the first judgment threshold, the trained yolov5 model is used as the final yolov5 model. Otherwise, the trained yolov5 model is used as the current pre-training model, and then return to step S5 to re-train the current pre-training model until the final yolov5 model is obtained.

[0012] S7: Detect the actual image of the conveyor belt based on the final YOLOv5 model and output the detection results. Each detection result includes detection category information, coordinate information, and confidence information. The process of outputting the detection results is as follows:

[0013] First, construct p detection result sets Hi, where p is the number of detection categories set by the final yolov5 model, i is between 0 and p-1, and one detection result set Hi corresponds to one detection category. The detection results are placed in the corresponding detection result set Hi according to the detection category.

[0014] Then, the detection results in the detection result set Hi are sorted in the order of confidence, with the sorting direction being from large to small;

[0015] Finally, the intersection-over-union ratio between each two detection results in each detection result set is calculated based on the coordinate information of the detection results. If the intersection-over-union ratio is higher than a second determination threshold, the two detection results are considered to overlap with each other, and the detection result with the lower confidence level is removed from the detection result set. Otherwise, both detection results are retained.

[0016] S8: Confirming the final detection result in step S7, wherein the confirmation is achieved by at least one of confidence screening, detection category screening, and target area screening;

[0017] The confidence screening process is as follows: setting a confidence threshold for a single detection category and removing detection results with a confidence less than the confidence threshold;

[0018] The process of the detection category screening is as follows: setting the target detection category and removing the detection results whose detection category is not in the target detection category;

[0019] The target area screening process is as follows: setting the target area and screening out the detection results that are not in the target area.

[0020] In one embodiment, the method for screening out abnormal conveyor belt images in step S2 is as follows:

[0021] Two labels, normal and abnormal, are set in the SVM classifier, and then the abnormal conveyor belt images in the conveyor belt images are screened out using the SVM classifier.

[0022] In a certain embodiment, M SVM classifiers are used to detect each conveyor belt image from M feature dimensions respectively. When the detection result of at least one SVM classifier is abnormal, the conveyor belt image is regarded as an abnormal conveyor belt image.

[0023] In some embodiments, in step S3, the image filtering includes first performing a median filter on the abnormal conveyor belt image to obtain a first filtered image, and then performing a Gaussian filter on the first filtered image to obtain a second filtered image;

[0024] The steps of median filtering are as follows:

[0025] First, use a g×g mask to traverse each pixel in the abnormal conveyor belt image in sequence, where g is an odd number greater than 1. The pixel and the (g×g-1) pixels around the pixel form a g×g pixel matrix, and the pixel is the center of the g×g pixel matrix. The pixels in the g×g pixel matrix are sorted by pixel intensity, and the median of the pixels in the g×g pixel matrix is ​​placed at the center pixel of the g×g pixel matrix. For edge pixels in the g×g pixel matrix, the missing pixels are padded with zeros.

[0026] The steps of the Gaussian filtering are as follows:

[0027] A convolution kernel of e×f is used to scan each pixel in the abnormal conveyor belt image, and a convolution operation is performed on each pixel, where both e and f are odd numbers.

[0028] In some embodiments, in step S3, the image enhancement process includes first performing brightness and contrast processing on the second filtered image to obtain a first enhanced image; and then performing histogram equalization processing on the first enhanced image to obtain a second enhanced image;

[0029] The brightness and contrast processing steps are as follows: the value of each pixel in the second filtered image is substituted into the formula y=ax+b for calculation, where a is the contrast adjustment coefficient, b is the brightness adjustment coefficient, x is the original pixel value, ranging from 0 to 255, and y is the adjusted pixel value, ranging from 0 to 255;

[0030] The steps of the histogram equalization process are as follows:

[0031] First, the original grayscale image of the conveyor belt image is obtained; then, each pixel of the original grayscale image is scanned in sequence to obtain the grayscale histogram of the original grayscale image; then, the cumulative distribution function of the grayscale histogram is calculated; and then, based on the cumulative distribution function and the histogram equalization principle, the input and output mapping function is obtained. The mapping function is as follows:

[0032]

[0033] Among them, S K is the value of the current gray level after being mapped by the cumulative distribution function, n is the sum of the pixels in the image, n j is the number of pixels at the current gray level, and L is the total number of gray levels in the image;

[0034] Finally, the first enhanced image is input into the mapping function, and a second enhanced image is obtained after being processed by the mapping function;

[0035] The image normalization process is used to process the second enhanced image to obtain the pre-processed image.

[0036] In a certain embodiment, in step S4, after the calibration images are obtained, sample expansion is performed on all calibration images by using a rotation enhancement method and / or a mosaic enhancement method, thereby increasing the number of calibration images.

[0037] In a certain embodiment, in step S5, when training the yolov5 model, the loss function of the yolov5 model is replaced with the CIOU_Loss loss function, the activation function of the yolov5 model is replaced with the ReLU activation function, the upsampling method of the yolov5 model is replaced with the bilinear interpolation method, and the SE attention mechanism is added to the C3 module of the yolov5 model. Beneficial effects

[0038] Compared with the prior art, the present invention has the following advantages: by obtaining abnormal conveyor belt images in advance, performing deep learning training of the yolov5 model based on the defect types on the abnormal conveyor belt images, and performing conveyor belt tear detection based on the trained yolov5 model, the present invention can reduce the impact of the environment and reduce the false alarm and missed alarm rates. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] FIG1 is a flow chart of the present invention;

[0040] FIG2 is a schematic diagram of pixel distribution;

[0041] FIG3 is a schematic diagram of performing zero padding on the pixel in the upper left corner of FIG2 during median filtering;

[0042] Figure 4 is a schematic diagram of using a rectangular frame to demarcate slender defects. Modes for Carrying Out the Invention

[0043] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0044] As shown in FIG1 , a method for detecting longitudinal tearing of a conveyor belt includes the following steps:

[0045] S1: Collect conveyor belt images.

[0046] Specifically, in actual use, a camera can be used to capture images of the conveyor belt. The camera can be mounted below the conveyor belt to capture images of the conveyor belt. The image capture frequency matches the conveyor belt's operating speed and the camera's frame rate. The faster the conveyor belt, the faster the image capture frequency; the slower the conveyor belt, the lower the image capture frequency. The higher the camera's frame rate, the faster the image capture frequency; the lower the camera's frame rate, the slower the image capture frequency.

[0047] S2: Screening the conveyor belt images to filter out abnormal conveyor belt images.

[0048] Since the collected conveyor belt images include normal conveyor belt images and abnormal conveyor belt images, and when performing model learning and training in the following steps, it is necessary to learn the abnormal conveyor belt images, so it is necessary to filter out the abnormal conveyor belt images.

[0049] Specifically, the abnormal conveyor belt images can be screened out by using an SVM classifier. Specifically, two labels, normal and abnormal, are set in the SVM classifier, and then the abnormal conveyor belt images are screened out by using the SVM classifier.

[0050] To improve the accuracy of screening abnormal conveyor belt images, M SVM classifiers can be used to detect each conveyor belt image in the image sample from M feature dimensions, where M is an integer greater than 1. When the detection result of at least one SVM classifier is abnormal, the conveyor belt image is considered an abnormal conveyor belt image. For example, four SVM classifiers are used to detect the conveyor belt image from four dimensions: hog features, hsv histogram features, cumulative histogram features, and raw pixel features. When the detection result of at least one of the four SVM classifiers is abnormal, the conveyor belt image is considered an abnormal conveyor belt image.

[0051] S3: performing image filtering, image enhancement, and image normalization processing on the abnormal conveyor belt image in sequence to obtain a preprocessed image. The image normalization processing is used to scale the abnormal conveyor belt image to the same size.

[0052] In actual use, step S3 can eliminate irrelevant information in the abnormal conveyor belt image, enhance the detectability of relevant information, and thus ensure the reliability of subsequent image recognition.

[0053] Specifically, image filtering involves first performing a median filter on the abnormal conveyor belt image to obtain a first filtered image, and then performing a Gaussian filter on the first filtered image to obtain a second filtered image. In practice, median filtering can remove impulse noise and salt-and-pepper noise while preserving edge details in the abnormal conveyor belt image, while Gaussian filtering can smooth the image.

[0054] In this embodiment, the steps of median filtering are as follows:

[0055] First, use a g×g mask to sequentially traverse each pixel in the abnormal conveyor belt image, where g is an odd number greater than 1. The pixel and the (g×g-1) pixels surrounding it form a g×g pixel matrix, with the pixel at the center of the g×g pixel matrix. The pixels in the g×g pixel matrix are sorted by pixel intensity, and the median value of the pixels in the g×g pixel matrix is ​​placed at the center pixel of the g×g pixel matrix. For edge pixels in the g×g pixel matrix, the missing pixels are padded with zeros. A 3×3 mask is used to perform a median filter on the pixel distribution diagram shown in Figure 2. For the pixel in the upper left corner, the pixel distribution diagram after zero padding is shown in Figure 3. In actual use, pixel zero padding can facilitate the use of the same calculation method during program calculations, eliminating the need to design different calculation methods for pixels in different positions.

[0056] In this embodiment, the steps of Gaussian filtering are as follows:

[0057] Use a convolution kernel of size e×f to scan each pixel in the abnormal conveyor belt image and perform a convolution operation on each pixel. Both e and f are odd numbers. Specifically, assuming the size of the convolution kernel is e×f, substitute the coordinates of each pixel in the abnormal conveyor belt image into the following formula to calculate the two-dimensional Gaussian convolution kernel. The formula is as follows:

[0058]

[0059] in

[0060]

[0061]

[0062] Specifically, the image enhancement process includes first performing brightness and contrast processing on the second filtered image to obtain a first enhanced image; and then performing histogram equalization processing on the first enhanced image to obtain a second enhanced image.

[0063] The brightness and contrast processing steps are as follows: the value of each pixel in the second filtered image is substituted into the formula y=ax+b for calculation, where a is the contrast adjustment coefficient, b is the brightness adjustment coefficient, x is the original pixel value, ranging from 0 to 255, and y is the adjusted pixel value, ranging from 0 to 255;

[0064] The steps of histogram equalization are as follows:

[0065] First, obtain the original grayscale image of the conveyor belt image; then scan each pixel of the original grayscale image in sequence to obtain the grayscale histogram of the original grayscale image; then calculate the cumulative distribution function of the grayscale histogram; then, based on the cumulative distribution function and the histogram equalization principle, obtain the mapping function between input and output. The mapping function is as follows:

[0066]

[0067] Among them, S K is the value of the current gray level after being mapped by the cumulative distribution function, n is the sum of the pixels in the image, n j is the number of pixels at the current gray level, and L is the total number of gray levels in the image;

[0068] Finally, the first enhanced image is input into the mapping function, and the second enhanced image is obtained after being processed by the mapping function.

[0069] The image normalization process is used to process the second enhanced image to obtain a preprocessed image.

[0070] S4: Manually calibrate the preprocessed image to identify defects in the preprocessed image and generate a calibrated image.

[0071] Since subsequent model training requires knowing the specific defects on the conveyor belt image and the corresponding type of each defect, the defect locations on the preprocessed image need to be calibrated. Specifically, the defects on the preprocessed image can be framed using a rectangular frame in the marking software. For slender defects with fewer pixels, continuous small frames can be used for calibration to increase the proportion of the defect within the rectangular frame. Specifically, a schematic diagram of using rectangular frames to calibrate slender defects is shown in Figure 4. In addition, each defect in the same preprocessed image must be calibrated separately, and defects in the same location in the same preprocessed image can only be marked as one defect.

[0072] In addition, when the number of preprocessed images cannot meet the requirements of model learning and training, all calibration images can be sampled and expanded through rotation enhancement and / or mosaic enhancement to increase the number of calibration images.

[0073] Specifically, rotation enhancement is to rotate the preprocessed image by a certain angle to increase the number of training samples. In addition, rotation enhancement can also add appropriate noise data.

[0074] Specifically, Mosaic enhancement randomly flips, crops, and scales four images, and then randomly arranges and splices them into one image. While enriching the data set, it also adds small sample targets.

[0075] S5: Divide the calibration images into training set, validation set and test set; train the yolov5 model based on the training set; and use the validation set to confirm the convergence effect of the yolov5 model.

[0076] Specifically, when training the YOLOv5 model, the relevant parameters are set as follows: the number of training rounds epochs is set to 100, the pre-training model uses a lightweight model, the batch training sample number batch_size is set to 64, the learning rate optimizer optimizer is set to "adam", and label smoothing is not enabled.

[0077] In addition, although the lightweight yolov5 model runs faster, its accuracy is relatively low. In order to improve the accuracy of the model, in this embodiment, the loss function of the yolov5 model is replaced with the CIOU_Loss loss function. The CIOU_Loss loss function takes into account the overlapping area and the center point distance. When the target box wraps the prediction box, the distance between the two boxes is directly measured. Therefore, the CIOU_Loss loss function converges faster; and by using the CIOU_Loss loss function, the problem of how to minimize the normalized distance between the prediction box and the target box, as well as the problem of how to regress more accurately when the prediction box and the target box overlap, is solved. The definition of the CIOU_Loss loss function is as follows:

[0078]

[0079] Among them, v is a parameter to measure the consistency of aspect ratio, which is defined as follows:

[0080]

[0081] If activation functions such as sigmoid are used in deep learning models, the amount of calculation will be relatively large and the gradient will disappear during back propagation, making it impossible to complete the training of the deep network. Based on this, in this embodiment, when training the yolov5 model, the activation function of the yolov5 model is replaced with the ReLU activation function. Compared with activation functions such as sigmoid, the use of the ReLU activation function has the following advantages: there is no gradient saturation problem, the calculation speed is fast, there is no saturation area, there is no gradient disappearance problem, it can prevent gradient diffusion, the sparsity of the network reduces the interdependence of parameters, alleviates the occurrence of overfitting problems, there is no complex exponential operation, the calculation is simple, the efficiency is improved, and the actual convergence speed is fast. In addition, due to its piecewise linear nature, the ReLU activation function causes its forward propagation, backward propagation, and derivation to be piecewise linear; while traditional functions such as sigmoid are easy to discard information during the propagation process due to saturation at both ends.

[0082] The expression of the ReLU activation function is as follows:

[0083]

[0084] Here, i represents different channels.

[0085] In order to improve the accuracy of the model, in this embodiment, the upsampling method of the yolov5 model is replaced with a bilinear interpolation method, which is specifically implemented as follows:

[0086] Suppose we want to find the value of an unknown function f at point P=(x, y). We know the values ​​of the function f at four points: Q11=(x1, y1), Q12 =(x1, y2), Q21=(x2, y1), and Q22=(x2, y2). Here, f(x, y) is the pixel value of a pixel.

[0087] First, perform linear interpolation in the x direction and obtain:

[0088]

[0089]

[0090] Then perform linear interpolation in the y direction to obtain:

[0091]

[0092] To sum up, this is the final result of bilinear interpolation:

[0093]

[0094] In order to improve the model accuracy, the SE attention mechanism is added to the C3 module of the yolov5 model. The specific implementation is as follows:

[0095] First obtain the h×w×c feature map;

[0096] Then perform global average pooling (pooling size is h×w) to obtain a 1×1×c feature map, and then obtain the global receptive field;

[0097] Then it goes through two fully connected layers in sequence. The number of neurons in the first fully connected layer is c / r (r is generally 16). The first fully connected layer can reduce the dimension of the feature map, thereby reducing the number of parameters and computational complexity. The second fully connected layer increases the dimension to c neurons and adds more nonlinear processing processes to fit the complex correlations between channels.

[0098] Then the normalized weights are obtained through the sigmoid nonlinear activation function to obtain a 1×1×c feature map;

[0099] Finally, the full multiplication operation of the original h×w×c feature map and the 1×1×c feature map is performed. Through the full multiplication operation, the 1×1×c feature map is weighted channel by channel to each channel of the h×w×c feature map, completing the recalibration of the original features by channel attention. This step restores the feature map to the original number of channels and completes the establishment of inter-channel correlation.

[0100] S6: After the training of the Yolov5 model is completed, the trained Yolov5 model is used to test the test set. If the test success rate of the test set is higher than the first judgment threshold, the trained Yolov5 model is used as the final Yolov5 model. Otherwise, the trained Yolov5 model is used as the current pre-training model, and then return to step S5 to retrain the current pre-training model until the final Yolov5 model is obtained. In actual use, the first judgment threshold is set according to the required accuracy of the model.

[0101] S7: Detect the actual image of the conveyor belt based on the final YOLOv5 model and output the detection results. Each detection result includes detection category information, coordinate information, and confidence information. The process of outputting the detection results is as follows:

[0102] First, construct p detection result sets Hi, where p is the number of detection categories set by the final yolov5 model, i is between 0 and p-1, and one detection result set Hi corresponds to one detection category. The detection results are placed in the corresponding detection result set Hi according to the detection category.

[0103] Then, the detection results in the detection result set Hi are sorted in the order of confidence, with the sorting direction being from large to small;

[0104] Finally, the intersection-over-union ratio between each two detection results in each detection result set is calculated based on the coordinate information of the detection results. If the intersection-over-union ratio is higher than the second judgment threshold, the two detection results are considered to overlap with each other, and the detection result with the lower confidence level is removed from the detection result set. Otherwise, both detection results are retained. The interpretation of the intersection-over-union ratio is as follows:

[0105] The rectangular area of ​​each detection result is calculated based on the coordinate information of the detection results. The intersection-over-union ratio refers to the overlapping area of ​​the rectangular areas corresponding to two detection results.

[0106] In actual use, for a single conveyor belt image, the trained yolov5 model may output multiple detection results after detection. However, step S7 can filter out overlapping detection results, thereby ensuring that the detection results output by the model are valid.

[0107] S8: Confirming the final detection result in step S7, the confirmation being achieved by at least one of confidence screening, detection category screening, and target area screening;

[0108] The confidence screening process is as follows: set the confidence threshold of a single detection category and remove the detection results with confidence less than the confidence threshold;

[0109] The process of detection category screening is as follows: set the target detection category and remove the detection results whose detection category is not in the target detection category;

[0110] The process of target area screening is as follows: set the target area and filter out the detection results that are not in the target area.

[0111] In actual use, the final test results can be further screened through step S8 to delete the test results that do not meet the requirements.

[0112] In summary, the present invention obtains abnormal conveyor belt images in advance, performs deep learning training of the yolov5 model based on the defect type on the abnormal conveyor belt image, and performs conveyor belt tear detection based on the trained yolov5 model, thereby reducing the impact of the environment and reducing the false alarm and missed alarm rates.

[0113] The above description is for inspiration. Based on the above description, relevant personnel can make various changes and modifications without departing from the technical concept of this invention. The technical scope of this invention is not limited to the content of the specification, but must be determined according to the scope of the claims.

Claims

1. A method for detecting longitudinal tearing of a conveyor belt, characterized in that: The steps include: S1: Collect conveyor belt images; S2: screening the conveyor belt images, thereby screening out abnormal conveyor belt images in the conveyor belt images; S3: performing image filtering, image enhancement and image normalization processing on the abnormal conveyor belt image in sequence to obtain a preprocessed image, wherein the image normalization processing is used to scale the abnormal conveyor belt image to the same size; S4: manually calibrating the preprocessed image, thereby calibrating defects in the preprocessed image and generating a calibrated image; S5: Divide all calibrated images into a training set, a validation set and a test set; perform yolov5 model training based on the training set; the validation set is used to confirm the convergence effect of the yolov5 model; S6: After the training of the yolov5 model is completed, the test set is tested using the trained yolov5 model. If the test success rate of the test set is higher than the first judgment threshold, the trained yolov5 model is used as the final yolov5 model. Otherwise, the trained yolov5 model is used as the current pre-trained model, and then the process returns to step S5 to retrain the current pre-trained model until the final yolov5 model is obtained. S7: Based on the final yolov5 model, the actual image of the conveyor belt is detected, and the detection results are output. Each detection result includes detection category information, coordinate information and confidence information. The process of outputting the detection results is as follows: First, construct p detection result sets Hi, where p is the number of detection categories set by the final yolov5 model, i is between 0 and p-1, and one detection result set Hi corresponds to one detection category. Put the detection results into the corresponding detection result set Hi according to the detection category; Then, the detection results in the detection result set Hi are sorted in the order of confidence, from large to small; Finally, the intersection-and-union ratio between every two detection results in each detection result set is calculated according to the coordinate information of the detection results. If the intersection-and-union ratio is higher than the second determination threshold, it is considered that the two detection results overlap with each other, and the detection result with low confidence in the two detection results is removed from the detection result set, otherwise the two detection results are retained; S8: Confirming the final detection result in step S7, wherein the confirmation is achieved by at least one of confidence screening, detection category screening, and target area screening; The confidence screening process is as follows: setting a confidence threshold for a single detection category, and removing detection results with a confidence less than the confidence threshold; The process of the detection category screening is as follows: setting the target detection category, and removing the detection results whose detection category is not in the target detection category; The target area screening process is as follows: setting the target area, and screening out the detection results that are not in the target area.

2. A method for detecting longitudinal tearing of a belt conveyor belt according to claim 1, characterized in that: The method for screening out abnormal conveyor belt images in the conveyor belt images in step S2 is as follows: Two labels, normal and abnormal, are set in the SVM classifier, and then the abnormal conveyor belt images in the conveyor belt images are screened out using the SVM classifier.

3. A method for detecting longitudinal tearing of a belt conveyor belt according to claim 2, characterized in that: M SVM classifiers are used to detect each conveyor belt image from M feature dimensions respectively. When the detection result of at least one SVM classifier is abnormal, the conveyor belt image is regarded as an abnormal conveyor belt image.

4. A method for detecting longitudinal tearing of a belt conveyor belt according to claim 1, characterized in that: In step S3, the image filtering includes first performing median filtering on the abnormal conveyor belt image to obtain a first filtered image, and then performing Gaussian filtering on the first filtered image to obtain a second filtered image; The steps of median filtering are as follows: First, use a g×g mask to traverse each pixel point in the abnormal conveyor belt image in sequence, where g is an odd number greater than 1, wherein the pixel point and (g×g-1) pixel points around the pixel point form a g×g pixel matrix, and the pixel point is the center of the g×g pixel matrix. The pixel points in the g×g pixel matrix are sorted by pixel point intensity, and the median value of the pixel points in the g×g pixel matrix is ​​placed at the center pixel point position of the g×g pixel matrix. For the edge pixel points in the g×g pixel matrix, the missing pixel points are filled with zeros; The steps of Gaussian filtering are as follows: A convolution kernel of e×f is used to scan each pixel in the abnormal conveyor belt image, and a convolution operation is performed on each pixel, where both e and f are odd numbers.

5. A method for detecting longitudinal tearing of a belt conveyor belt according to claim 4, characterized in that: In step S3, the image enhancement process includes first performing brightness and contrast processing on the second filtered image to obtain a first enhanced image; then performing histogram equalization processing on the first enhanced image to obtain a second enhanced image; The brightness and contrast processing steps are as follows: the value of each pixel in the second filtered image is substituted into the formula y=ax+b for calculation, wherein a is the contrast adjustment coefficient, b is the brightness adjustment coefficient, x is the original pixel value, and the value range is 0 to 255, and y is the adjusted pixel value, and the value range is 0 to 255; The steps of the histogram equalization process are as follows: First, the original grayscale image of the conveyor belt image is obtained; then, each pixel of the original grayscale image is scanned in sequence to obtain the grayscale histogram of the original grayscale image; then, the cumulative distribution function of the grayscale histogram is calculated; then, the mapping function of input and output is obtained according to the cumulative distribution function and the histogram equalization principle, and the mapping function is as follows: Among them, S K is the value of the current gray level after being mapped by the cumulative distribution function, n is the sum of the pixels in the image, n j is the number of pixels at the current gray level, and L is the total number of gray levels in the image; Finally, the first enhanced image is input into the mapping function, and a second enhanced image is obtained after being processed by the mapping function; The image normalization process is used to process the second enhanced image to obtain the preprocessed image.

6. A method for detecting longitudinal tearing of a conveyor belt according to claim 1, characterized in that: In step S4, after the calibration images are obtained, sample expansion is performed on all calibration images by means of a rotation enhancement method and / or a Mosaic enhancement method, thereby increasing the number of calibration images.

7. A method for detecting longitudinal tearing of a conveyor belt according to claim 1, characterized in that: In step S5, when training the yolov5 model, the loss function of the yolov5 model is replaced with the CIOU_Loss loss function, the activation function of the yolov5 model is replaced with the ReLU activation function, the upsampling method of the yolov5 model is replaced with the bilinear interpolation method, and the SE attention mechanism is added to the C3 module of the yolov5 model.

Citation Information

Patent Citations

  • Image detection method based on defect detection and segmentation of deep convolutional neural network

    CN112906689A

  • Intrinsic safety type mining conveyor belt surface damage detection system and detection method

    CN113702393A

  • Multi-scale tiny flaw detection method based on attention mechanism

    CN115170529A

  • Substantially seamless electrostatographic member fabrication apparatus

    US20020074520A1

Cited By

  • Image recognition method for titanium alloy online detection sensor calibration

    CN119395040A

  • Noodle moisture online detection system and method

    CN120404718A

  • Coal conveying belt deviation detection method and system

    CN120495790A

  • Robot vision detection method and system based on image enhancement

    CN120563475A

  • Intelligent monitoring and detecting method and system of belt conveyor and electronic equipment

    CN121612380A