A coal flow impurity rotating detection method based on physical blur modeling and scene customized attention

Abstract

This invention provides a method for detecting rotating debris in coal flow based on physical fuzzy modeling and scene-customized attention. The method includes: acquiring a clear coal flow image and then labeling non-coal debris targets in the image; establishing a dataset; constructing a debris detection network; constructing a physics-driven motion fuzzy mapping model; using the motion fuzzy mapping model to dynamically blur the clear coal flow image, generating training samples simulating motion fuzziness phenomena in a conveyor belt; training and optimizing the debris detection network; inputting a real-time coal flow image into the optimized network to obtain candidate detection results for non-coal debris targets; post-processing the candidate detection results and outputting the final filtered detection results. This invention can improve the stability and accuracy of debris detection under complex conditions of high-speed coal flow, reduce the influence of motion fuzziness and background interference, and provide technical support for the safety monitoring and automatic sorting of coal mine conveying systems.

Description

Technical Field

[0001] This invention relates to the technical field of coal flow debris detection, and more specifically, to a method for detecting rotating coal flow debris based on physical fuzzy modeling and scene-customized attention. Background Technology

[0002] Coal debris detection is a key technology in intelligent coal mine sorting systems, and its accuracy and stability directly affect the reliability of subsequent sorting and handling operations. Existing technologies mainly rely on computer vision methods to achieve target recognition and localization, but under conditions of high-speed conveyor belt transportation and complex coal quality backgrounds, traditional image processing algorithms and general deep learning models have shown significant limitations in adaptability.

[0003] Existing methods for detecting debris in coal flow mainly include background subtraction, edge detection, and general target detection models. Traditional methods typically rely on color differences or gradient information between the target and the background for discrimination. However, in underground environments, these methods are easily affected by dust obstruction, uneven lighting, and significant interference from low-contrast scenes, resulting in poor robustness and stability. While general deep learning detection models improve recognition accuracy to some extent, their training process is often based on clear or static images, limiting their generalization ability in dynamic coal flow scenarios, and their target localization accuracy is insufficient to meet industrial-grade grasping and sorting requirements.

[0004] Furthermore, existing coal flow debris detection technologies mainly suffer from the following shortcomings: (1) Limited resistance to motion ambiguity. Most algorithms do not consider the physical trailing effect caused by relative motion under high-speed conditions (such as conveyor belt speed greater than 1 m / s). There is a significant difference in domain distribution between training data and actual online inference data, which makes it easy for debris targets under motion ambiguity conditions to be missed or falsely detected; (2) Insufficient ability to represent weak texture features. In the same color but different material scenarios (such as coal and gangue, black rubber, etc.), the multiple downsampling operations of the general network are prone to cause the loss of key surface texture and edge detail information, making the debris features submerged by the complex coal flow background and difficult to achieve effective differentiation; (3) Lack of rotational attitude information. Existing methods generally use horizontal bounding boxes for target localization. For slender or arbitrarily positioned debris, the detection box often contains a large number of redundant background areas, which not only reduces the positioning accuracy but also fails to provide target main axis direction information, making it difficult to provide effective support for the back-end robotic arm to achieve accurate attitude perception and stable grasping.

[0005] In summary, existing coal flow debris detection technologies still have systemic shortcomings in terms of motion fuzziness adaptability, weak feature extraction capability, and precise positioning of rotating targets, making it difficult to meet the comprehensive application requirements of intelligent sorting systems for high precision, high robustness, and engineering feasibility under complex coal flow conditions. Summary of the Invention

[0006] To address the issues of poor anti-motion blurring capability and unstable extraction of weak features in existing visual inspection technologies under high-speed coal flow conditions, this invention provides a method for detecting rotating debris in coal flow based on physical fuzzy modeling and scene-customized attention.

[0007] To achieve the above objectives, the present invention provides a method for detecting rotating debris in coal flow based on physical fuzzy modeling and scene-customized attention. This method for detecting rotating coal flow debris based on physical fuzzy modeling and scene-customized attention includes: Step S1: acquiring a clear coal flow image of the conveyor belt under static or running conditions, and then labeling non-coal debris targets in the clear coal flow image using rotated bounding boxes to generate labeling information corresponding to the non-coal debris targets; Step S2: establishing a dataset based on the labeled coal flow image; Step S3: constructing a debris detection network including a feature extraction module, a feature fusion module, and a prediction output module; wherein, the feature extraction module is used to receive the clear coal flow image and perform convolution processing on the clear coal flow image to extract visual features and output a multi-scale feature map; the feature fusion module is used to fuse deep and shallow features of the multi-scale feature map; the feature fusion module includes a scene-customized attention module, which is used to enhance the features of non-coal debris targets and suppress coal flow background interference, and output enhanced fused features; the prediction output module is used to predict and output candidate detection results of non-coal debris targets based on the enhanced fused features; the candidate detection results include: non-coal debris targets. Step S4: Based on the operating parameters of the conveyor belt and the imaging system parameters for obtaining the clear coal flow image, construct a physical-driven motion blur mapping model; wherein, the motion blur mapping model is used to describe the mapping relationship between the clear image and the blurred image of the coal flow image; Step S5: Use the motion blur mapping model to perform dynamic blur processing on the clear coal flow image in the dataset to generate training samples simulating the motion blur phenomenon of the conveyor belt; Step S6: Based on the training samples and loss function, train and optimize the debris detection network; Step S7: Obtain real-time coal flow images of the conveyor belt under static or running conditions, input the real-time coal flow images into the optimized debris detection network, and obtain candidate detection results of non-coal debris targets; Step S8: Post-process the candidate detection results to filter out overlapping and redundant detection results, and output the final screening detection results containing the non-coal debris target category, the confidence level corresponding to the non-coal debris target category, the non-coal debris target rotation bounding box positioning information, and the actual physical size of the non-coal debris target.

[0008] Preferably, in step S1, the annotation of non-coal debris targets in the clear coal flow image specifically includes: the category label of the non-coal debris target, the geometric center coordinates of the rotated bounding box, the length of the long side of the rotated bounding box, the length of the short side of the rotated bounding box, and the rotation angle of the rotated bounding box.

[0009] The mathematical expression corresponding to the annotation information is:

[0010] (1)

[0011] In equation (1), A labeled truth vector representing a single non-coal debris target; The pixel coordinates representing the geometric center of the non-coal debris target; This represents the length of the longest side in pixels of the rotated bounding box of the non-coal debris target; This represents the shortest side length in pixels of the rotated bounding box of the non-coal debris target; This represents the angle between the longest side of the rotated bounding box of non-coal debris targets and the horizontal axis of the clear coal flow image. The range of values ​​is .

[0012] Preferably, in step S3, the scene-customized attention module includes: a local contrast feature enhancement branch module and a serial attention weighting branch module; wherein, the local contrast feature enhancement branch module is used to receive features output from the feature extraction module to the feature fusion module as input features of the scene-customized attention module; the local contrast feature enhancement branch module is also used to smooth the input features to extract local smooth background features; the local contrast feature enhancement branch module is also used to generate texture mutation features based on the difference between the input features and the local smooth background features; the serial attention weighting branch module is used to perform channel dimension weighting and spatial dimension weighting on the texture mutation features to obtain weighted saliency features; the serial attention weighting branch module is also used to fuse the weighted saliency features with the input features of the scene-customized attention module to output fused enhanced features;

[0013] The expression corresponding to the enhanced features after fusion is:

[0014] (2)

[0015] In equation (2), This represents the enhanced features output by the serial attention-weighted branch module; This represents the input features of the scene-customized attention module; This represents the significance of the weighted feature.

[0016] Preferably, the local contrast feature enhancement branch module is specifically used to: perform a convolution on the input features of the scene-customized attention module using formula (3), with a kernel size of [missing value]. The depthwise separable convolution operation is then sequentially passed through a batch normalization layer and... The activation function layer generates locally smoothed background features; the expression corresponding to the locally smoothed background features is:

[0017] (3)

[0018] In equation (3), Indicates localized smooth background features; express Activation function; Indicates batch normalization; Indicates the kernel size as Depth-separable convolution; This represents the input features of the scene-customized attention module;

[0019] Using formula (4), the absolute difference between the input features of the scene-customized attention module and the local smooth background features is calculated to generate texture mutation features; wherein, the expression corresponding to the texture mutation features is:

[0020] (4)

[0021] In equation (4), Indicates texture abrupt change features; This represents the input features of the scene-customized attention module; Indicates localized smooth background features; This represents the absolute value operation element by element.

[0022] Preferably, the serial attention weighted branch module is specifically used for: modeling the channel dependency of the texture mutation feature using formula (5): performing global average pooling on the texture mutation feature to aggregate spatial information, then inputting a multilayer perceptron containing dimensionality reduction convolutional layers and dimensionality increase convolutional layers to generate a channel weight vector, and multiplying the channel weight vector with the texture mutation feature channel by channel to obtain the channel weighted feature; wherein, the expression corresponding to the channel weighted feature is:

[0023] (5)

[0024] In equation (5), Indicates channel weighting characteristics; Indicates texture abrupt change features; This indicates a global average pooling operation; and These are the weight parameters for the dimensionality reduction and dimensionality increase convolutional layers, respectively. and They are respectively Activation function and Activation function;

[0025] Spatial saliency focusing is performed on the channel-weighted features using formula (6): Max pooling and average pooling operations are performed on the channel-weighted features along the channel dimension. The two sets of feature maps are then concatenated and a spatial saliency mask is generated through a two-dimensional convolution operation. The spatial saliency mask is then multiplied element-wise by the channel-weighted features to obtain the weighted saliency features. The expression corresponding to the weighted saliency features is:

[0026] (6)

[0027] In equation (6), This represents the weighted saliency feature; Indicates channel weighting characteristics; and These represent average pooling and max pooling operations, respectively. Indicates feature concatenation operation; Indicates the kernel size as Two-dimensional convolution.

[0028] Preferably, in step S4, the construction of the physically driven motion blur mapping model specifically includes the following process: Constructing a light energy accumulation model in the continuous spatiotemporal domain using formula (7): The light signal in the instantaneously clear scene undergoes relative displacement along the motion velocity vector over time, and within the shutter exposure time of the imaging system acquiring the clear coal flow image, the photon energy at the pixel coordinates of the sensor plane of the imaging system is integrated and accumulated to obtain the blurred image light intensity in the continuous domain; wherein, the expression corresponding to the blurred image light intensity is:

[0029] (7)

[0030] In equation (7), Represents a two-dimensional spatial position coordinate vector The light intensity of the blurred image at that location; Represents the irradiance distribution of a momentarily clear scene; Represents a two-dimensional spatial position coordinate vector; Represents a time variable; Represents the velocity vector; This indicates the shutter exposure time of the imaging system;

[0031] By using formula (8), a two-dimensional convolution operation is performed on the original clear coal flow image in the dataset using the point spread function determined by the physical parameters, so as to superimpose the motion blur effect onto the static image;

[0032] (8)

[0033] In equation (8), This represents the generated discretized motion-blurred image; This represents the original, sharp images in the static benchmark dataset; This represents the two-dimensional convolution operator. A point spread function convolution kernel representing linear motion blur; Indicates the pixel length of the convolution kernel; This indicates the direction angle of motion of the convolution kernel.

[0034] Preferably, in step S5, the process of dynamically blurring the clear coal flow image in the dataset using the motion blur mapping model specifically includes: setting the range of values ​​for the conveyor belt speed and the mechanical vibration disturbance angle using formula (9); and obtaining the conveyor belt speed and vibration disturbance angle of the current iteration random sampling using a uniform distribution sampling function during the data loading stage of the debris detection network training; wherein, the expressions corresponding to the conveyor belt speed and vibration disturbance angle of the current iteration random sampling are respectively:

[0035] (9)

[0036] In equation (9), This indicates the conveyor belt speed during the current iteration of random sampling; This represents the vibration disturbance angle of the current iteration's random sampling; Represents a uniform distribution sampling function; and The lower and upper limits of the operating speed are not specified for the preset conditions under which the conveyor belt produces motion blur. Indicates the maximum mechanical vibration amplitude;

[0037] Using formula (10), the corresponding blur kernel is calculated based on the sampled conveyor belt speed and vibration disturbance angle, and then convolved with the clear coal flow image in the dataset to synthesize a dynamic blur image for the current batch input, so as to guide the debris detection network to learn feature representations with stable responses to different blur conditions; wherein, the expression corresponding to the synthesized dynamic blur image for the current batch input is:

[0038] (10)

[0039] In equation (10), This represents the synthesized, dynamically blurred image used for the current batch of inputs; This represents the original, sharp images in the static benchmark dataset; This represents the two-dimensional convolution operator. A point spread function convolution kernel representing linear motion blur; The function representing the kernel pixel length calculated based on the sampled conveyor belt speed; This indicates the conveyor belt speed during the current iteration of random sampling; Indicates the original angle at which the clutter was placed; This indicates the vibration disturbance angle of the sample.

[0040] Preferably, in step S6, the expression for the loss function is:

[0041] (11)

[0042] In equation (11), This is expressed as the total loss value; , , These represent the balance coefficients for classification, regression, and angle loss, respectively. Represents the classification loss function; This represents the regression loss function for the geometric parameters of the rotated bounding box. This represents the rotation angle regression loss function; , , These represent the category probabilities predicted by the clutter detection network, the bounding box geometric parameters, and the rotation angle, respectively. , , Indicates the corresponding labeled truth value;

[0043] The expression for the rotation angle regression loss function is:

[0044] (12)

[0045] In equation (12), This represents the cosine trigonometric function, and the loss function is configured to be applied when the predicted rotation angle is... with labeled truth value Overlap or difference When the loss is an integer multiple, the loss value of this loss function is 0.

[0046] Preferably, in step S8, the candidate detection results are post-processed, specifically including: using formula (13), treating the rotated bounding box as a polygon, and calculating the overlap between the current highest confidence detection box and the remaining candidate boxes; wherein, the expression corresponding to the overlap is:

[0047] (13)

[0048] In equation (13), This indicates the degree of overlap between the current highest confidence detection box and the other candidate boxes; this degree of overlap is the intersection-union ratio (IU). This represents the baseline box with the highest confidence among the remaining candidate boxes. This represents the candidate box to be filtered. Represents the area of ​​a polygon; Represents the polygon intersection operation;

[0049] Non-maximum suppression logic is executed using formula (14); the expression corresponding to formula (14) is:

[0050] (14)

[0051] In equation (14), This indicates the degree of overlap between the current highest confidence detection box and the other candidate boxes; This indicates the preset overlap rejection threshold;

[0052] If the calculation results of the candidate box and the reference box satisfy the above formula (14), then the candidate box is determined to be a redundant target and the redundant target is removed; otherwise, it is retained.

[0053] Preferably, in step S8, the output includes the detection results containing the object target category, category confidence, rotated bounding box positioning information, and corresponding physical size parameters. Specifically, this includes: converting the detected object pixel geometric parameters into actual physical dimensions using formula (15); using the spatial mapping scale of the imaging system that acquires clear coal flow images, mapping the pixel length and width values ​​of the rotated bounding box to the physical length and width values ​​in the conveyor belt coordinate system, thus completing the physical quantization of the detection data; wherein, the expressions corresponding to the actual physical length and width of the non-coal object target are respectively:

[0054] (15)

[0055] In equation (15), and These represent the actual physical length and width of the non-coal debris target, respectively. and The unit is millimeters; and Represents the longest and shortest pixel side lengths of the rotated bounding box of the non-coal debris target; Represents the physical-pixel space mapping scale;

[0056] The final screening and detection results are constructed using formula (16); the expression corresponding to the final screening and detection results is:

[0057] (16)

[0058] In equation (16), This is the output set corresponding to the final screening and detection results; Indicates the target category label for non-coal debris; This represents the confidence score corresponding to the non-coal debris target category; and The pixel x-coordinate and pixel y-coordinate represent the geometric center; and These represent the actual physical length and width of the non-coal debris target, respectively. and The unit is millimeters; This indicates the rotation angle of the final selected non-coal debris targets.

[0059] The beneficial effects of this invention are:

[0060] This invention provides a coal flow debris rotation detection method based on physical fuzzy modeling and scene-customized attention. By constructing a physically driven motion fuzzy mapping model and dynamically blurring clear coal flow images in the dataset during the training phase, it effectively alleviates the domain distribution difference between clear training data and blurred images under motion fuzzy conditions. This allows the debris detection network to maintain stable detection performance even under complex conditions such as physical trailing caused by conveyor belt operation and mechanical vibration disturbances. In terms of feature representation, a scene-customized attention module is introduced, which can address the problem of coal flow backgrounds of the same color and weak texture targets being easily submerged. It can directionally enhance the local texture abrupt changes and saliency features of debris while suppressing background interference from the coal flow. This reduces the probability of missed and false detections, and improves the robustness of recognition in complex backgrounds. In terms of output, by combining the overlap calculation of the rotated bounding box and the non-maximum suppression logic, overlapping and redundant candidate detection results can be effectively eliminated, improving the uniqueness and usability of the detection results. Furthermore, the pixel-level geometric parameters are quantized into the actual physical dimensions in the conveyor belt coordinate system according to the physical-pixel space mapping scale. The final screening detection results include the non-coal debris target category, the confidence level corresponding to the non-coal debris target category, the rotation bounding box positioning information of the non-coal debris target, and the actual physical dimensions of the non-coal debris target. This improves the engineering application value of the detection results and provides data support for subsequent sorting and grasping operations.

[0061] Additional aspects and advantages of the invention will become apparent from the description which follows, or may be learned by practice of the invention. Attached Figure Description

[0062] Figure 1 This diagram illustrates the overall process of a coal flow debris rotation detection method based on physical fuzzy modeling and scene-customized attention, according to an embodiment of the present invention.

[0063] Figure 2 A schematic diagram of the scene-customized attention module according to an embodiment of the present invention is shown;

[0064] Figure 3 This diagram illustrates the working principle of a physics-driven motion fuzzy mapping model according to an embodiment of the present invention.

[0065] Figure 4 A schematic diagram showing the final screening and detection results of coal flow impurities according to an embodiment of the present invention is illustrated. Detailed Implementation

[0066] To better understand the above-mentioned objects, features, and advantages of the present invention, such as Figures 1 to 4 As shown in the accompanying drawings and specific embodiments, the present invention will be further described in detail below. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0067] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0068] The technical solution of the present invention will be illustrated below with a specific embodiment. For example... Figure 1 As shown, the coal flow debris rotation detection method based on physical fuzzy modeling and scene-customized attention in this specific embodiment is implemented through the following technical solution:

[0069] 1) Step S1: Acquire and annotate the coal flow image. Specifically, acquire a clear coal flow image of the conveyor belt under static or running conditions, and then annotate the non-coal debris targets in the clear coal flow image using a rotated bounding box to generate annotation information corresponding to the non-coal debris targets;

[0070] In step S1, the annotation of non-coal debris targets in the clear coal flow image specifically includes: the category label of the non-coal debris target, the geometric center coordinates of the rotated bounding box, the length of the long side of the rotated bounding box, the length of the short side of the rotated bounding box, and the rotation angle of the rotated bounding box.

[0071] The mathematical expression corresponding to the annotation information is:

[0072] (1)

[0073] In equation (1), A labeled truth vector representing a single non-coal debris target; The pixel coordinates representing the geometric center of the non-coal debris target; This represents the length of the longest side in pixels of the rotated bounding box of the non-coal debris target; This represents the shortest side length in pixels of the rotated bounding box of the non-coal debris target; The angle between the longest side of the rotated bounding box of the non-coal debris target and the horizontal axis of the clear coal flow image is used. To avoid multiple solutions for the angle representation of the same rotated bounding box and to ensure the uniqueness of the training labels, this invention limits the rotation angle to a specific value. Within this range. When the actual angle exceeds this range, the long and short sides of the bounding box are adjusted and the angle is equivalently transformed to remap it back to this range, thus ensuring that each rotated target corresponds to only one unique parameter representation. Therefore, The range of values ​​is ;

[0074] In step S1, clear coal flow images are acquired at the coal mine site using an industrial camera fixed above the conveyor belt. The coal flow area is imaged while the conveyor belt is stationary or running at low speed. Images obtained under these conditions exhibit weak motion blur and can relatively completely preserve the geometric contours and pose information of debris targets, making them suitable as clear benchmark samples for model training.

[0075] For the acquired clear coal flow images, non-coal debris targets are finely labeled using rotated bounding boxes. The rotated bounding box includes at least the category label of the debris target, the geometric center coordinates, the length of the long side, the length of the short side, and the rotation angle information, so as to accurately represent the arbitrary posture and orientation distribution of debris in the coal flow and reduce the redundant background interference introduced by traditional horizontal bounding boxes.

[0076] 2) Step S2: Establish a dataset based on the labeled coal flow images.

[0077] 3) Step S3: Construct a debris detection network comprising a feature extraction module, a feature fusion module, and a prediction output module; wherein, the feature extraction module is used to receive the clear coal flow image and perform convolution processing on the clear coal flow image to extract visual features and output a multi-scale feature map; the feature fusion module is used to fuse deep and shallow features of the multi-scale feature map; the feature fusion module includes a scene-customized attention module, which is used to enhance the features of non-coal debris targets and suppress coal flow background interference, and output the enhanced fused features; the prediction output module is used to predict and output candidate detection results of non-coal debris targets based on the enhanced fused features; the candidate detection results include: the class probability of the non-coal debris target, the geometric parameters of the rotated bounding box, and the rotation angle;

[0078] In step S3, as Figure 2 As shown, the scene-customized attention module includes: a local contrast feature enhancement branch module and a serial attention weighting branch module;

[0079] The local contrast feature enhancement branch module is used to receive features output from the feature extraction module to the feature fusion module as input features of the scene customization attention module; the local contrast feature enhancement branch module is also used to smooth the input features to extract local smooth background features; the local contrast feature enhancement branch module is also used to generate texture mutation features based on the difference between the input features and the local smooth background features.

[0080] The serial attention weighting branch module is used to perform channel dimension weighting and spatial dimension weighting on the texture mutation features to obtain weighted salient features; the serial attention weighting branch module is also used to fuse the weighted salient features with the input features of the scene-customized attention module to output the fused enhanced features;

[0081] The expression corresponding to the enhanced features after fusion is:

[0082] (2)

[0083] In equation (2), This represents the enhanced features output by the serial attention-weighted branch module; This represents the input features of the scene-customized attention module; This represents the weighted saliency feature, which is generated by refining the texture abrupt change feature through a secondary process involving channel and spatial dimensions.

[0084] Furthermore, the local contrast feature enhancement branch module is specifically used for:

[0085] Using formula (3), the input features of the scene-customized attention module are convolved with a kernel size of [size missing]. The depthwise separable convolution operation is then sequentially passed through a batch normalization layer and... The activation function layer generates locally smoothed background features; the expression corresponding to the locally smoothed background features is:

[0086] (3)

[0087] In equation (3), Indicates localized smooth background features; express Activation function; Indicates batch normalization; Indicates the kernel size as Depth-separable convolution; The input features of the scene-customized attention module are represented; the construction of the above-mentioned local smooth background features enables the slowly changing coal flow background information in the original features to be effectively modeled, thereby providing a comparative benchmark for highlighting the texture abrupt change of the debris surface relative to the background through differential operation.

[0088] Using formula (4), the absolute difference between the input features of the scene-customized attention module and the local smooth background features is calculated to generate texture mutation features; wherein, the expression corresponding to the texture mutation features is:

[0089] (4)

[0090] In equation (4), Indicates texture abrupt change features; This represents the input features of the scene-customized attention module; Indicates localized smooth background features; This represents the absolute value operation element by element;

[0091] Since the texture mutation features still contain redundant channel information and non-target region responses, it is necessary to further introduce an attention mechanism to selectively enhance the clutter-related features.

[0092] Furthermore, the serial attention weighted branch module is specifically used for:

[0093] The texture mutation feature is modeled using formula (5) to model channel dependencies: global average pooling is performed on the texture mutation feature to aggregate spatial information, and then a multilayer perceptron containing dimensionality reduction and dimensionality increase convolutional layers is input to generate a channel weight vector. The channel weight vector is then multiplied by the texture mutation feature channel by channel to obtain the channel-weighted feature. The expression corresponding to the channel-weighted feature is:

[0094] (5)

[0095] In equation (5), Indicates channel weighting characteristics; Indicates texture abrupt change features; This indicates a global average pooling operation; and These are the weight parameters for the dimensionality reduction and dimensionality increase convolutional layers, respectively. and They are respectively Activation function and Activation function;

[0096] Spatial saliency focusing is performed on the channel-weighted features using formula (6): Max pooling and average pooling operations are performed on the channel-weighted features along the channel dimension. The two sets of feature maps are then concatenated and a spatial saliency mask is generated through a two-dimensional convolution operation. The spatial saliency mask is then multiplied element-wise by the channel-weighted features to obtain the weighted saliency features. The expression corresponding to the weighted saliency features is:

[0097] (6)

[0098] In equation (6), This represents the weighted saliency feature; Indicates channel weighting characteristics; and These represent average pooling and max pooling operations, respectively. Indicates feature concatenation operation; Indicates the kernel size as Two-dimensional convolution;

[0099] In step S3, the debris detection network adopts an end-to-end deep learning structure, including a feature extraction module for extracting multi-scale semantic information, a feature fusion module for fusing features from different levels, and a prediction output module for outputting debris categories and rotation bounding box parameters. To adapt to the complex background texture and low contrast between debris and coal in coal flow scenarios, a scene-customized attention module is introduced in the feature fusion stage to perform targeted enhancement of significant features related to debris.

[0100] Using formula (2), the saliency features are fused with the input features of the scene-customized attention module using a residual connection method, and the output is an enhanced feature that maintains the stability of the deep feature flow, i.e., the enhanced feature output by the serial attention weighted branch module. .

[0101] 4) Step S4: Based on the operating parameters of the conveyor belt and the imaging system parameters for obtaining the clear coal flow image, construct a physical-driven motion fuzzy mapping model; wherein, the motion fuzzy mapping model is used to describe the mapping relationship between the clear image of the coal flow image and the fuzzy image of the coal flow image;

[0102] Step S4 above mainly enhances the distinguishability of debris in complex coal flow background from the feature expression level. However, under high-speed operating conditions, due to the unavoidable motion blur effect during the imaging process, relying solely on feature enhancement is still difficult to completely eliminate the detection performance degradation caused by data distribution differences. Therefore, it is necessary to further characterize the high-speed operating conditions from the data generation and physical modeling level.

[0103] In step S4, the construction of the physical-driven motion fuzzy mapping model specifically includes the following processes:

[0104] To characterize the motion blur effect of coal flow debris during imaging under high-speed conveyor belt operation conditions, a light energy accumulation model is first established in the continuous spatiotemporal domain. Specifically, the light energy accumulation model in the continuous spatiotemporal domain is constructed using formula (7): the light signal in the instantaneously clear scene undergoes relative displacement along the motion velocity vector over time, and the photon energy at the pixel coordinates of the sensor plane of the imaging system is integrated and accumulated within the shutter exposure time of the imaging system acquiring the clear coal flow image to obtain the blurred image light intensity in the continuous domain; wherein, the expression corresponding to the blurred image light intensity is:

[0105] (7)

[0106] In equation (7), Represents a two-dimensional spatial position coordinate vector The light intensity of the blurred image at that location; Represents the irradiance distribution of a momentarily clear scene; Represents a two-dimensional spatial position coordinate vector; Represents a time variable; Represents the velocity vector; This indicates the shutter exposure time of the imaging system;

[0107] By using formula (8), a two-dimensional convolution operation is performed on the original clear coal flow image in the dataset using the point spread function determined by the physical parameters, so as to superimpose the motion blur effect onto the static image;

[0108] (8)

[0109] In equation (8), This represents the generated discretized motion-blurred image; This represents the original, sharp images in the static benchmark dataset; This represents the two-dimensional convolution operator. A point spread function convolution kernel representing linear motion blur; Indicates the pixel length of the convolution kernel; Indicates the direction angle of motion of the convolution kernel;

[0110] By constructing the aforementioned physical-driven motion blur mapping model, step S4 achieves a controllable transformation from a static, clear coal flow image to a high-speed motion blur image, providing a physically consistent data foundation for network training and parameter optimization based on dynamic blur samples in subsequent steps.

[0111] like Figure 3 As shown, the working principle of the physics-driven motion blur mapping model is as follows: taking a clear coal flow image as input, under the action of the conveyor belt speed, the image is displaced and light energy is accumulated during the exposure time, forming a blurred imaging process; further, according to the motion parameters, a corresponding point spread function is constructed, and convolution operation is performed on the clear image to realize the mapping relationship between the clear image and the blurred image.

[0112] 5) Step S5: Generate training samples for simulating motion blur conditions. Specifically, the motion blur mapping model is used to perform dynamic blurring processing on the clear coal flow images in the dataset to generate training samples simulating motion blur phenomena caused by the conveyor belt;

[0113] In step S5, the process of dynamically blurring the clear coal flow images in the dataset using the motion blur mapping model specifically includes:

[0114] To enable the detection network to fully adapt to the diverse motion fuzzy features generated under different high-speed operating conditions, an online physical fuzzy adaptation strategy is introduced during the training phase. Specifically, using formula (9), the range of values ​​for the conveyor belt speed and the mechanical vibration disturbance angle is set. During the data loading phase of the debris detection network training, a uniform distribution sampling function is used to obtain the conveyor belt speed and vibration disturbance angle randomly sampled in the current iteration. The expressions corresponding to the conveyor belt speed and vibration disturbance angle randomly sampled in the current iteration are as follows:

[0115] (9)

[0116] In equation (9), This indicates the conveyor belt speed during the current iteration of random sampling; This represents the vibration disturbance angle of the current iteration's random sampling; Represents a uniform distribution sampling function; and The lower and upper limits of the operating speed are not specified for the preset conditions under which the conveyor belt produces motion blur. Indicates the maximum mechanical vibration amplitude;

[0117] Using formula (10), the corresponding blur kernel is calculated based on the sampled conveyor belt speed and vibration disturbance angle, and then convolved with the clear coal flow image in the dataset to synthesize a dynamic blur image for the current batch input, so as to guide the debris detection network to learn feature representations with stable responses to different blur conditions; wherein, the expression corresponding to the synthesized dynamic blur image for the current batch input is:

[0118] (10)

[0119] In equation (10), This represents the synthesized, dynamically blurred image used for the current batch of inputs; This represents the original, sharp images in the static benchmark dataset; This represents the two-dimensional convolution operator. A point spread function convolution kernel representing linear motion blur; The function representing the kernel pixel length calculated based on the sampled conveyor belt speed; This indicates the conveyor belt speed during the current iteration of random sampling; Indicates the original angle at which the clutter was placed; Indicates the sampling vibration disturbance angle;

[0120] The above-mentioned continuous introduction of dynamic blurred samples during the training phase enables the detection network to simultaneously face input data with different blur intensities and directions during parameter updates, providing a representative training sample basis for subsequent classification, localization, and angle prediction constrained by a multi-task joint loss function.

[0121] 6) Step S6: Based on the training samples and loss function, train and optimize the clutter detection network;

[0122] In step S6, the expression for the loss function is:

[0123] (11)

[0124] In equation (11), This is expressed as the total loss value; , , These represent the balance coefficients for classification, regression, and angle loss, respectively. Represents the classification loss function; This represents the regression loss function for the geometric parameters of the rotated bounding box. This represents the rotation angle regression loss function; , , These represent the category probabilities predicted by the clutter detection network, the bounding box geometric parameters, and the rotation angle, respectively. , , Indicates the corresponding labeled truth value;

[0125] The expression for the rotation angle regression loss function is:

[0126] (12)

[0127] In equation (12), This represents the cosine trigonometric function, and the loss function is configured to be applied when the predicted rotation angle is... with labeled truth value Overlap or difference When the loss is an integer multiple of the integer, the loss value of this loss function is 0;

[0128] In step S6, a multi-task joint loss function including category classification, bounding box regression, and angle prediction is constructed using formula (11). The network prediction output is compared with the ground truth label, and the classification loss, regression loss, and angle loss are weighted and summed to calculate the total loss value. To drive backpropagation; calculate the rotation angle regression loss considering the periodicity of the angle using formula (12). To address the periodicity of the rotation angle definition, the periodicity of the cosine function is used to calculate the deviation between the predicted angle and the true angle.

[0129] 7) Step S7: Detect the input real-time coal flow image. Specifically, acquire real-time coal flow images of the conveyor belt under static or running conditions, input the real-time coal flow images into the optimized debris detection network, and obtain candidate detection results for non-coal debris targets.

[0130] 8) Step S8: Perform post-processing and output the final detection result. Specifically, perform post-processing on the candidate detection results to filter out overlapping and redundant detection results, and output the final screening detection result containing the non-coal debris target category, the confidence level corresponding to the non-coal debris target category, the non-coal debris target rotation bounding box positioning information, and the actual physical size of the non-coal debris target;

[0131] Through the above-mentioned dynamic training process based on physical fuzzy modeling, the detection network has acquired the ability to adapt to motion fuzzy features under high-speed operating conditions, thus enabling stable detection of real-time coal flow images during the online inference stage.

[0132] In this embodiment, when the conveyor belt is operating at high speed, an industrial camera captures real-time images of the coal flow. These images are then input into a trained debris detection network for forward inference. The network outputs candidate detection results containing the probability of the debris target category, the geometric parameters of the rotated bounding box, and the rotation angle. Since multiple debris targets may exist simultaneously in a high-speed coal flow scenario, and the network output may contain overlapping or redundant rotated detection boxes, further filtering of the candidate detection results is necessary.

[0133] Further, in step S8, the candidate detection results are post-processed, specifically including:

[0134] Using formula (13), the rotated bounding box is treated as a polygon, and the overlap between the current highest confidence detection box and the other candidate boxes is calculated; wherein, the expression corresponding to the overlap is:

[0135] (13)

[0136] In equation (13), This indicates the degree of overlap between the current highest confidence detection box and the other candidate boxes; this degree of overlap is the intersection-union ratio (IU). This represents the baseline box with the highest confidence among the remaining candidate boxes. This represents the candidate box to be filtered. Represents the area of ​​a polygon; Represents the polygon intersection operation;

[0137] Non-maximum suppression logic is executed using formula (14); the expression corresponding to formula (14) is:

[0138] (14)

[0139] In equation (14), This indicates the degree of overlap between the current highest confidence detection box and the other candidate boxes; This indicates the preset overlap rejection threshold;

[0140] If the calculation results of the candidate box and the reference box satisfy the above formula (14), then the candidate box is determined to be a redundant target and the redundant target is removed; otherwise, it is retained.

[0141] In step S8, the detection results, including the object category, category confidence, rotated bounding box positioning information, and corresponding physical size parameters, are output, specifically including:

[0142] Formula (15) is used to convert the geometric parameters of the detected debris pixels into actual physical dimensions. Using the spatial mapping scale of the imaging system that acquires clear coal flow images, the pixel length and width values ​​of the rotated bounding box are mapped to the physical length and width values ​​in the conveyor belt coordinate system, thus completing the physical quantification of the detection data. The expressions for the actual physical length and width of the non-coal debris targets are as follows:

[0143] (15)

[0144] In equation (15), and These represent the actual physical length and width of the non-coal debris target, respectively. and The unit is millimeters; and Represents the longest and shortest pixel side lengths of the rotated bounding box of the non-coal debris target; Represents the physical-pixel space mapping scale;

[0145] The final screening and detection results are constructed using formula (16); the expression corresponding to the final screening and detection results is:

[0146] (16)

[0147] In equation (16), This is the output set corresponding to the final screening and detection results; Indicates the target category label for non-coal debris; This represents the confidence score corresponding to the non-coal debris target category; and The pixel x-coordinate and pixel y-coordinate represent the geometric center; and These represent the actual physical length and width of the non-coal debris target, respectively. and The unit is millimeters; This indicates the rotation angle of the final selected non-coal debris targets.

[0148] The final screening and testing results of coal flow debris are as follows: Figure 4 As shown. Figure 4 The image shows the final screening and detection results of the method for identifying debris targets in coal flow, illustrating that the method can achieve debris identification and rotational bounding box localization.

[0149] In this specific implementation, the coal flow debris rotation detection method based on physical fuzzy modeling and scene-customized attention utilizes a physically driven motion fuzzy mapping model to dynamically blur clear coal flow images in the dataset, enabling the debris detection network to learn and adapt feature representations under different fuzzy conditions. Combined with a scene-customized attention module, it enhances the features of non-coal debris targets and suppresses background interference from the coal flow, thereby improving the representational ability of debris features. By rotating the bounding box to model the geometric shape and posture of the debris, and combining a multi-task learning mechanism to achieve category recognition and position regression, the method outputs detection results including rotation angle and spatial position. This method can adapt to motion fuzziness and background interference of the same color generated during conveyor belt operation, improving the accuracy and stability of coal flow debris detection in complex scenes.

[0150] In summary, the coal flow debris rotation detection method based on physical fuzzy modeling and scene-customized attention provided by this invention has unparalleled technical advantages.

[0151] 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 rotating debris in coal flow based on physical fuzzy modeling and scene-customized attention, characterized in that, include: Step S1: Obtain a clear coal flow image of the conveyor belt under static or running conditions, and then use a rotated bounding box to annotate the non-coal debris targets in the clear coal flow image to generate annotation information corresponding to the non-coal debris targets; Step S2: Establish a dataset based on the labeled coal flow images; Step S3: Construct a debris detection network comprising a feature extraction module, a feature fusion module, and a prediction output module; wherein, the feature extraction module is used to receive the clear coal flow image and perform convolution processing on the clear coal flow image to extract visual features and output a multi-scale feature map; the feature fusion module is used to fuse deep and shallow features of the multi-scale feature map; the feature fusion module includes a scene-customized attention module, which is used to enhance the features of non-coal debris targets and suppress coal flow background interference, and output the enhanced fused features; the prediction output module is used to predict and output candidate detection results of non-coal debris targets based on the enhanced fused features; the candidate detection results include: the class probability of the non-coal debris target, the geometric parameters of the rotated bounding box, and the rotation angle; Step S4: Based on the operating parameters of the conveyor belt and the imaging system parameters for obtaining the clear coal flow image, construct a physical-driven motion fuzzy mapping model; wherein, the motion fuzzy mapping model is used to describe the mapping relationship between the clear image of the coal flow image and the fuzzy image of the coal flow image; Step S5: Use the motion blur mapping model to perform dynamic blur processing on the clear coal flow images in the dataset to generate training samples for simulating the motion blur phenomenon of the conveyor belt. Step S6: Based on the training samples and loss function, train and optimize the debris detection network; Step S7: Obtain real-time coal flow images of the conveyor belt under static or running conditions, and input the real-time coal flow images into the optimized debris detection network to obtain candidate detection results of non-coal debris targets; Step S8: Post-process the candidate detection results to remove overlapping and redundant detection results, and output the final screening detection results, which include the non-coal debris target category, the confidence level corresponding to the non-coal debris target category, the non-coal debris target rotation bounding box positioning information, and the actual physical size of the non-coal debris target.

2. The method for detecting rotating debris in coal flow based on physical fuzzy modeling and scene-customized attention as described in claim 1, characterized in that, In step S1, the annotation of non-coal debris targets in the clear coal flow image specifically includes: the category label of the non-coal debris target, the geometric center coordinates of the rotated bounding box, the length of the long side of the rotated bounding box, the length of the short side of the rotated bounding box, and the rotation angle of the rotated bounding box. The mathematical expression corresponding to the annotation information is: （1） In equation (1), A labeled truth vector representing a single non-coal debris target; The pixel coordinates representing the geometric center of the non-coal debris target; This represents the length of the longest side in pixels of the rotated bounding box of the non-coal debris target; This represents the shortest side length in pixels of the rotated bounding box of the non-coal debris target; This represents the angle between the longest side of the rotated bounding box of non-coal debris targets and the horizontal axis of the clear coal flow image. The range of values ​​is .

3. The method for detecting rotating debris in coal flow based on physical fuzzy modeling and scene-customized attention as described in claim 2, characterized in that, In step S3, the scene-customized attention module includes: a local contrast feature enhancement branch module and a serial attention weighting branch module; The local contrast feature enhancement branch module is used to receive features output from the feature extraction module to the feature fusion module as input features of the scene customization attention module; the local contrast feature enhancement branch module is also used to smooth the input features to extract local smooth background features; the local contrast feature enhancement branch module is also used to generate texture mutation features based on the difference between the input features and the local smooth background features. The serial attention weighting branch module is used to perform channel dimension weighting and spatial dimension weighting on the texture mutation features to obtain weighted salient features; the serial attention weighting branch module is also used to fuse the weighted salient features with the input features of the scene-customized attention module to output the fused enhanced features; The expression corresponding to the enhanced features after fusion is: （2） In equation (2), This represents the enhanced features output by the serial attention-weighted branch module; This represents the input features of the scene-customized attention module; This represents the significance of the weighted feature.

4. The method for detecting rotating debris in coal flow based on physical fuzzy modeling and scene-customized attention as described in claim 3, characterized in that, The local contrast feature enhancement branch module is specifically used for: Using formula (3), the input features of the scene-customized attention module are convolved with a kernel size of [size missing]. The depthwise separable convolution operation is then sequentially passed through a batch normalization layer and... The activation function layer generates locally smoothed background features; the expression corresponding to the locally smoothed background features is: （3） In equation (3), Indicates localized smooth background features; express Activation function; Indicates batch normalization; Indicates the kernel size as Depth-separable convolution; This represents the input features of the scene-customized attention module; Using formula (4), the absolute difference between the input features of the scene-customized attention module and the local smooth background features is calculated to generate texture mutation features; wherein, the expression corresponding to the texture mutation features is: （4） In equation (4), Indicates texture abrupt change features; This represents the input features of the scene-customized attention module; Indicates localized smooth background features; This represents the absolute value operation element by element.

5. The method for detecting rotating debris in coal flow based on physical fuzzy modeling and scene-customized attention as described in claim 3, characterized in that, The serial attention weighted branch module is specifically used for: The texture mutation feature is modeled using formula (5) to model channel dependencies: global average pooling is performed on the texture mutation feature to aggregate spatial information, and then a multilayer perceptron containing dimensionality reduction and dimensionality increase convolutional layers is input to generate a channel weight vector. The channel weight vector is then multiplied by the texture mutation feature channel by channel to obtain the channel-weighted feature. The expression corresponding to the channel-weighted feature is: （5） In equation (5), Indicates channel weighting characteristics; Indicates texture abrupt change features; This indicates a global average pooling operation; and These are the weight parameters for the dimensionality reduction and dimensionality increase convolutional layers, respectively. and They are respectively Activation function and Activation function; Spatial saliency focusing is performed on the channel-weighted features using formula (6): Max pooling and average pooling operations are performed on the channel-weighted features along the channel dimension. The two sets of feature maps are then concatenated and a spatial saliency mask is generated through a two-dimensional convolution operation. The spatial saliency mask is then multiplied element-wise by the channel-weighted features to obtain the weighted saliency features. The expression corresponding to the weighted saliency features is: （6） In equation (6), This represents the weighted saliency feature; Indicates channel weighting characteristics; and These represent average pooling and max pooling operations, respectively. Indicates feature concatenation operation; Indicates the kernel size as Two-dimensional convolution.

6. The method for detecting rotating debris in coal flow based on physical fuzzy modeling and scene-customized attention as described in claim 1, characterized in that, In step S4, the construction of the physical-driven motion fuzzy mapping model is specifically... The process includes the following: Using formula (7), a light energy accumulation model in the continuous spatiotemporal domain is constructed: the light signal in the instantaneously clear scene undergoes relative displacement along the motion velocity vector over time, and within the shutter exposure time of the imaging system acquiring the clear coal flow image, the photon energy at the pixel coordinates of the sensor plane of the imaging system is integrated and accumulated to obtain the light intensity of the blurred image in the continuous domain; wherein, the expression corresponding to the light intensity of the blurred image is: （7） In equation (7), Represents a two-dimensional spatial position coordinate vector The light intensity of the blurred image at that location; Represents the irradiance distribution of a momentarily clear scene; Represents a two-dimensional spatial position coordinate vector; Represents a time variable; Represents the velocity vector; This indicates the shutter exposure time of the imaging system; By using formula (8), a two-dimensional convolution operation is performed on the original clear coal flow image in the dataset using the point spread function determined by the physical parameters, so as to superimpose the motion blur effect onto the static image; （8） In equation (8), This represents the generated discretized motion-blurred image; This represents the original, sharp images in the static benchmark dataset; This represents the two-dimensional convolution operator. A point spread function convolution kernel representing linear motion blur; Indicates the pixel length of the convolution kernel; This indicates the direction angle of motion of the convolution kernel.

7. The method for detecting rotating debris in coal flow based on physical fuzzy modeling and scene-customized attention as described in claim 1, characterized in that, In step S5, the process of dynamically blurring the clear coal flow images in the dataset using the motion blur mapping model specifically includes: Using formula (9), the range of values ​​for conveyor belt speed and mechanical vibration disturbance angle is set. During the data loading stage of the debris detection network training, a uniform distribution sampling function is used to obtain the conveyor belt speed and vibration disturbance angle of the current iteration random sampling. The expressions corresponding to the conveyor belt speed and vibration disturbance angle of the current iteration random sampling are as follows: （9） In equation (9), This indicates the conveyor belt speed during the current iteration of random sampling; This represents the vibration disturbance angle of the current iteration's random sampling; Represents a uniform distribution sampling function; and The lower and upper limits of the operating speed are not specified for the preset conditions under which the conveyor belt produces motion blur. Indicates the maximum mechanical vibration amplitude; Using formula (10), the corresponding blur kernel is calculated based on the sampled conveyor belt speed and vibration disturbance angle, and then convolved with the clear coal flow image in the dataset to synthesize a dynamic blur image for the current batch input, so as to guide the debris detection network to learn feature representations with stable responses to different blur conditions; wherein, the expression corresponding to the synthesized dynamic blur image for the current batch input is: （10） In equation (10), This represents the synthesized, dynamically blurred image used for the current batch of inputs; This represents the original, sharp images in the static benchmark dataset; This represents the two-dimensional convolution operator. A point spread function convolution kernel representing linear motion blur; The function representing the kernel pixel length calculated based on the sampled conveyor belt speed; This indicates the conveyor belt speed during the current iteration of random sampling; Indicates the original angle at which the clutter was placed; This indicates the vibration disturbance angle of the sample.

8. The method for detecting rotating debris in coal flow based on physical fuzzy modeling and scene-customized attention as described in claim 1, characterized in that, In step S6, the expression for the loss function is: （11） In equation (11), This is expressed as the total loss value; , , These represent the balance coefficients for classification, regression, and angle loss, respectively. Represents the classification loss function; This represents the regression loss function for the geometric parameters of the rotated bounding box. This represents the rotation angle regression loss function; , , These represent the category probabilities predicted by the clutter detection network, the bounding box geometric parameters, and the rotation angle, respectively. , , Indicates the corresponding labeled truth value; The expression for the rotation angle regression loss function is: （12） In equation (12), This represents the cosine trigonometric function, and the loss function is configured to be applied when the predicted rotation angle is... with labeled truth value Overlap or difference When the loss is an integer multiple, the loss value of this loss function is 0.

9. The method for detecting rotating debris in coal flow based on physical fuzzy modeling and scene-customized attention as described in claim 1, characterized in that, In step S8, the candidate detection results are post-processed, specifically including: Using formula (13), the rotated bounding box is treated as a polygon, and the overlap between the current highest confidence detection box and the other candidate boxes is calculated; wherein, the expression corresponding to the overlap is: （13） In equation (13), This indicates the degree of overlap between the current highest confidence detection box and the other candidate boxes; this degree of overlap is the intersection-union ratio (IU). This represents the baseline box with the highest confidence among the remaining candidate boxes. This represents the candidate box to be filtered. Represents the area of ​​a polygon; Represents the polygon intersection operation; Non-maximum suppression logic is executed using formula (14); the expression corresponding to formula (14) is: （14） In equation (14), This indicates the degree of overlap between the current highest confidence detection box and the other candidate boxes; This indicates the preset overlap rejection threshold; If the calculation results of the candidate box and the reference box satisfy the above formula (14), then the candidate box is determined to be a redundant target and the redundant target is removed; otherwise, it is retained.

10. The method for detecting rotating debris in coal flow based on physical fuzzy modeling and scene-customized attention according to claim 1, characterized in that, In step S8, the detection results, including the object category, category confidence, rotated bounding box positioning information, and corresponding physical size parameters, are output, specifically including: Formula (15) is used to convert the geometric parameters of the detected debris pixels into actual physical dimensions. Using the spatial mapping scale of the imaging system that acquires clear coal flow images, the pixel length and width values ​​of the rotated bounding box are mapped to the physical length and width values ​​in the conveyor belt coordinate system, thus completing the physical quantification of the detection data. The expressions for the actual physical length and width of the non-coal debris targets are as follows: （15） In equation (15), and These represent the actual physical length and width of the non-coal debris target, respectively. and The unit is millimeters; and Represents the longest and shortest pixel side lengths of the rotated bounding box of the non-coal debris target; Represents the physical-pixel space mapping scale; The final screening and detection results are constructed using formula (16); the expression corresponding to the final screening and detection results is: （16） In equation (16), This is the output set corresponding to the final screening and detection results; Indicates the target category label for non-coal debris; This represents the confidence score corresponding to the non-coal debris target category; and The pixel x-coordinate and pixel y-coordinate represent the geometric center; and These represent the actual physical length and width of the non-coal debris target, respectively. and The unit is millimeters; This indicates the rotation angle of the final selected non-coal debris targets.

Images