A belt foreign matter detection method and system based on deep feature reconstruction and a storage medium

Abstract

This invention discloses a method, system, and storage medium for foreign object detection on conveyor belts based on deep feature reconstruction. The method includes: acquiring positive sample images of the conveyor belt during normal operation without foreign objects and performing size normalization processing; performing feature normalization and alignment preprocessing on the positive sample images; inputting the preprocessed images into a foreign object detection model, using positive samples for training to learn the feature distribution of the conveyor belt during normal operation and complete image reconstruction; inputting the image to be detected into the trained model to obtain the reconstructed image, calculating the reconstruction error, and determining the foreign object region based on a threshold; performing connected component analysis on the foreign object region, and implementing graded alarms based on area and shape parameters. This invention eliminates the need for foreign object samples and annotations, reducing implementation costs, has strong resistance to environmental and material fluctuation interference, low false alarm and false negative rates, high automation, and is adaptable to various industrial conveyor belt transportation scenarios, possessing extremely high engineering application value.

Description

Technical Field

[0001] This invention relates to the field of belt conveyor safety detection technology, specifically to a belt conveyor foreign object detection method, system and storage medium based on depth feature reconstruction, applicable to industrial scenarios such as coal mines, metal mines, ports, sintering plants, and thermal power plants that use belt conveyors for material transportation. Background Technology

[0002] As a core piece of equipment for continuous material transport in the industrial field, belt conveyors have advantages such as high transport efficiency, low energy consumption, and wide applicability, and are widely used in mining, port loading and unloading, power plant coal transportation, and other scenarios. During actual operation, foreign objects such as anchor bolts, channel steel, wood blocks, large pieces of gangue, and metal debris can easily enter the belt conveyor system along with the materials. These foreign objects can cause serious damage to the belt conveyor: small, sharp objects can easily scratch the belt surface, shortening its service life; large, long objects can easily get stuck between rollers and idlers, causing belt tearing, misalignment, and even equipment shutdowns, personnel injuries, and other safety accidents, resulting in huge economic losses and safety hazards.

[0003] Currently, conveyor belt foreign object detection mainly employs methods such as manual inspection, traditional machine vision inspection, and supervised inspection based on deep learning. Manual inspection relies heavily on the experience and responsibility of inspectors, resulting in low efficiency, high false negative rates, high labor intensity, and significant susceptibility to human factors, making it difficult to meet the inspection demands of modern industrial automation and intelligence. Traditional machine vision inspection methods are mostly based on simple feature comparisons (such as grayscale thresholding and edge detection), easily affected by changes in ambient lighting, dust interference, and fluctuations in material particle size and texture, leading to a high false alarm rate and insufficient detection accuracy for complex working conditions. While supervised inspection methods based on deep learning offer higher detection accuracy, they require collecting a large number of foreign object samples of different types and shapes and manually labeling them to ensure the model's generalization ability and detection accuracy. However, in actual industrial settings, foreign objects are rare and diverse, making it time-consuming and labor-intensive to obtain sufficient and comprehensive samples, with extremely high labeling costs. Furthermore, when new types of foreign objects appear, the model needs to be retrained, resulting in poor adaptability and difficulty in meeting the requirements for long-term stable operation. Furthermore, some existing unsupervised detection methods rely solely on local feature comparisons and fail to effectively normalize the features of normal samples. This makes it impossible to effectively eliminate the impact of material fluctuations and environmental interference, resulting in insufficient robustness of the model to normal operating conditions, a high false alarm rate, and difficulty in adapting to the complex environment of industrial sites.

[0004] Therefore, in view of the problems of difficulty in obtaining samples, high annotation costs, weak anti-interference ability, high false alarm and false negative rates, and poor adaptability of existing technologies, it has become an urgent technical problem to be solved in this field to develop a belt foreign object detection method that does not require foreign object samples, can be trained solely based on normal operating images, has strong anti-interference ability, high detection accuracy, and high practicality. Summary of the Invention

[0005] The purpose of this invention is to provide a foreign object detection method for belts based on deep feature reconstruction, aiming to solve the problems of existing detection methods such as reliance on foreign object samples, high annotation costs, weak anti-interference ability, and high false alarm and false negative rates. This method achieves high-precision, low-false-alarm, automated detection and graded early warning of foreign objects on belts. The specific technical solution is as follows: A method for detecting foreign objects on a belt based on depth feature reconstruction includes: Collect positive sample images of the conveyor belt during normal operation without foreign objects and perform size normalization processing; The positive sample images are preprocessed with feature normalization and alignment. The preprocessed image is input into the foreign object detection model, and positive samples are used for training to learn the feature distribution when the belt is running normally and to complete image reconstruction. The image to be detected is input into the trained foreign object detection model to obtain a reconstructed image. The reconstruction error between the image to be detected and its reconstructed image is calculated. The foreign object region is determined based on the reconstruction error and the set error threshold. Connectivity analysis is performed on the foreign object area, and hierarchical alarms are implemented based on the area and shape parameters of the foreign object area.

[0006] Preferably, the positive sample images are acquired using a camera perpendicular to the belt plane, wherein the field of view completely covers the width direction of the belt, and the camera frame rate is... satisfy , To capture the range of the field of view along the direction of belt movement, This represents the maximum operating speed of the belt.

[0007] Preferably, the feature normalization alignment preprocessing is implemented using histogram specification, specifically as follows: Extract the R, G, and B color channels of all training positive sample images respectively, and statistically analyze the histogram distribution of each channel; The average value of all histograms for each channel is calculated to obtain the target histograms for the R, G, and B channels respectively. Then, the histograms of the R, G, and B channels of each training image are mapped and transformed into the target histograms of the corresponding channels.

[0008] Preferably, the foreign object detection model includes a VGG16 multi-scale feature extraction network, a multi-scale feature alignment and fusion module, and a convolutional autoencoder; The VGG16 multi-scale feature extraction network extracts shallow, mid, and deep features from the preprocessed image. The multi-scale feature alignment and fusion module then merges these features by aligning their sizes and unifying their channels to obtain multi-scale fused features. These multi-scale fused features are then input into a convolutional autoencoder for feature compression and image reconstruction. Positive sample training enables the foreign object detection model to learn the feature distribution during normal belt operation, thus completing the image reconstruction.

[0009] Preferably, the foreign object region is determined based on the reconstruction error and a set error threshold, specifically: The reconstruction error between the image to be detected and its reconstructed image is calculated pixel by pixel. Pixels with reconstruction errors greater than a set error threshold are identified as foreign object pixels, and a binary image is generated to distinguish the foreign object region from the normal background region.

[0010] Preferably, a connected component analysis is performed on the foreign object region, and a tiered alarm is implemented based on the area and shape parameters of the foreign object region. Specifically: Connected component detection is performed on the binary image to identify all independent foreign object regions. The number of connected components is counted, the area of ​​each connected component is calculated, and a minimum bounding ellipse is fitted to each connected component to obtain the major and minor axis parameters of the ellipse. The area of ​​each connected component is then used as the basis for further analysis. The ratio of the major axis to the minor axis of the ellipse Implement tiered alarms, among which This is the number of the connected component.

[0011] Preferably, the graded alarm specifically refers to: If the area of ​​the connected region < And the ratio of its major axis to its minor axis is < If the fluctuation is normal, it is considered a false alarm and no alarm is triggered. If the area of ​​the connected region < And the ratio of its major axis to its minor axis is ≥ If so, it is determined to be a long and thin foreign object, triggering a normal alarm; If the area of ​​the connected region ≥ And the ratio of its major axis to its minor axis is < If the object is large and without sharp foreign objects, a normal alarm will be triggered. If the area of ​​the connected region ≥ And the ratio of its major axis to its minor axis is ≥ If so, it is determined to be a large, long, and dangerous foreign object, triggering a serious alarm; in, This is the area alarm threshold. The alarm threshold is the ratio of the major axis to the minor axis.

[0012] This invention also provides a belt foreign object detection system based on depth feature reconstruction. This system employs the aforementioned belt foreign object detection method and includes: Image acquisition module: used to acquire positive sample images of the conveyor belt during normal operation without foreign objects and on-site images to be inspected, and output image data to the preprocessing module; Preprocessing module: Performs size normalization and feature normalization alignment preprocessing on the acquired images; Model training and inference module: The foreign object detection model is trained using the preprocessed image to learn the feature distribution when the belt is running normally; and features are extracted and image reconstruction is performed on the image to be detected, and the reconstructed image corresponding to the image to be detected is output. Foreign object detection and judgment module: used to calculate the reconstruction error between the image to be detected and its reconstructed image pixel by pixel, determine the foreign object region according to the error threshold and generate a binary image; Connectivity Analysis and Alarm Module: Used to perform connectivity detection on foreign object regions, calculate the area of ​​foreign object regions, fit the minimum circumscribed ellipse of foreign object regions to obtain the major-minor axis ratio parameter, and execute hierarchical alarm output according to the area of ​​foreign object regions and the major-minor axis ratio parameter; Control output module: Used to upload alarm information to the host computer or belt control system.

[0013] The present invention also provides another belt foreign object detection system based on deep feature reconstruction, including a memory and a processor, wherein the memory stores a computer program, and the computer program executes the belt foreign object detection method when it is run.

[0014] The present invention also provides a storage medium storing a computer program, which, when run, executes the aforementioned belt foreign object detection method.

[0015] The application of the technical solution of the present invention has the following beneficial effects: This invention adopts an unsupervised learning approach, requiring only the collection of images of the conveyor belt under normal operating conditions without foreign objects as training samples. There is no need to collect or label foreign object samples, nor is there any need for manual participation in the data labeling process. This solves the problems of traditional supervised detection methods that rely on a large number of foreign object samples, are difficult to obtain, have high labeling costs, and have long cycles. It is particularly suitable for industrial conveyor belt transportation scenarios where foreign object types are varied and the probability of occurrence is low.

[0016] This invention performs feature normalization and alignment processing on positive sample images through histogram specification, effectively eliminating the feature dispersion problem caused by changes in lighting, material particle size, texture density, dust and exposure differences between different samples. This makes the features of normal conveyor belts and materials more concentrated and stable, greatly improving the robustness of the model to normal working conditions, reducing the false alarm rate from the source, and meeting the needs of long-term stable operation in industrial sites.

[0017] This invention extracts shallow, medium and deep multi-scale features through the VGG16 network, and performs size alignment, channel unification and channel dimension fusion on features at different levels. This allows the model to simultaneously retain image edge texture, local material structure and overall semantic information. Compared with single-scale features, it has stronger feature expression capabilities and high sensitivity to small foreign objects, long foreign objects and large foreign objects, effectively avoiding missed detections.

[0018] This invention utilizes a convolutional autoencoder to learn and reconstruct the feature distribution of normal conveyor belts and materials. Foreign objects, because they do not belong to the normal feature distribution, will have a significantly higher reconstruction error than normal areas. This reconstruction error is used to accurately distinguish foreign objects from the normal background. This method is not limited by material type, color, or particle shape, has strong generalization ability, and its detection logic is stable and reliable.

[0019] This invention performs connectivity analysis on the foreign object area to obtain quantitative parameters such as the foreign object area and the major and minor axes of the minimum circumscribed ellipse. It then implements graded alarms based on the area alarm threshold and the major and minor axis ratio alarm threshold, which can effectively distinguish between normal material fluctuations, ordinary foreign objects, and high-risk long foreign objects, avoid invalid alarms, improve on-site operation and maintenance efficiency, and provide a reliable basis for the safety control of belt conveyors.

[0020] This invention requires no structural modification to existing belt conveyor systems and can be deployed using only image acquisition equipment and computing units. It is adaptable to various complex industrial environments such as coal mines, metal mines, ports, power plants, and sintering plants. The model training and inference process is highly automated, operates stably, and has strong engineering application value and promotion prospects.

[0021] In addition to the objectives, features, and advantages described above, the present invention has other objectives, features, and advantages. The invention will now be described in further detail with reference to the figures. Attached Figure Description

[0022] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a flowchart of the belt foreign object detection method based on depth feature reconstruction in Example 1; Figure 2 This is a schematic diagram showing the arrangement of the cameras above the belt. Detailed Implementation

[0023] To facilitate understanding of the present invention, a more complete description is provided below, along with preferred embodiments. However, the present invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to provide a thorough and complete understanding of the disclosure of the present invention.

[0024] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention.

[0025] Example: See Figure 1 This embodiment provides a method for detecting foreign objects on a belt based on depth feature reconstruction, including: Collect positive sample images of the conveyor belt during normal operation without foreign objects and perform size normalization processing; The positive sample images are preprocessed with feature normalization and alignment. The preprocessed image is input into the foreign object detection model, and positive samples are used for training to learn the feature distribution when the belt is running normally and to complete image reconstruction. The image to be detected is input into the trained foreign object detection model to obtain a reconstructed image. The reconstruction error between the image to be detected and its reconstructed image is calculated. The foreign object region is determined based on the reconstruction error and the set error threshold. Connectivity analysis is performed on the foreign object area, and hierarchical alarms are implemented based on the area and shape parameters of the foreign object area.

[0026] The belt foreign object detection method in this embodiment adopts an unsupervised learning approach, relying solely on images of the belt running normally without foreign objects to complete model training. Through deep feature reconstruction, reconstruction error judgment, and graded alarm, it achieves high-precision and low-false-alarm detection of foreign objects on the belt.

[0027] The following is a detailed explanation of each step of the belt foreign object detection method in this embodiment: Preferably, positive sample images of the conveyor belt during normal operation without foreign objects are acquired and their size is normalized. Specifically: This step is used to obtain standard samples containing only belts and normal materials, providing stable and uniform input data for subsequent unsupervised model training.

[0028] like Figure 2As shown, in this embodiment, image acquisition utilizes a high-definition industrial camera. The camera is centrally positioned directly above the belt, perpendicular to the belt plane, and captures images with a field of view that completely covers the width of the belt. Preferably, the field of view covers the entire width of the belt. The camera should be 10-50cm wider than the conveyor belt. Later, simply trimming is sufficient to obtain the target acquisition area, ensuring that the acquisition area only includes the conveyor belt and the transported materials, eliminating interference from the machine frame, environmental background, etc. Furthermore, the camera frame rate... With the maximum running speed of the belt Matching settings to allow the belt to run at maximum speed. At the same time, overlapping areas are preserved between adjacent frames to prevent foreign objects from appearing in the frame gaps and causing missed detection. Specifically, the camera frame rate satisfy ,in The field of view is set to the range along the belt's movement direction. Simultaneously, a dedicated, uniform light source is used at the acquisition site to stabilize light intensity and reduce the impact of dust scattering, sudden changes in illumination, and uneven exposure on image quality.

[0029] Under normal operating conditions with no foreign objects on the conveyor belt, images were continuously acquired at different time periods, with different material loads and different equipment operating conditions, with a sample size of no less than 5000 images (those skilled in the art can flexibly adjust the sample size as needed), forming the model training sample set. To standardize the model input specifications and reduce the computational load of depth feature extraction, all original images were uniformly normalized to a fixed size of 3×1024×1024, completing the standardization processing of positive samples.

[0030] Preferably, since the materials transported by belt conveyors have natural differences in particle size, shape, and density distribution, and factors such as ambient light, dust, and shooting angle can cause fluctuations in image brightness and contrast, normal samples exhibit a certain degree of dispersion in pixel distribution, which can easily lead to model misjudgment. Therefore, this embodiment performs feature normalization and alignment preprocessing on positive sample images to eliminate apparent differences between samples, enhance the compactness of normal feature clusters, and significantly reduce the probability of false alarms.

[0031] In this embodiment, histogram specification is used to achieve feature normalization and alignment, specifically: First, extract the R, G, and B color channels of all training positive sample images respectively, and statistically analyze the histogram distribution of each channel. Subsequently, the average value of all histograms for each channel is calculated to obtain the target histograms for the three channels R, G, and B. Finally, the R, G, and B channel histograms of each training image are mapped and transformed into the target histograms of the corresponding channels. After this step, a new training sample set is obtained. In the new training sample set, the feature differences between normal samples caused by material texture, particle size, and lighting changes are significantly eliminated. All training samples maintain a high degree of consistency in color distribution and brightness distribution, which provides a foundation for the subsequent model to learn stable and compact normal features.

[0032] Preferably, the preprocessed image is input into the foreign object detection model, and positive samples are used for training to learn the feature distribution during normal belt operation and complete image reconstruction. Specifically: The foreign object detection model in this embodiment includes a VGG16 multi-scale feature extraction network, a multi-scale feature alignment and fusion module, and a convolutional autoencoder. The VGG16 extracts shallow, medium, and deep features from the preprocessed image. After aligning the size and unifying the channels of the shallow, medium, and deep features, the model is fused to obtain multi-scale fused features. The multi-scale fused features are then input into the convolutional autoencoder for feature compression and image reconstruction. Positive sample training is used to enable the model to learn the feature distribution when the belt is running normally, thereby achieving accurate reconstruction of normal images.

[0033] Preferably, VGG16 multi-scale feature extraction specifically involves: A pre-trained VGG16 network was used as the backbone for feature extraction, extracting three key features: Shallow features: taken from Conv2 layer, with a feature dimension of 128×256×256 (number of channels × height × width), mainly used to capture detailed information such as image edges, textures, and colors, and can reflect subtle features such as belt surface texture and material particle outlines; Mid-layer features: taken from Conv4 layer, with a feature dimension of 512×64×64, mainly used to extract local structural information of materials, which can reflect the aggregation state, local morphology and other features of materials; Deep features: taken from Conv5 layer, with a feature dimension of 512×32×32, mainly used to extract the overall semantic information of the image, which can reflect the overall distribution pattern when the belt is running normally.

[0034] Preferably, since the spatial dimensions and channel numbers of shallow, middle, and deep features are inconsistent, they cannot be directly fused. Therefore, size alignment and channel unification are required. In this embodiment, multi-scale feature alignment and fusion specifically involve: The mid-level and deep features are upsampled to the shallow feature size (256×256) through deconvolution; then, the number of channels is adjusted by 1×1 convolution, unifying the 128 channels in the shallow layer, 512 channels in the mid-level layer, and 512 channels in the deep layer to 256 channels; finally, they are concatenated along the channel dimension to obtain multi-scale fused features (i.e., ... This fusion feature retains detailed texture, local structure, and overall semantic information, possessing stronger feature expression capabilities and more comprehensively reflecting the characteristics of the belt during normal operation.

[0035] The preferred convolutional autoencoder structure and training are as follows: The convolutional autoencoder comprises an encoder and a decoder, primarily used for feature compression and image reconstruction. The encoder inputs a feature tensor (i.e., multi-scale fusion features) of 768×256×256, which is converted to 512×128×128 after convolution and max pooling. A second convolution and max pooling operation yields 256×64×64, resulting in a final bottleneck layer feature of 128×32×32. This bottleneck layer feature serves as the core feature output by the encoder, preserving the key distribution patterns during normal belt operation. The decoder inputs the bottleneck layer feature of 128×32×32, which is deconvolutioned and upsampled to 256×64×64. A second deconvolution and upsampled result in 512×128×128, ultimately outputting a reconstructed image of 3×1024×1024, consistent with the input image size.

[0036] Preferably, the foreign object detection model is trained using mean squared error (MSE) as the loss function to minimize the pixel difference between the original image and the reconstructed image, so that the model can fully learn the feature distribution when the belt is running normally and achieve accurate reconstruction of normal images.

[0037] Preferably, the image to be detected is input into a trained foreign object detection model to obtain a reconstructed image, the reconstruction error between the image to be detected and its reconstructed image is calculated, and the foreign object region is determined based on the reconstruction error and a set error threshold. Specifically: The image to be detected undergoes the same preprocessing as the training samples: first, it is normalized to a size of 3×1024×1024, and then histogram-defined feature alignment is performed. The processed image to be detected is then input into the trained model to obtain the reconstructed image.

[0038] Since foreign objects are abnormal targets that the model has not learned, the model cannot effectively fit the foreign object region, and the reconstruction error of the foreign object region is much larger than that of the normal region. The reconstruction error between the image to be detected and the reconstructed image is calculated pixel by pixel. Pixels with reconstruction errors greater than a set error threshold are identified as foreign objects, and a binary image is generated. The white area is the foreign object region, and the black area is the normal background, thus completing the segmentation and localization of the foreign object region.

[0039] Preferably, a connected component analysis is performed on the foreign object region, and a tiered alarm is implemented based on the area and shape parameters of the foreign object region. Specifically: Connected component detection is performed on the binary image (e.g., using an eight-neighbor connected component detection algorithm) to identify all independent foreign object regions. The number of connected components is counted, the area of ​​each connected component is calculated, and a minimum bounding ellipse is fitted to each connected component to obtain the major and minor axis parameters of the ellipse. The area of ​​each connected component is then used as the basis for further analysis. The ratio of the major axis to the minor axis of the ellipse (in The alarm is graded based on the ratio of the major axis length to the minor axis length. This is the number of the connected component.

[0040] Preferably, an area alarm threshold is set. alarm threshold based on ratio of major to minor axis This embodiment implements tiered alarms according to the following rules: If the area of ​​the connected region < And the ratio of its major axis to its minor axis is < If the fluctuation is normal, it is considered a false alarm and no alarm is triggered. If the area of ​​the connected region < And the ratio of its major axis to its minor axis is ≥ If so, it is determined to be a long and thin foreign object, triggering a normal alarm; If the area of ​​the connected region ≥ And the ratio of its major axis to its minor axis is < If the object is large and without sharp foreign objects, a normal alarm will be triggered. If the area of ​​the connected region ≥ And the ratio of its major axis to its minor axis is ≥ If it is identified as a large, long, and dangerous foreign object, a serious alarm will be triggered.

[0041] Preferably, the above threshold and It can be flexibly adjusted according to the type of materials, belt specifications and safety level requirements on site to ensure stable and reliable test results.

[0042] In conveyor belt foreign object detection scenarios, the difficulty in comprehensively collecting foreign object samples and the complexity and variety of foreign object types make it impossible to obtain sufficient and diverse images for model training, hindering the model from achieving ideal detection rates and robustness. This embodiment fully leverages the relatively stable types of materials transported by conveyor belts and the high feature similarity between normal materials at different times. It adopts a detection approach based on deep feature reconstruction: by collecting a large number of images of conveyor belts under normal operating conditions without foreign objects to construct a training sample set, a foreign object detection model relying solely on normal samples is trained. Simultaneously, feature normalization and distribution alignment are combined to reduce the feature distribution differences between normal samples, enhancing the compactness and stability of normal feature clusters. This achieves accurate online detection of non-transported materials (i.e., foreign objects), effectively reducing feature interference caused by differences in particle size and shape of normal materials, significantly reducing false alarms, and improving the applicability and reliability of the detection system in complex industrial environments.

[0043] Example 2: This embodiment provides a belt foreign object detection system based on depth feature reconstruction. The system in this embodiment adopts the belt foreign object detection method of Embodiment 1, and the system includes: Image acquisition module: Used to acquire positive sample images of the conveyor belt during normal operation without foreign objects, as well as on-site images to be inspected, and output image data to the preprocessing module. Preferably, the image acquisition module includes a high-definition industrial camera configured to be mounted perpendicular to the conveyor belt plane and a dedicated uniform light source.

[0044] Preprocessing module: This module performs size normalization on the acquired images, unifying them to 3×1024×1024, and performs histogram specification on the positive sample images to achieve feature normalization and alignment.

[0045] Model training and inference module: This module includes VGG16 multi-scale feature extraction network, multi-scale feature alignment and fusion module, and convolutional autoencoder. This module uses preprocessed positive samples to complete model training, learns the feature distribution when the belt is running normally, and performs feature extraction and image reconstruction on the image to be detected, outputting the reconstructed image.

[0046] Foreign object detection and judgment module: used to calculate the reconstruction error between the image to be detected and its reconstructed image pixel by pixel, determine the foreign object region based on the error threshold and generate a binary image.

[0047] Connectivity Analysis and Alarm Module: Used to perform connectivity detection on foreign object regions, calculate the area of ​​the foreign object region, fit the minimum circumscribed ellipse to obtain the major and minor axis ratio parameters, and execute hierarchical alarm output according to the area and major and minor axis ratio parameters.

[0048] Control output module: Used to upload alarm information to the host computer or belt control system, supporting audible and visual alarms and belt emergency stop linkage. The alarm information generally includes the location of the foreign object and the shape parameters of the foreign object, such as the area of ​​the foreign object region and the major and minor axes of the smallest circumscribed ellipse.

[0049] Example 3: This embodiment provides a belt foreign object detection system based on deep feature reconstruction, including a memory and a processor. The memory stores a computer program, which, when run, executes the belt foreign object detection method in Embodiment 1.

[0050] Example 4: This embodiment provides a storage medium storing a computer program, which, when run, executes the belt foreign object detection method in Embodiment 1.

[0051] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for detecting foreign objects on a belt based on depth feature reconstruction, characterized in that, include: Collect positive sample images of the conveyor belt during normal operation without foreign objects and perform size normalization processing; The positive sample images are preprocessed with feature normalization and alignment. The preprocessed image is input into the foreign object detection model, and positive samples are used for training to learn the feature distribution when the belt is running normally and to complete image reconstruction. The image to be detected is input into the trained foreign object detection model to obtain a reconstructed image. The reconstruction error between the image to be detected and its reconstructed image is calculated. The foreign object region is determined based on the reconstruction error and the set error threshold. Connectivity analysis is performed on the foreign object area, and hierarchical alarms are implemented based on the area and shape parameters of the foreign object area.

2. The method for detecting foreign objects in a belt based on depth feature reconstruction according to claim 1, characterized in that, The positive sample images are acquired using a camera perpendicular to the belt plane, wherein the field of view completely covers the width direction of the belt, and the camera frame rate is... satisfy , To capture the range of the field of view along the direction of belt movement, This represents the maximum operating speed of the belt.

3. The method for detecting foreign objects in a belt based on depth feature reconstruction according to claim 1, characterized in that, The feature normalization alignment preprocessing is implemented using histogram specification, specifically as follows: Extract the R, G, and B color channels of all training positive sample images respectively, and statistically analyze the histogram distribution of each channel; The average value of all histograms for each channel is calculated to obtain the target histograms for the R, G, and B channels respectively. Then, the histograms of the R, G, and B channels of each training image are mapped and transformed into the target histograms of the corresponding channels.

4. The method for detecting foreign objects in a belt based on depth feature reconstruction according to claim 1, characterized in that, The foreign object detection model includes a VGG16 multi-scale feature extraction network, a multi-scale feature alignment and fusion module, and a convolutional autoencoder. The VGG16 multi-scale feature extraction network extracts shallow, mid, and deep features from the preprocessed image. The multi-scale feature alignment and fusion module then merges these features by aligning their sizes and unifying their channels to obtain multi-scale fused features. These multi-scale fused features are then input into a convolutional autoencoder for feature compression and image reconstruction. Positive sample training enables the foreign object detection model to learn the feature distribution during normal belt operation, thus completing the image reconstruction.

5. The method for detecting foreign objects in a belt based on depth feature reconstruction according to claim 1, characterized in that, The foreign object region is determined based on the reconstruction error and the set error threshold, specifically: The reconstruction error between the image to be detected and its reconstructed image is calculated pixel by pixel. Pixels with reconstruction errors greater than a set error threshold are identified as foreign object pixels, and a binary image is generated to distinguish the foreign object region from the normal background region.

6. The method for detecting foreign objects in a belt based on depth feature reconstruction according to claim 5, characterized in that, Connectivity analysis is performed on the foreign object region, and a tiered alarm is implemented based on the area and shape parameters of the foreign object region. Specifically: Connected component detection is performed on the binary image to identify all independent foreign object regions. The number of connected components is counted, the area of ​​each connected component is calculated, and a minimum bounding ellipse is fitted to each connected component to obtain the major and minor axis parameters of the ellipse. The area of ​​each connected component is then used as the basis for further analysis. The ratio of the major axis to the minor axis of the ellipse Implement tiered alarms, among which This is the number of the connected component.

7. The method for detecting foreign objects in a belt based on depth feature reconstruction according to claim 6, characterized in that, The tiered alarm system specifically refers to: If the area of ​​the connected region < And the ratio of its major axis to its minor axis is < If the fluctuation is normal, it is considered a false alarm and no alarm is triggered. If the area of ​​the connected region < And the ratio of its major axis to its minor axis is ≥ If so, it is determined to be a long and thin foreign object, triggering a normal alarm; If the area of ​​the connected region ≥ And the ratio of its major axis to its minor axis is < If the object is large and without sharp foreign objects, a normal alarm will be triggered. If the area of ​​the connected region ≥ And the ratio of its major axis to its minor axis is ≥ If so, it is determined to be a large, long, and dangerous foreign object, triggering a serious alarm; in, This is the area alarm threshold. The alarm threshold is the ratio of the major axis to the minor axis.

8. A belt foreign object detection system based on depth feature reconstruction, characterized in that, The system employs the belt foreign object detection method as described in any one of claims 1-7, and the system comprises: Image acquisition module: used to acquire positive sample images of the conveyor belt during normal operation without foreign objects and on-site images to be inspected, and output image data to the preprocessing module; Preprocessing module: Performs size normalization and feature normalization alignment preprocessing on the acquired images; Model training and inference module: The foreign object detection model is trained using the preprocessed image to learn the feature distribution when the belt is running normally; and features are extracted and image reconstruction is performed on the image to be detected, and the reconstructed image corresponding to the image to be detected is output. Foreign object detection and judgment module: used to calculate the reconstruction error between the image to be detected and its reconstructed image pixel by pixel, determine the foreign object region according to the error threshold and generate a binary image; Connectivity Analysis and Alarm Module: Used to perform connectivity detection on foreign object regions, calculate the area of ​​foreign object regions, fit the minimum circumscribed ellipse of foreign object regions to obtain the major-minor axis ratio parameter, and execute hierarchical alarm output according to the area of ​​foreign object regions and the major-minor axis ratio parameter; Control output module: Used to upload alarm information to the host computer or belt control system.

9. A belt foreign object detection system based on depth feature reconstruction, characterized in that, It includes a memory and a processor, wherein the memory stores a computer program, which, when executed, performs the belt foreign object detection method as described in any one of claims 1-7.

10. A storage medium, characterized in that, The storage medium contains a computer program that, when executed, performs the belt foreign object detection method as described in any one of claims 1-7.

Images