Target detection method in mining belt conveying system

Abstract

The invention discloses a target detection method in a mining belt type conveying system, and the specific process comprises the steps: 1, obtaining an image of a coal mine conveyed by a conveyor, carrying out the manual foreign matter marking, obtaining a data set, dividing the data set into a training set, a verification set and a test set, and guaranteeing the uniform distribution of data in each subset; 2, replacing an original first convolutional layer at the beginning part of a backbone network of the YOLOV8, and adding an improved IAT network as a preprocessing module; 3, replacing an original YOLOV8 detection head with a dynamic detection head; 4, setting a loss function; and 5, training the improved YOLOV8 model by adopting the training set, and performing test verification on the trained improved YOLOV8 model by using the verification set and the test set. According to the target detection method in the mining belt type conveying system, an original YOLOV8 detection head is replaced with a dynamic detection head, and the multi-scale target detection and feature expression capability is remarkably improved.

Description

Technical Field

[0001] The invention belongs to the technical field of computer image processing, and in particular relates to a target detection method in a mining belt conveyor system. Background Art

[0002] China is the world's largest coal producer and consumer. Rich in coal, poor in oil and little in gas are the basic national conditions of our country's energy. In the field of coal mine transportation, coal conveyor belts are one of the important equipment in coal mine production, responsible for transporting mined coal from mines to related processing facilities. However, due to the complexity of the operating environment of coal mine conveyor belts, they often face the problem of foreign matter mixing. The presence of these foreign objects not only causes damage to coal mine equipment, such as abnormal belt sound, belt deviation, conveyor slippage, belt breakage, etc., but may also cause production accidents and pose a threat to the safety of miners and equipment. When detecting foreign objects in coal mine conveyors, due to the harsh underground environment, insufficient lighting conditions, and severe noise in coal mines, the image quality obtained is poor, which seriously affects the accuracy of target detection. Summary of the invention

[0003] The purpose of the present invention is to provide a target detection method in a mining belt conveyor system, which solves the problem of poor accuracy of target detection in existing coal mine conveyors.

[0004] The technical solution adopted by the present invention is a target detection method in a mining belt conveyor system, which is specifically implemented according to the following steps: Step 1: Obtain images of coal transported by conveyors, manually annotate foreign objects to obtain a data set, and divide the data set into a training set, a validation set, and a test set to ensure that the data in each subset is evenly distributed; Step 2: At the beginning of the backbone network of YOLOV8, replace the original first convolutional layer and add the improved IAT network as a preprocessing module; Step 3: Replace the original YOLOV8 detection head with the dynamic detection head to obtain the improved YOLOV8 model; Step 4: Set the loss function; Step 5: Use the training set to train the improved YOLOV8 model, and use the validation set and test set to test and verify the trained improved YOLOV8 model. The present invention is also characterized in that: The specific process of step 1 is: The images of conveyors transporting coal mines are obtained, and the foreign objects are manually annotated to obtain the data set, which is divided into training set, validation set and test set. The training set accounts for 70% of the total set, the validation set accounts for 20% of the total set, and the test set accounts for 10% of the total set.

[0005] The specific process of adding the improved IAT network as a preprocessing module in step 2 is: The existing IAT network model is used, which is divided into local branches and global branches. The pixel-level enhancement module of the local branch is replaced by the SE module. In the IAT network model, the input feature map is X, and the size is , where H and W represent the height and width of the feature map respectively, and C represents the number of channels; For local branches, convolution operations are used to extract local features: (1); After local feature extraction, the self-attention mechanism is used to enhance feature representation: (2); For the global branch, it is implemented through global pooling operation and fully connected layer, and the global features are extracted by global pooling operation: (3); The global features are further processed through the fully connected layer: (4); The SE module is introduced, and the steps of the SE module include compression and excitation; Perform global average pooling on the input feature map to compress the spatial information of each channel, where a and b are the indexes of the spatial dimensions of the input feature map: a represents the index of the height dimension of the feature map, from 1 to H, where H is the height of the feature map; b represents the index of the width dimension of the feature map, from 1 to W, where W is the width of the feature map: (5); The compression operation obtains a vector 𝑧 with a channel number of 𝐶; The compressed features are recalibrated through two fully connected layers: (6); in, and are the weight matrices of the two fully connected layers, 𝛿 represents the ReLU activation function, and 𝜎 represents the Sigmoid activation function; Multiply the excited feature 𝑠 by the input feature map 𝑋 channel by channel: (7); Among them, ⊙ represents the channel-by-channel multiplication operation; Merge the features of local branches and global branches: (8); The final output feature map is the fused feature Y.

[0006] The specific process of step 3 is: Step 3.1: Replace the three detection heads of YOLOV8 with the improved dynamic detection head, and change the deformable convolution in the dynamic detection head from DCNv2 to DCNv; extract features of different scales from Backbone, and adjust the scale by upsampling and downsampling. After adjustment, pass it into the detection head, and the input feature is a four-dimensional tensor ,in ; L, S, C represent scale, space and channel respectively; Step 3.2: Constructing a scale attention mechanism , dynamically fuse features of different scales according to semantic importance, first use a 1*1 convolutional layer to adaptively learn the attention weights of each scale, then introduce an adaptive sigmoid activation function to assign weights to features of each scale, and finally apply the normalized attention weight vector to the multi-scale feature map and generate the corresponding output results. The specific formula is as follows: (9); Function f( ) is the linear function of the convolutional layer, is the adaptive sigmoid activation function, and its expression is , by introducing additional parameters a and b for adaptive adjustment, it is suitable for dynamically adjusting weights according to input features; Step 3.3: Spatial-aware attention fuses the multi-layer features after the previous step, learns the offset of the feature map through deformable convolution, and then adaptively aggregates multiple levels of features to learn more discriminative representations; the specific expression is as follows: (10); Where k represents the number of sparse sampling spatial locations, Represents the feature map after deformable convolution, which can focus on a specific spatial location more flexibly while taking into account the dynamically learned offset. and They represent weights and dynamically learned feature representations, which can be dynamically adjusted according to specific inputs to improve the effectiveness of features. is the spatial offset learned by the network, Indicates the offset; Step 3.4: Combine multiple detection tasks by dynamically controlling feature channels. (11); in Represents the feature map The specific feature channel extracted from , , , Represents a vector that controls the activation threshold. By taking the maximum value, the model can select the feature channel combination that performs best in the current task; Step 3.5: Finally, we introduce scale-aware attention, spatial position-aware attention, and task-aware attention into the dynamic detection head. The three-dimensional feature tensor at the detection layer is F∈R L×S×C , the attention function calculation formula is as follows, where L is the level, S is the space, and C is the channel: (12); pass , , The modules are stacked one after another, so that the detection head can better integrate contextual information and enhance feature extraction.

[0007] The specific process of step 4 is: define a multi-task loss function, which includes the following three parts: bounding box regression loss, category classification loss and confidence loss; Step 4.1: In the bounding box regression loss, L1 loss or L2 loss is used to measure the difference between the predicted bounding box and the true bounding box. The L1 loss formula is as follows: (13); in, is the i-th predicted bounding box, is the i-th ground-truth bounding box, N is the total number of objects, represents the L1 norm; Step 4.2: The formula for class classification loss is as follows: (14); is the true category label of the i-th target, is the probability that the i-th target is predicted to be category c, and M is the total number of categories; Step 4.3: The confidence loss formula is as follows: (15); in, is the true confidence label of the i-th target, is the confidence prediction of the i-th target; Step 4.4: The total loss function is the weighted sum of the above three losses, and its formula is as follows: (16); in, , and is a hyperparameter that controls the weight of each loss term.

[0008] The specific contents of step 5 are: The training set is used to train the improved YOLOV8 model, and then the validation set and test set are used to evaluate the trained improved YOLOV8 model. The compiler is pycharm, and the learning framework of Python3.9.7+torch1.12+cu113 is used to train and detect the model. The initial learning rate is 0.01, the number of iterations is 300, and the evaluation indicators are recall rate R, precision rate P, average precision rate AP, average precision rate mAP and transmission frame rate per second FPS. The calculation formula is as follows: (17); (18); (19); (20); Among them, TP is a true positive example, FP is a false positive example, TN is a true negative example, and FN is a false negative example. Finally, the model that passes the evaluation is deployed in practical applications for reasoning on foreign body target detection tasks in mining belt conveyors.

[0009] The beneficial effects of the present invention are: By replacing the original YOLOV8 detection head with a dynamic detection head, the multi-scale target detection and feature expression capabilities have been significantly improved. The dynamic detection head introduces a dynamic convolution mechanism that can adaptively adjust the convolution kernel parameters according to the input features, improving the flexibility and expression of feature extraction. Compared with the traditional fixed convolution kernel detection head, the dynamic detection head can better capture target features in different scales and complex scenes, and enhance the detection accuracy of small and large targets. In addition, the feature fusion strategy of the dynamic detection head enables the model to make more full use of multi-level feature information, improving the richness and accuracy of feature representation. DETAILED DESCRIPTION

[0010] The present invention is described in detail below in conjunction with specific implementation modes.

[0011] The target detection method in the mining belt conveyor system is implemented according to the following steps: Step 1: Obtain images of coal transported by conveyors, manually annotate foreign objects to obtain a data set, and divide the data set into a training set, a validation set, and a test set, where the training set accounts for 70% of the total set, the validation set accounts for 20% of the total set, and the test set accounts for 10% of the total set.

[0012] Step 2: At the beginning of the backbone network of YOLOV8, replace the original first convolutional layer and add the improved IAT network as a preprocessing module; the specific process of adding the improved IAT network as a preprocessing module is as follows: The existing IAT network model is used, which is divided into local branches and global branches. The pixel-level enhancement module of the local branch is replaced by the SE module. In the IAT network model, the input feature map is X, and the size is , where H and W represent the height and width of the feature map respectively, and C represents the number of channels; For local branches, convolution operations are used to extract local features: (1); After local feature extraction, the self-attention mechanism is used to enhance feature representation: (2); For the global branch, it is implemented through global pooling operation and fully connected layer, and the global features are extracted by global pooling operation: (3); The global features are further processed through the fully connected layer: (4); The SE module is introduced, and the steps of the SE module include compression and excitation; Perform global average pooling on the input feature map to compress the spatial information of each channel, where a and b are the indexes of the spatial dimensions of the input feature map: a represents the index of the height dimension of the feature map, from 1 to H, where H is the height of the feature map; b represents the index of the width dimension of the feature map, from 1 to W, where W is the width of the feature map: (5); The compression operation obtains a vector 𝑧 with a channel number of 𝐶; The compressed features are recalibrated through two fully connected layers: (6); in, and are the weight matrices of the two fully connected layers, 𝛿 represents the ReLU activation function, and 𝜎 represents the Sigmoid activation function; Multiply the excited feature 𝑠 by the input feature map 𝑋 channel by channel: (7); Among them, ⊙ represents the channel-by-channel multiplication operation; Merge the features of local branches and global branches: (8); The final output feature map is the fused feature Y.

[0013] Step 3: Replace the original YOLOV8 detection head with the dynamic detection head to obtain the improved YOLOV8 model; the specific process is: Step 3.1: Replace the three detection heads of YOLOV8 with the improved dynamic detection head, and change the deformable convolution in the dynamic detection head from DCNv2 to DCNv; extract features of different scales from Backbone, and adjust the scale by upsampling and downsampling. After adjustment, pass it into the detection head, and the input feature is a four-dimensional tensor ,in ; L, S, C represent scale, space and channel respectively; Step 3.2: Constructing a scale attention mechanism , dynamically fuse features of different scales according to semantic importance, first use a 1*1 convolutional layer to adaptively learn the attention weights of each scale, then introduce an adaptive sigmoid activation function to assign weights to features of each scale, and finally apply the normalized attention weight vector to the multi-scale feature map and generate the corresponding output results. The specific formula is as follows: (9); Function f( ) is the linear function of the convolutional layer, is the adaptive sigmoid activation function, and its expression is , by introducing additional parameters a and b for adaptive adjustment, it is suitable for dynamically adjusting weights according to input features; Step 3.3: Spatial-aware attention fuses the multi-layer features after the previous step, learns the offset of the feature map through deformable convolution, and then adaptively aggregates multiple levels of features to learn more discriminative representations; the specific expression is as follows: (10); Where k represents the number of sparse sampling spatial locations, Represents the feature map after deformable convolution, which can focus on a specific spatial location more flexibly while taking into account the dynamically learned offset. and They represent weights and dynamically learned feature representations, which can be dynamically adjusted according to specific inputs to improve the effectiveness of features. is the spatial offset learned by the network, Indicates the offset; Step 3.4: Combine multiple detection tasks by dynamically controlling feature channels. (11); in Represents the feature map The specific feature channel extracted from , , , Represents a vector that controls the activation threshold. By taking the maximum value, the model can select the feature channel combination that performs best in the current task; Step 3.5: Finally, we introduce scale-aware attention, spatial position-aware attention, and task-aware attention into the dynamic detection head. The three-dimensional feature tensor at the detection layer is F∈R L×S×C , the attention function calculation formula is as follows, where L is the level, S is the space, and C is the channel: (12); pass , , The modules are stacked one after another, so that the detection head can better integrate contextual information and enhance feature extraction.

[0014] Step 4: Set the loss function; the specific process is: define a multi-task loss function, which includes the following three parts: bounding box regression loss, category classification loss and confidence loss; Step 4.1: In the bounding box regression loss, L1 loss or L2 loss is used to measure the difference between the predicted bounding box and the true bounding box. The L1 loss formula is as follows: (13); in, is the i-th predicted bounding box, is the i-th ground-truth bounding box, N is the total number of objects, represents the L1 norm; Step 4.2: The formula for class classification loss is as follows: (14); is the true category label of the i-th target, is the probability that the i-th target is predicted to be category c, and M is the total number of categories; Step 4.3: The confidence loss formula is as follows: (15); in, is the true confidence label of the i-th target, is the confidence prediction of the i-th target; Step 4.4: The total loss function is the weighted sum of the above three losses, and its formula is as follows: (16); in, , and is a hyperparameter that controls the weight of each loss term.

[0015] The dynamic monitoring head provides size, task and space awareness at the same time, and each attention is focused on a specific direction with the dimension of horizontal × spatial × channel. The scale-aware attention module fuses features of different scales according to their semantic importance. The spatial-aware attention module first uses deformable convolution to learn sparsity and then aggregates cross-layer features at the same spatial location. The task-aware attention module dynamically opens or closes feature channels and selects different tasks, which makes the input data adaptive.

[0016] Step 5: Use the training set to train the improved YOLOV8 model, and use the validation set and test set to test and verify the trained improved YOLOV8 model. The specific contents are: The training set is used to train the improved YOLOV8 model, and then the validation set and test set are used to evaluate the trained improved YOLOV8 model. The compiler is pycharm, and the learning framework of Python3.9.7+torch1.12+cu113 is used to train and detect the model. The initial learning rate is 0.01, the number of iterations is 300, and the evaluation indicators are recall rate R, precision rate P, average precision rate AP, average precision rate mAP and transmission frame rate per second FPS. The calculation formula is as follows: (17); (18); (19); (20); Among them, TP is a true positive example, FP is a false positive example, TN is a true negative example, and FN is a false negative example. Finally, the model that passes the evaluation is deployed in practical applications for reasoning on foreign body target detection tasks in mining belt conveyors.

[0017] The target detection method in the mining belt conveyor system of the present invention replaces the pixel-level enhancement module (PEM) with the SE module (Squeeze-and-Excitation module) in the IAT, which significantly improves the feature expression ability and adaptability of the model. The SE module enhances the focus on useful features and improves the feature representation ability of the model by adaptively reweighting the importance of each channel. Compared with the PEM, the SE module has lower computational complexity and parameter quantity, making the IAT more efficient in processing high-resolution images. At the same time, the adaptive mechanism of the SE module enables the model to better adapt to changes in different lighting conditions and background complexity, thereby improving the preprocessing effect.

[0018] The improved IAT (Illumination Adaptive Transformer) module can perform illumination adaptive processing on the input image, effectively solving the problem of image quality degradation caused by illumination changes. Through the collaborative work of local branches and global branches, IAT can extract more robust features, thereby improving the performance of the model under various lighting conditions.

[0019] By replacing the original YOLOV8 detection head with the dynamic detection head Dyhead, DI-YOLOV8 has been significantly improved in multi-scale target detection and feature expression capabilities. The Dyhead detection head introduces a dynamic convolution mechanism that can adaptively adjust the convolution kernel parameters according to the input features, improving the flexibility and expression of feature extraction. Compared with the traditional fixed convolution kernel detection head, Dyhead can better capture target features in different scales and complex scenes, and enhance the detection accuracy of small and large targets. In addition, Dyhead's feature fusion strategy enables the model to make more full use of multi-level feature information, improving the richness and accuracy of feature representation. Combined with the preprocessing effect of IAT, DI-YOLOV8 shows higher detection accuracy and robustness in multi-scale target detection tasks, and its adaptability and versatility have been further improved.

[0020] Example 1 The target detection method in the mining belt conveyor system of the present invention is specifically implemented according to the following steps: Step 1: Obtain images of coal transported by conveyors, manually annotate foreign objects to obtain a data set, and divide the data set into a training set, a validation set, and a test set to ensure that the data in each subset is evenly distributed; Step 2: At the beginning of the backbone network of YOLOV8, replace the original first convolutional layer and add the improved IAT network as a preprocessing module; Step 3: Replace the original YOLOV8 detection head with the dynamic detection head to obtain the improved YOLOV8 model; Step 4: Set the loss function; Step 5: Use the training set to train the improved YOLOV8 model, and use the validation set and test set to test and verify the trained improved YOLOV8 model.

[0021] Example 2 In the target detection method in the mining belt conveyor system of the present invention, the specific process of step 1 is as follows: The images of conveyors transporting coal mines are obtained, and the foreign objects are manually annotated to obtain the data set, which is divided into training set, validation set and test set. The training set accounts for 70% of the total set, the validation set accounts for 20% of the total set, and the test set accounts for 10% of the total set.

[0022] The specific process of adding the improved IAT network as a preprocessing module in step 2 is: The existing IAT network model is used, which is divided into local branches and global branches. The pixel-level enhancement module of the local branch is replaced by the SE module. In the IAT network model, the input feature map is X, and the size is , where H and W represent the height and width of the feature map respectively, and C represents the number of channels; For local branches, convolution operations are used to extract local features: (1); After local feature extraction, the self-attention mechanism is used to enhance feature representation: (2); For the global branch, it is implemented through global pooling operation and fully connected layer, and the global features are extracted by global pooling operation: (3); The global features are further processed through the fully connected layer: (4); The SE module is introduced, and the steps of the SE module include compression and excitation; Perform global average pooling on the input feature map to compress the spatial information of each channel, where a and b are the indexes of the spatial dimensions of the input feature map: a represents the index of the height dimension of the feature map, from 1 to H, where H is the height of the feature map; b represents the index of the width dimension of the feature map, from 1 to W, where W is the width of the feature map: (5); The compression operation obtains a vector 𝑧 with a channel number of 𝐶; The compressed features are recalibrated through two fully connected layers: (6); in, and are the weight matrices of the two fully connected layers, 𝛿 represents the ReLU activation function, and 𝜎 represents the Sigmoid activation function; Multiply the excited feature 𝑠 by the input feature map 𝑋 channel by channel: (7); Among them, ⊙ represents the channel-by-channel multiplication operation; Merge the features of local branches and global branches: (8); The final output feature map is the fused feature Y.

[0023] The present invention replaces the PEM module with the SE module (Squeeze-and-Excitation module), which enhances the attention ability between channels by adaptively adjusting the importance of each channel, thereby more effectively capturing the correlation between channels and improving the characterization ability of features. Compared with the PEM module, the SE module implements the attention mechanism between channels through global pooling and two fully connected layers, with low computational complexity and parameter amount, which improves computational efficiency. The SE module only needs to introduce a small number of parameters, which significantly reduces the storage requirements and training difficulty of the model. It is simple to implement and easy to integrate into the existing network architecture without large-scale structural adjustment. By introducing the SE module, the model's attention to important features in the image can be effectively enhanced, and the performance and robustness of the model can be improved. The SE module is suitable for various network architectures and tasks, has good versatility, and can achieve good performance improvement on different models and data sets.

[0024] Example 3 In the target detection method in the mining belt conveyor system of the present invention, the specific process of step 3 is as follows: Step 3.1: Replace the three detection heads of YOLOV8 with the improved dynamic detection head, and change the deformable convolution in the dynamic detection head from DCNv2 to DCNv4. Compared with the original DCNv2, DCNv4 can learn more complex spatial deformations and support the extraction of features of different scales, improving the adaptability to diverse targets; extract features of different scales from Backbone, and adjust the scale by upsampling and downsampling. After adjustment, pass it into the detection head, and the input feature is a four-dimensional tensor ,in ; L, S, C represent scale, space and channel respectively; Step 3.2: Constructing a scale attention mechanism , dynamically fuses features of different scales according to semantic importance, which enables the importance of various feature levels to adapt to the input. First, a 1*1 convolutional layer is used to adaptively learn the attention weights of each scale, and then an adaptive sigmoid activation function is introduced to assign weights to features of each scale. Finally, the normalized attention weight vector is applied to the multi-scale feature map and the corresponding output is generated. The specific formula is as follows: (9); Function f( ) is the linear function of the convolutional layer, is the adaptive sigmoid activation function, and its expression is , by introducing additional parameters a and b for adaptive adjustment, it is suitable for dynamically adjusting weights according to input features; Step 3.3: Spatial-aware attention fuses the multi-layer features after the previous step, learns the offset of the feature map through deformable convolution, and then adaptively aggregates multiple levels of features to learn more discriminative representations; the specific expression is as follows: (10); Where k represents the number of sparse sampling spatial locations, Represents the feature map after deformable convolution, which can focus on a specific spatial location more flexibly while taking into account the dynamically learned offset. and They represent weights and dynamically learned feature representations, which can be dynamically adjusted according to specific inputs to improve the effectiveness of features. is the spatial offset learned by the network, Indicates the offset; Step 3.4: Combine multiple detection tasks by dynamically controlling feature channels. (11); in Represents the feature map The specific feature channel extracted from , , , Represents a vector that controls the activation threshold. By taking the maximum value, the model can select the feature channel combination that performs best in the current task; Step 3.5: Finally, we introduce scale-aware attention, spatial position-aware attention, and task-aware attention into the dynamic detection head. The three-dimensional feature tensor at the detection layer is F∈R L×S×C , the attention function calculation formula is as follows, where L is the level, S is the space, and C is the channel: (12); pass , , The modules are stacked one after another, so that the detection head can better integrate contextual information and enhance feature extraction.

[0025] The dynamic monitoring head provides size, task and space awareness at the same time, and each attention is focused on a specific direction with the dimension of horizontal × spatial × channel. The scale-aware attention module fuses features of different scales according to their semantic importance. The spatial-aware attention module first uses deformable convolution to learn sparsity and then aggregates cross-layer features at the same spatial location. The task-aware attention module dynamically opens or closes feature channels and selects different tasks, which makes the input data adaptive.

[0026] Example 4 The target detection method in the mining belt conveyor system of the present invention, the specific process of step 4 is: defining a multi-task loss function, the loss function includes the following three parts: bounding box regression loss, category classification loss and confidence loss; Step 4.1: In the bounding box regression loss, L1 loss or L2 loss is used to measure the difference between the predicted bounding box and the true bounding box. The L1 loss formula is as follows: (13); in, is the i-th predicted bounding box, is the i-th ground-truth bounding box, N is the total number of objects, represents the L1 norm; Step 4.2: The formula for class classification loss is as follows: (14); is the true category label of the i-th target, is the probability that the i-th target is predicted to be category c, and M is the total number of categories; Step 4.3: The confidence loss formula is as follows: (15); in, is the true confidence label of the i-th target, is the confidence prediction of the i-th target; Step 4.4: The total loss function is the weighted sum of the above three losses, and its formula is as follows: (16); in, , and is a hyperparameter that controls the weight of each loss term.

[0027] Example 5 In the target detection method in the mining belt conveyor system of the present invention, the specific content of step 5 is: The training set is used to train the improved YOLOV8 model, and then the validation set and test set are used to evaluate the trained improved YOLOV8 model. The compiler is pycharm, and the learning framework of Python3.9.7+torch1.12+cu113 is used to train and detect the model. The initial learning rate is 0.01, the number of iterations is 300, and the evaluation indicators are recall rate R, precision rate P, average precision rate AP, average precision rate mAP and transmission frame rate per second FPS. The calculation formula is as follows: (17); (18); (19); (20); Among them, TP is a true positive example, FP is a false positive example, TN is a true negative example, and FN is a false negative example. Finally, the model that passes the evaluation is deployed in practical applications for reasoning on foreign body target detection tasks in mining belt conveyors.

[0028] The target detection method in the mining belt conveyor system of the present invention replaces the pixel-level enhancement module (PEM) with the SE module (Squeeze-and-Excitation module) in the IAT, which significantly improves the feature expression ability and adaptability of the model. The SE module enhances the focus on useful features and improves the feature representation ability of the model by adaptively reweighting the importance of each channel. Compared with the PEM, the SE module has lower computational complexity and parameter quantity, making the IAT more efficient in processing high-resolution images. At the same time, the adaptive mechanism of the SE module enables the model to better adapt to changes in different lighting conditions and background complexity, thereby improving the preprocessing effect.

[0029] The improved IAT (Illumination Adaptive Transformer) module can perform illumination adaptive processing on the input image, effectively solving the problem of image quality degradation caused by illumination changes. Through the collaborative work of local branches and global branches, IAT can extract more robust features, thereby improving the performance of the model under various lighting conditions.

[0030] By replacing the original YOLOV8 detection head with the dynamic detection head Dyhead, DI-YOLOV8 has been significantly improved in multi-scale target detection and feature expression capabilities. The Dyhead detection head introduces a dynamic convolution mechanism that can adaptively adjust the convolution kernel parameters according to the input features, improving the flexibility and expression of feature extraction. Compared with the traditional fixed convolution kernel detection head, Dyhead can better capture target features in different scales and complex scenes, and enhance the detection accuracy of small and large targets. In addition, Dyhead's feature fusion strategy enables the model to make more full use of multi-level feature information, improving the richness and accuracy of feature representation. Combined with the preprocessing effect of IAT, DI-YOLOV8 shows higher detection accuracy and robustness in multi-scale target detection tasks, and its adaptability and versatility have been further improved.

Claims

1. A target detection method in a mining belt conveyor system, characterized in that: Follow the steps below to implement it: Step 1: Obtain images of coal transported by conveyors, manually annotate foreign objects to obtain a data set, and divide the data set into a training set, a validation set, and a test set to ensure that the data in each subset is evenly distributed; Step 2: At the beginning of the backbone network of YOLOV8, replace the original first convolutional layer and add the improved IAT network as a preprocessing module; Step 3: Replace the original YOLOV8 detection head with the dynamic detection head to obtain the improved YOLOV8 model; Step 4: Set the loss function; Step 5: Use the training set to train the improved YOLOV8 model, and use the validation set and test set to test and verify the trained improved YOLOV8 model.

2. The target detection method in a mining belt conveyor system according to claim 1, characterized in that: The specific process of step 1 is: The images of conveyors transporting coal mines are obtained, and the foreign objects are manually annotated to obtain the data set, which is divided into training set, validation set and test set. The training set accounts for 70% of the total set, the validation set accounts for 20% of the total set, and the test set accounts for 10% of the total set.

3. The target detection method in a mining belt conveyor system according to claim 2, characterized in that: The specific process of adding the improved IAT network as a preprocessing module in step 2 is: The existing IAT network model is used, which is divided into local branches and global branches. The pixel-level enhancement module of the local branch is replaced by the SE module. In the IAT network model, the input feature map is X, and the size is , where H and W represent the height and width of the feature map respectively, and C represents the number of channels; For local branches, convolution operations are used to extract local features: (1)； After local feature extraction, the self-attention mechanism is used to enhance feature representation: (2)； For the global branch, it is implemented through global pooling operation and fully connected layer, and the global features are extracted by global pooling operation: (3)； The global features are further processed through the fully connected layer: (4)； The SE module is introduced, and the steps of the SE module include compression and excitation; Perform global average pooling on the input feature map to compress the spatial information of each channel, where a and b are the indexes of the spatial dimensions of the input feature map: a represents the index of the height dimension of the feature map, from 1 to H, where H is the height of the feature map; b represents the index of the width dimension of the feature map, from 1 to W, where W is the width of the feature map: (5)； The compression operation obtains a vector 𝑧 with 𝐶 channels; The compressed features are recalibrated through two fully connected layers: (6)； in, and are the weight matrices of the two fully connected layers, 𝛿 represents the ReLU activation function, and 𝜎 represents the Sigmoid activation function; Multiply the excited feature 𝑠 by the input feature map 𝑋 channel by channel: (7)； Among them, ⊙ represents the channel-by-channel multiplication operation; Merge the features of local branches and global branches: (8)； The final output feature map is the fused feature Y.

4. The target detection method in a mining belt conveyor system according to claim 3, characterized in that: The specific process of step 3 is as follows: Step 3.1: Replace the three detection heads of YOLOV8 with the improved dynamic detection head, and change the deformable convolution in the dynamic detection head from DCNv2 to DCNv; extract features of different scales from Backbone, and adjust the scale by upsampling and downsampling. After adjustment, pass it into the detection head, and the input feature is a four-dimensional tensor ,in ; L, S, C represent scale, space and channel respectively; Step 3.2: Constructing a scale attention mechanism , dynamically fuse features of different scales according to semantic importance, first use a 1*1 convolutional layer to adaptively learn the attention weights of each scale, then introduce an adaptive sigmoid activation function to assign weights to features of each scale, and finally apply the normalized attention weight vector to the multi-scale feature map and generate the corresponding output results. The specific formula is as follows: （9）； Function f( ) is the linear function of the convolutional layer, is the adaptive sigmoid activation function, and its expression is , by introducing additional parameters a and b for adaptive adjustment, it is suitable for dynamically adjusting weights according to input features; Step 3.3: Spatial-aware attention fuses the multi-layer features after the previous step, learns the offset of the feature map through deformable convolution, and then adaptively aggregates multiple levels of features to learn more discriminative representations; the specific expression is as follows: （10）； Where k represents the number of sparse sampling spatial locations, Represents the feature map after deformable convolution, which can focus on a specific spatial location more flexibly while taking into account the dynamically learned offset. and They represent weights and dynamically learned feature representations, which can be dynamically adjusted according to specific inputs to improve the effectiveness of features. is the spatial offset learned by the network, Indicates the offset; Step 3.4: Combine multiple detection tasks by dynamically controlling feature channels. （11）； in Represents the feature map The specific feature channel extracted from , , , Represents a vector that controls the activation threshold. By taking the maximum value, the model can select the feature channel combination that performs best in the current task; Step 3.5: Finally, we introduce scale-aware attention, spatial position-aware attention, and task-aware attention into the dynamic detection head. The three-dimensional feature tensor at the detection layer is F∈R L×S×C , the attention function calculation formula is as follows, where L is the level, S is the space, and C is the channel: （12）； pass , , The modules are stacked one after another, so that the detection head can better integrate contextual information and enhance feature extraction.

5. The target detection method in a mining belt conveyor system according to claim 1, characterized in that: The specific process of step 4 is: defining a multi-task loss function, which includes the following three parts: bounding box regression loss, category classification loss and confidence loss; Step 4.1: In the bounding box regression loss, use L1 loss or L2 loss to measure the difference between the predicted bounding box and the true bounding box. The L1 loss formula is as follows: (13)； in, is the i-th predicted bounding box, is the i-th ground-truth bounding box, N is the total number of objects, represents the L1 norm; Step 4.2: The formula for class classification loss is as follows: (14)； is the true category label of the i-th target, is the probability that the i-th target is predicted to be category c, and M is the total number of categories; Step 4.3: The confidence loss formula is as follows: (15)； in, is the true confidence label of the i-th target, is the confidence prediction of the i-th target; Step 4.4: The total loss function is the weighted sum of the above three losses, and its formula is as follows: (16)； in, , and is a hyperparameter that controls the weight of each loss term.

6. The target detection method in a mining belt conveyor system according to claim 1, characterized in that: The specific content of step 5 is: The training set is used to train the improved YOLOV8 model, and then the validation set and test set are used to evaluate the trained improved YOLOV8 model. The compiler is pycharm, and the learning framework of Python3.9.7+torch1.12+cu113 is used to train and detect the model. The initial learning rate is 0.01, the number of iterations is 300, and the evaluation indicators are recall rate R, precision rate P, average precision rate AP, average precision rate mAP and transmission frame rate per second FPS. The calculation formula is as follows: （17）； （18）； （19）； （20）； Among them, TP is a true positive example, FP is a false positive example, TN is a true negative example, and FN is a false negative example. Finally, the model that passes the evaluation is deployed in practical applications for reasoning on foreign body target detection tasks in mining belt conveyors.

Images