An industrial-grade mine belt boundary identification method based on multi-frame time sequence fusion

Abstract

The present application relates to the technical field of intelligent control and industrial visual detection of beneficiation process, and discloses an industrial-grade ore belt boundary identification method based on multi-frame time sequence fusion. A multi-branch hollow boundary enhancement unit is constructed based on deep convolution, different convolution kernel sizes, and a Sobel operator. A feature extraction and fusion module is constructed based on unbiased convolution, time sequence cross-channel difference, the multi-branch hollow boundary enhancement unit, and multi-head attention. A feature enhancement module is constructed based on deep separable convolution and the multi-branch hollow boundary enhancement unit. A prediction module is constructed based on the multi-branch hollow boundary enhancement unit. An identification model is constructed based on the feature extraction and fusion module, the feature enhancement module, and the prediction module. Continuous multi-frame pictures in an industrial field are acquired, and after preprocessing, the multi-frame pictures are input into the identification model to obtain ore belt boundary identification results. Based on the ore belt boundary identification results, sorting is performed, and the problem that existing table sorting and identification methods cannot effectively identify ore belt boundaries is solved.

Description

Technical Field

[0001] This invention relates to the fields of intelligent control of mineral processing and industrial visual inspection technology, and in particular to a method for identifying the boundary of ore zones on a shaking table based on machine vision and a method for constructing a mineral boundary detection model. Background Technology

[0002] Shaking table separation is the most widely used fine separation equipment in gravity separation processes, and its separation effect is highly dependent on the accurate identification of the ore zone boundary. However, in actual mineral processing plants, shaking table ore zone images are subject to complex interferences such as uneven lighting, water film reflection, local overexposure, and vibration blurring, making it difficult for existing vision methods to meet the accuracy and real-time requirements of industrial production.

[0003] First, industrial cameras are usually installed at a height of 1-1.5m. Due to limitations in resolution and lighting conditions, the boundary of the mining zone (only 0.5-2mm wide) appears as an extremely narrow structure in the image. Traditional edge detection methods (such as Canny and Sobel) can have a false detection rate of 12%-18% in areas with strong reflection and weak texture, making it difficult to obtain continuous and stable boundaries.

[0004] Secondly, the shaking table experiences periodic vibrations of 8-14Hz during operation, and the slurry flow exhibits significant time-varying characteristics, meaning that a single frame image often fails to reflect the true boundary positions. Existing single-frame convolutional neural networks (such as U-Net and DeepLabV3+) lack temporal modeling capabilities, resulting in frame jump rates exceeding 30% under vibration conditions. This causes oscillations in the PLC's actuator, affecting the stable output of concentrate.

[0005] Secondly, mineral belts are composed of both coarse-grained and fine-grained minerals, with significant differences in scale. Coarse-grained mineral belts have strong textures and large areas, while fine-grained mineral belts have weak textures and blurred boundaries. A single-scale convolutional structure is difficult to take into account both types of mineral belts, and problems such as misidentification of coarse-grained mineral belts or breakage of fine-grained mineral belts are prone to occur.

[0006] Furthermore, industrial environments demand stringent real-time performance. Shaking table sorting typically requires real-time setup with image input at 20-30fps, necessitating a single-frame inference time for the visual model of no more than 50ms. However, mainstream high-precision segmentation networks (such as DeepLabV3+) generally have inference times exceeding 90–120ms, making them unsuitable for industrial deployments.

[0007] In summary, existing technologies have significant shortcomings in terms of boundary accuracy, temporal stability, multi-scale adaptability, and real-time performance, failing to meet the demands of ore dressing workshops for automated and intelligent sorting control. Therefore, it is necessary to propose a method for identifying ore zone boundaries that can achieve high precision, high robustness, and real-time inference under complex industrial interference conditions. Summary of the Invention

[0008] This invention provides an industrial-grade mineral zone boundary identification method based on multi-frame temporal fusion to solve the problem that existing shaking table sorting and identification methods cannot effectively identify mineral zone boundaries.

[0009] To achieve the above objectives, the present invention employs the following technical solution:

[0010] This invention provides an industrial-grade mineral belt boundary identification method based on multi-frame temporal fusion, comprising the following steps: Step 1: Construct a multi-branch dilated boundary enhancement unit based on depthwise convolution, dilated convolution with different kernel sizes, and the Sobel operator using a multi-branch progressive merging structure to enhance feature boundaries. Construct a feature extraction and fusion module based on unbiased convolution and temporal cross-channel difference, combined with the multi-branch dilated boundary enhancement unit and multi-head attention. Construct a feature enhancement module based on depthwise separable convolution and multi-branch dilated boundary enhancement unit stacked and combined with downsampling and residual connection. Construct a prediction module based on multi-branch dilated boundary enhancement unit and multiple perceptron. Construct a recognition model based on the feature extraction and fusion module, feature enhancement module, and prediction module. Step 2: Acquire multiple consecutive frames of images during the industrial shaking table sorting process, preprocess the multiple frames of images to obtain multiple preprocessed images, input the multiple preprocessed images into the recognition model to obtain the mineral zone boundary recognition result, and perform sorting based on the mineral zone boundary recognition result.

[0011] Furthermore, the multi-branch cavity boundary enhancement unit performs a channel number consistency check on the input features and splits the input features into fine-grained mineral zone features, coarse-grained mineral zone features, and real-time features. It performs cavity convolution enhancement and directional gradient enhancement processing on the fine-grained and coarse-grained mineral zone features with different receptive fields, and performs depth convolution and convolution channel processing on the real-time features. It then fuses the enhanced fine-grained and coarse-grained mineral zone features to obtain a first fused feature, and then fuses the processed real-time features with the first fused feature to obtain a second fused feature. Finally, it performs convolution channel processing on the second fused feature and outputs the feature boundary enhanced feature.

[0012] Furthermore, the multi-branch void boundary enhancement unit introduces consistency constraints to constrain features that are in different time sequences after feature boundary enhancement.

[0013] Through the above design, the three-frame temporal difference and cross-frame consistency constraints are introduced into the shaking table ore zone identification task. By capturing the ore zone movement trend and the illumination change pattern, the boundary jump caused by 8-14Hz shaking table vibration and slurry flow is effectively suppressed.

[0014] Furthermore, the multi-branch hole boundary enhancement unit includes a feature verification splitting layer, a multi-processing branch structure, a first fusion layer, a second fusion layer, and a first channel convolutional layer. The multi-processing branch structure includes a first processing branch, a second processing branch, and a third processing branch. The feature verification splitting layer is connected to the input of the multi-processing branch structure. The outputs of the first and second branches of the multi-processing branch are connected to the input of the first fusion layer. The outputs of the first fusion layer and the third branch are connected to the input of the second fusion layer. The output of the second fusion layer is connected to the input of the first channel convolutional layer.

[0015] Furthermore, the first processing branch sequentially includes a dilated convolution with a kernel of 5 and a dilation rate of 2, and a Soble operator; The second processing branch includes, in sequence, a dilated convolution with a kernel of 7 and a dilation rate of 3, and a Soble operator; The third processing branch includes a depthwise convolutional layer with a kernel of 3 and a compensation of 1, and a second-channel convolutional layer.

[0016] Furthermore, the feature extraction and fusion module extracts basic features from the input multi-frame preprocessed images through multiple channels to obtain basic features of multiple frames. Based on the basic features of the previous frame and the basic features of the current frame, it combines absolute difference and downsampling to obtain the first difference feature. Based on the basic features of the current frame and the basic features of the next frame, it combines absolute difference and downsampling to obtain the second difference feature. The first difference feature, the second difference feature, and the downsampled basic features of the current frame are subjected to feature boundary enhancement processing. The first difference feature, the second difference feature, and the basic features of the current frame after boundary enhancement are combined with multi-head attention to obtain the feature extraction and fusion feature.

[0017] Furthermore, the feature extracted and fused by combining the first difference feature, the second difference feature, and the basic feature of the current frame with the feature obtained through multi-head attention includes: The base features of the current frame after feature boundary enhancement are used as the query vector of multi-head attention. The first and second difference features after feature boundary enhancement are used as the key vector of multi-head attention. The first and second difference features after feature boundary enhancement, as well as the base features of the current frame, are used as the value vector. The multi-head attention output features are then combined with the features extracted and fused.

[0018] Furthermore, the feature extraction and fusion module includes a multi-branch extraction structure, a first cross-channel difference layer, a second cross-channel difference layer, multiple downsampling layers, multiple multi-branch hole boundary enhancement units, and multi-head attention; The multi-branch extraction structure includes a first extraction branch, a second extraction branch, and a third extraction branch. The first extraction branch, the second extraction branch, and the third extraction branch are all constructed based on an unbiased convolution with a kernel of 3. The input terminal of the first cross-channel difference layer is connected to the output terminals of the first extraction branch and the second extraction branch. The input terminal of the second cross-channel difference layer is connected to the output terminals of the second extraction branch and the third extraction branch. The output terminals of the second extraction branch, the first cross-channel difference layer, and the second cross-channel difference layer are sequentially connected to the downsampling layer, the multi-branch hole boundary enhancement unit, and the multi-head attention layer.

[0019] Furthermore, the feature enhancement module repeatedly downsamples and enhances the feature boundaries of the feature extraction and fusion module to obtain enhanced features at different scales. It then performs convolutional channel processing on the enhanced features at different scales and combines upsampling and residual connection to obtain the third fused feature after feature enhancement.

[0020] Furthermore, the feature enhancement module includes a downsampling backbone, a main branch, an auxiliary branch, upsampling, and residual connections; The downsampling backbone is constructed based on stacked downsampling enhancement submodules, which are constructed based on downsampling layers and multi-branch hole boundary enhancement units. The input ends of the main branch and auxiliary branch are connected to the multi-branch hole boundary enhancement unit in different downsampling enhancement submodules. Both the main branch and auxiliary branch are constructed based on channel convolution. The output end of the auxiliary branch is connected to the residual of the main branch after upsampling.

[0021] Furthermore, the prediction module includes a multi-branch cavity boundary enhancement unit, a first prediction branch, and a second prediction branch. Both the first and second prediction branches are constructed based on a multiple perceptron. The output of the multi-branch cavity boundary enhancement unit is connected to the input of the first and second prediction branches.

[0022] Furthermore, the ore zone boundary identification results include anomaly region boxes, ore zone boundary probability maps, and statistical time. The first prediction branch uses a multiple perceptron to acquire abnormal region boxes, and performs non-maximum suppression optimization on the abnormal region boxes output by the first prediction branch. The second prediction branch uses a convolutional layer to generate a probability map of ore zone boundaries and performs morphological closing operations to optimize the probability map of ore zone boundaries. The recognition model outputs a statistical time for inference time, and the model reduces model complexity by combining the statistical time with a statistical threshold and a lightweight switching mechanism.

[0023] Furthermore, the sorting based on the mineral belt boundary identification results includes: performing a quality score based on the mineral belt boundary probability map, calculating the center line position of the boundary line based on the mineral belt boundary probability map, obtaining a first control quantity, obtaining a PLC control quantity based on the first control quantity combined with proportional mapping, and encapsulating and outputting the abnormal area box, mineral belt boundary probability map, quality score, and PLC control quantity to the PLC, and the PLC performing sorting based on the encapsulated package.

[0024] Through the above design, the abnormal area bounding box and the probability map of the ore zone boundary can be directly mapped to PLC control commands, realizing automatic adjustment of the shaking table sorting process. The system has an extremely short end-to-end response time, which can meet the requirements of industrial control systems for rapid response, effectively reduce manual intervention, and improve production stability.

[0025] Beneficial effects: This invention provides an industrial-grade mineral zone boundary identification method based on multi-frame temporal fusion. It employs a multi-branch progressive merging structure using dilated convolutions with different kernel sizes to construct multi-branch dilated boundary enhancement units, effectively strengthening extremely narrow mineral zone boundaries of only 0.5-2 mm in width. This ensures that the boundary structure remains clear and separable even under conditions of low contrast, weak texture, and reflective interference. Industrial verification shows that the boundary positioning error is stably controlled within ≤2 mm, representing an improvement of approximately 60%-70% compared to traditional edge detection methods and approximately 30%-40% compared to mainstream deep learning methods. This significantly overcomes the performance bottleneck of existing technologies in extremely narrow boundary identification and solves the problem that existing shaking table sorting and identification methods cannot effectively identify mineral zone boundaries. The multi-head attention fusion in the feature extraction and fusion module can adaptively focus on the strong texture regions of coarse-grained mineral zones and the weak texture regions of fine-grained mineral zones, and achieve the collaborative expression of multi-scale features through the dual-branch structure of the feature enhancement module. This structure effectively avoids problems such as the breakage of fine-grained mineral zones and the offset of coarse-grained mineral zones, significantly improving the integrity and continuity of the overall boundary recognition, and its adaptability is better than that of existing single-scale convolutional structures. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the network structure of the multi-branch void boundary enhancement unit in the recognition model of this invention embodiment; Figure 2 This is a schematic diagram of the network structure of the recognition model in an embodiment of the present invention. Detailed Implementation

[0027] The technical solution of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0028] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms "an" or "a" and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms "connected" or "linked" and similar terms are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. "Up," "down," "left," "right," etc., are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship also changes accordingly.

[0029] This application provides an industrial-grade mineral belt boundary identification method based on multi-frame temporal fusion, comprising the following steps: Step 1: Based on depthwise convolution, dilated convolution with different kernel sizes, and the Sobel operator, a multi-branch progressive merging structure is used to construct a multi-branch dilated boundary enhancement unit to enhance feature boundaries. Based on unbiased convolution and temporal cross-channel difference, a multi-branch dilated boundary enhancement unit and multi-head attention are combined to construct a feature extraction and fusion module. Based on depthwise separable convolution and multi-branch dilated boundary enhancement units, a stacked combination of downsampling and residual connection is used to construct a feature enhancement module. Based on multi-branch dilated boundary enhancement units and multiple perceptrons, a prediction module is constructed. Based on the feature extraction and fusion module, the feature enhancement module, and the prediction module, a recognition model is constructed. Please see Figure 1 The multi-branch void boundary enhancement unit performs a channel number consistency check on the input features to adapt to fine-grained ore zones, coarse-grained ore zones, and real-time features. It splits the input features into fine-grained ore zone features, coarse-grained ore zone features, and real-time features. The channel number consistency check enables ore zone features of different scales to be modeled in independent channel spaces, improving the stability of subsequent boundary enhancement. After the consistency check, the fine-grained ore zone features and coarse-grained ore zone features are subjected to void convolution enhancement and directional gradient enhancement processing with different receptive fields. The real-time features are subjected to depth convolution and convolution channel processing. The enhanced fine-grained ore zone features and coarse-grained ore zone features are fused to obtain the first fused feature. Then, the processed real-time features are fused with the first fused feature to obtain the second fused feature. After the second fused feature is processed by convolution channel, the feature boundary enhanced feature is output. The specific network structure of the multi-branch hole boundary enhancement unit includes a feature verification splitting layer, a multi-processing branch structure, a first fusion layer, a second fusion layer, and a first channel convolutional layer. The multi-processing branch structure includes a first processing branch, a second processing branch, and a third processing branch. The first fusion layer and the second fusion layer are implemented using element-wise multiplication. The feature verification splitting layer is connected to the input of the multi-processing branch structure. The outputs of the first and second branches of the multi-processing branch are connected to the input of the first fusion layer. The outputs of the first fusion layer and the third branch are connected to the input of the second fusion layer. The output of the second fusion layer is connected to the input of the first channel convolutional layer. In this embodiment, the first processing branch includes a dilated convolution with a kernel size of 5 and a dilation rate of 2, and a Soble operator. The dilated convolution expands the effective receptive field while keeping the feature size unchanged. The kernel size and dilation rate effectively cover the complete boundary region of the fine-grained mineral zone, avoiding boundary breakage due to insufficient receptive field. The Soble operator effectively enhances the true boundary of the fine-grained mineral zone, thereby suppressing water film reflection and high-frequency noise, making the boundary line clearer. The second processing branch includes a dilated convolution with a kernel size of 7 and a void ratio of 3, and the Soble operator. The design of the second processing branch, the selection of the kernel size and void ratio, and the application of the Soble operator enable the boundary of the coarse-grained mineral zone to remain stable under the conditions of light reflection and changes in slurry thickness, thereby reducing the boundary missing rate and improving the positioning accuracy. The third processing branch includes a depthwise convolution with a kernel of 3 and a compensation of 1, and a second-channel convolutional layer. The design of the third processing branch is to select depthwise convolution to reduce the amount of computation while maintaining the texture details of the mineral belt for real-time features. The second-channel convolutional layer can also reduce redundant computation, which is more in line with the video data input of high frame rate industrial cameras. It also ensures the consistency of the number of subsequent channels, enabling real-time features to have cross-channel expressive capabilities. For the enhanced features output by the multi-branch hole boundary enhancement unit, a consistency constraint is introduced to constrain the enhanced features at different time sequences. This ensures the stability of the multi-branch hole boundary enhancement unit's output between consecutive frames and avoids boundary jumps caused by illumination fluctuations or vibrations. The consistency constraint can be expressed as the following formula: ; in, Indicating multi-branch void boundary reinforcement elements in Characteristics of frame output; Indicating multi-branch void boundary reinforcement elements in Characteristics of frame output; Indicates the cross-frame consistency threshold; Please see Figure 2For the feature extraction and fusion module of the recognition model, it extracts basic features from the input multi-frame preprocessed images through multiple channels to obtain basic features of multiple frames. Based on the basic features of the previous frame and the basic features of the current frame, it combines absolute difference and downsampling to obtain the first difference feature. Based on the basic features of the current frame and the basic features of the next frame, it combines absolute difference and downsampling to obtain the second difference feature. Feature boundary enhancement processing is performed on the first difference feature, the second difference feature, and the downsampled basic features of the current frame. The basic features of the current frame with enhanced feature boundaries are used as the query vector for multi-head attention, and the first and second difference features with enhanced feature boundaries are used as the key vector for multi-head attention. The features, the second difference features, and the basic features of the current frame are used as value vectors. Combined with the features extracted and fused from the multi-head attention output, the number of multi-head attention is set to 8 in this embodiment, corresponding to the ore strip width distribution in the industrial scenario of shaking table sorting. The attention softmax temperature coefficient is set to 0.5 to enhance the weight response of key boundary areas. Finally, a random deactivation mechanism is introduced after the attention weight calculation to adapt to the uneven labeling of industrial data and frequent changes in slurry state. The deactivation probability of this random deactivation mechanism is set to 0.1 in this embodiment. In other embodiments, the random deactivation mechanism can be removed based on the model's computing power requirements, and the features extracted and fused can be directly output.

[0030] In the feature extraction and fusion module, the use of absolute value difference avoids feature cancellation caused by small-amplitude movement of the ore belt during vibration, while downsampling reduces the computational load and enhances the significance of local changes. The absolute value difference and downsampling play a synergistic role in filtering static backgrounds such as the shaking table frame and support, so that the moving area of ​​the ore belt is prominently expressed in the feature space. For the industrial scenario of shaking table sorting, the use of multi-branch void boundary enhancement unit for feature boundary enhancement processing effectively solves the problems of blurred ore belt boundaries and gradient weakening caused by reflective water film. The specific design of query, key and value vectors in multi-head attention enables both the wide boundary of coarse ore belt and the narrow boundary of fine ore belt to obtain stable response, avoiding missed detection or boundary breakage caused by single-scale features.

[0031] The specific network structure of the feature extraction and fusion module includes a multi-branch extraction structure, a first cross-channel difference layer, a second cross-channel difference layer, multiple downsampling layers, multiple multi-branch hole boundary enhancement units, and multi-head attention. The multi-branch extraction structure includes a first extraction branch, a second extraction branch, and a third extraction branch. The first extraction branch, the second extraction branch, and the third extraction branch are all constructed based on unbiased convolution with a kernel of 3. The unbiased convolution here is mainly aimed at the dark current drift of industrial cameras in the industrial scenario of shaking table mining belt, thereby improving the stability of subsequent temporal difference. The input of the first cross-channel differential layer is connected to the output of the first extraction branch and the second extraction branch. The input of the second cross-channel differential layer is connected to the output of the second extraction branch and the third extraction branch. The output of the second extraction branch, the output of the first cross-channel differential layer, and the output of the second cross-channel differential layer are all connected to the downsampling layer, the multi-branch hole boundary enhancement unit, and the multi-head attention in sequence. For the feature enhancement module, it repeatedly performs downsampling and feature boundary enhancement processing on the features output by the feature extraction and fusion module to obtain enhanced features at different scales. It then performs convolution channel processing on the enhanced features at different scales, and finally combines upsampling and residual connection to obtain the third fusion feature after feature enhancement. The repeated downsampling and feature boundary enhancement are employed. Downsampling expands the effective receptive field, ensuring the full representation of wide boundaries for coarse-grained mineral zones and providing stable global boundary representation capabilities. For fine-grained mineral zones, it ensures the boundary lines remain clear even under low contrast conditions. The repeated use of feature boundary enhancement reduces computational load and detail loss caused by repeated downsampling. Upsampling and residual connections address the issue of detail loss in mineral zones due to resolution limitations of industrial cameras. They restore spatial details and maintain consistency across scales. In this embodiment, a feature validity check is introduced for the third fused feature output by the feature enhancement module. Features with abnormal amplitudes or values ​​of 0 are removed, primarily to prevent abnormal features from entering the final task header and avoid false detection boxes or boundary jumps. In other embodiments, based on computational requirements, the feature validity check can be removed, and the third fused feature can be directly output.

[0032] The specific network structure of the feature enhancement module includes downsampling backbone, main branch, auxiliary branch, upsampling, and residual connections; The downsampling backbone is constructed based on stacked downsampling enhancement submodules, which are constructed based on downsampling layers and multi-branch hole boundary enhancement units. The input ends of the main branch and auxiliary branch are connected to the multi-branch hole boundary enhancement unit in different downsampling enhancement submodules. Both the main branch and auxiliary branch are constructed based on channel convolution. The output end of the auxiliary branch is connected to the residual of the main branch after upsampling. The prediction module includes a multi-branch cavity boundary enhancement unit, a first prediction branch, and a second prediction branch. Both the first and second prediction branches are constructed based on a multiple perceptron. The output of the multi-branch cavity boundary enhancement unit is connected to the input of the first and second prediction branches. The first prediction branch uses a multiple perceptron to acquire the abnormal region box, and performs non-maximum suppression optimization on the abnormal region box output by the first prediction branch. Non-maximum suppression processing avoids false alarms caused by multiple interference sources in the industrial field. In this embodiment, the overlap suppression threshold for non-maximum suppression processing is set to 0.5, and the confidence screening threshold is set to 0.8. The second prediction branch uses a convolutional layer to generate a probability map of mineral zone boundaries and performs morphological closing operations to optimize the probability map of mineral zone boundaries. The size of the structuring element in the morphological closing operation is set to 3 to correspond to the width of the mineral zone boundaries. The recognition model also performs statistics on inference time to output statistical time. The recognition model reduces model complexity by combining statistical time with statistical thresholds and a lightweight switching mechanism. Finally, a recognition model is constructed based on the feature extraction and fusion module, the feature enhancement module, and the prediction module. After the model is constructed, an industrial camera is used to acquire continuous image frames at 25fps. The continuous image frames are preprocessed, including dehazing, grayscale normalization, smoothing, and overexposure repair. The dehazing threshold is set to 0.9 to adapt to the environment with high humidity and severe fogging. The grayscale normalization is set to [0.05, 0.8] to enhance the contrast between minerals and gangue. The standard deviation of Gaussian smoothing is set to 0.18 to suppress the high-frequency noise generated by fine-grained minerals under vibration conditions. Overexposure is to address the water film reflection problem and avoid false strong responses in this area during subsequent boundary enhancement. After preprocessing, the continuous image frames are divided into a validation set and a training set in a 1:9 ratio. The training set is input into the recognition model for training. During training, the Adam optimizer is used to iteratively update the network parameters. The initial learning rate is set to 0.001, the batch size is 16, and the number of training epochs is 100. To avoid overfitting caused by uneven distribution of industrial data, an early stopping strategy is introduced on the validation set. Training is automatically terminated when the validation loss no longer decreases within a certain number of consecutive epochs. After training is completed, the optimal weights are fixed and deployed to edge computing devices.

[0033] During the inference phase, the system receives images of the ore belt from an industrial camera, outputs abnormal area boxes, ore belt boundary probability maps, and statistical times through the recognition model, and uses the PLC control interface to guide the automatic adjustment of the shaking table sorting mechanism, realizing the "recognition-control" closed-loop operation.

[0034] Step 2: Acquire multiple consecutive frames of images during the industrial shaking table sorting process, preprocess the multiple frames of images to obtain multiple preprocessed images, input the multiple preprocessed images into the recognition model to obtain the mineral zone boundary recognition result, and perform sorting based on the mineral zone boundary recognition result.

[0035] For the preprocessing here, the same dehazing, grayscale normalization, smoothing, and overexposure repair were used, with the parameters remaining consistent.

[0036] The ore zone boundary identification results include abnormal region boxes, ore zone boundary probability maps, and statistical time. The specific process for sorting based on the ore zone boundary results includes: Quality scoring is performed based on the probability map of the ore zone boundary. Then, the position of the center line of the boundary line is calculated based on the probability map of the ore zone boundary to obtain the first control quantity. The PLC control quantity is obtained based on the first control quantity and the proportional mapping. The abnormal area box, the probability map of the ore zone boundary, the quality score, and the PLC control quantity are packaged and output to the PLC. The PLC performs sorting based on the package. The quality score is used to evaluate the continuity, integrity, and reliability of the boundary line, with a pass threshold of 0.9. The proportional mapping is set according to the response characteristics of the actuator, such as the correspondence between the output range and the physical action, which is set to 0~1 in this embodiment. The linear mapping is the PLC control range of 0~100. Finally, in this embodiment, a consistency constraint is introduced again for the ore zone boundary probability map, with a continuity threshold of 0.8, to achieve smooth constraint on the ore zone boundary probability maps of different time series. In other embodiments, this consistency constraint can also be removed based on computing power issues.

[0037] Finally, to verify the industrial-grade mineral boundary identification method based on multi-frame temporal fusion and its proposed identification model, this embodiment selects three representative mainstream methods in the current field of mineral boundary identification as comparison objects, including: traditional gradient operator methods (Canny, Sobel), single-frame convolutional neural network methods (U-Net, DeepLabV3+), and lightweight real-time segmentation networks (ENet, BiSeNet). All methods are trained and tested on the same shaking table dataset. Evaluation indicators include boundary localization error, false detection rate, cross-frame stability, inference speed, and process improvement effect. For specific comparison results, please refer to Table 1.

[0038] Table 1: Performance comparison of different methods on the task of identifying mineral belt boundaries.

[0039]

[0040] As shown in Table 1, our proposed method significantly outperforms the comparison methods in all key metrics, specifically: Regarding boundary positioning accuracy: The average error of the method in this embodiment is controlled at 1.3-1.6mm, which is about 70% higher than the traditional method and about 40% higher than the mainstream CNN method, meeting the strict industrial requirements for accuracy of ≤2mm.

[0041] In terms of robustness: Under conditions of strong reflection, weak light and vibration interference, the false detection rate of the method in this paper is only 0.9%-1.2%, which is significantly lower than other methods (minimum 3%).

[0042] In terms of cross-frame stability: thanks to three-frame differential, the cross-frame jump amplitude is reduced by more than 55%, effectively avoiding boundary jitter caused by vibration.

[0043] In terms of real-time performance: the inference time is stable at 38-45ms, which is better than high-precision networks such as U-Net and DeepLabV3+, and comparable to lightweight networks such as ENet and BiSeNet.

[0044] In terms of process improvement: In actual production environments, the method presented in this paper increases the recovery rate of tin concentrate by 3.4%-4.1%, which is significantly better than other methods (0-1.5%).

[0045] In summary, an industrial-grade ore zone boundary identification method based on multi-frame temporal fusion and its proposed identification model demonstrate significant advantages in accuracy, robustness, cross-frame stability, and process benefits. It is particularly suitable for mineral processing shaking table scenarios characterized by complex lighting, strong vibrations, and extremely narrow ore zones. Compared with existing mainstream methods, this method exhibits higher reliability and practical value in industrial applications.

[0046] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. An industrial-grade mineral belt boundary identification method based on multi-frame temporal fusion, characterized in that, Includes the following steps: Step 1: Construct a multi-branch dilated boundary enhancement unit based on depthwise convolution, dilated convolution with different kernel sizes, and the Sobel operator using a multi-branch progressive merging structure to enhance feature boundaries. Construct a feature extraction and fusion module based on unbiased convolution and temporal cross-channel difference, combined with the multi-branch dilated boundary enhancement unit and multi-head attention. Construct a feature enhancement module based on depthwise separable convolution and multi-branch dilated boundary enhancement unit stacked and combined with downsampling and residual connection. Construct a prediction module based on multi-branch dilated boundary enhancement unit and multiple perceptron. Construct a recognition model based on the feature extraction and fusion module, feature enhancement module, and prediction module. Step 2: Acquire multiple consecutive frames of images during the industrial shaking table sorting process, preprocess the multiple frames of images to obtain multiple preprocessed images, input the multiple preprocessed images into the recognition model to obtain the mineral zone boundary recognition result, and perform sorting based on the mineral zone boundary recognition result.

2. The industrial-grade mineral belt boundary identification method based on multi-frame temporal fusion according to claim 1, characterized in that, The multi-branch cavity boundary enhancement unit performs a channel number consistency check on the input features and splits the input features into fine-grained mineral zone features, coarse-grained mineral zone features, and real-time features. It performs cavity convolution enhancement and directional gradient enhancement on the fine-grained and coarse-grained mineral zone features with different receptive fields, and performs depth convolution and convolution channel processing on the real-time features. It then fuses the enhanced fine-grained and coarse-grained mineral zone features to obtain a first fused feature, and then fuses the processed real-time features with the first fused feature to obtain a second fused feature. Finally, it performs convolution channel processing on the second fused feature and outputs the feature boundary enhanced feature.

3. The industrial-grade mineral belt boundary identification method based on multi-frame temporal fusion according to claim 2, characterized in that, The multi-branch void boundary enhancement unit introduces consistency constraints to constrain features that are at different time sequences after feature boundary enhancement.

4. The industrial-grade mineral belt boundary identification method based on multi-frame temporal fusion according to claim 2, characterized in that, The multi-branch hole boundary enhancement unit includes a feature verification splitting layer, a multi-processing branch structure, a first fusion layer, a second fusion layer, and a first channel convolutional layer. The multi-processing branch structure includes a first processing branch, a second processing branch, and a third processing branch. The feature verification splitting layer is connected to the input of the multi-processing branch structure. The outputs of the first and second branches of the multi-processing branch are connected to the input of the first fusion layer. The outputs of the first fusion layer and the third branch are connected to the input of the second fusion layer. The output of the second fusion layer is connected to the input of the first channel convolutional layer.

5. The industrial-grade mineral belt boundary identification method based on multi-frame temporal fusion according to claim 4, characterized in that, The first processing branch includes, in sequence, a dilated convolution with a kernel of 5 and a dilation rate of 2, and a Soble operator; The second processing branch includes, in sequence, a dilated convolution with a kernel of 7 and a dilation rate of 3, and a Soble operator; The third processing branch includes a depthwise convolutional layer with a kernel of 3 and a compensation of 1, and a second-channel convolutional layer.

6. The industrial-grade mineral belt boundary identification method based on multi-frame temporal fusion according to claim 1, characterized in that, The feature extraction and fusion module extracts basic features from the input multi-frame preprocessed images through multiple channels to obtain basic features of multiple frames. Based on the basic features of the previous frame and the basic features of the current frame, it combines absolute difference and downsampling to obtain the first difference feature. Based on the basic features of the current frame and the basic features of the next frame, it combines absolute difference and downsampling to obtain the second difference feature. The first difference feature, the second difference feature, and the downsampled basic features of the current frame are subjected to feature boundary enhancement processing. The first difference feature, the second difference feature, and the basic features of the current frame after boundary enhancement are combined with multi-head attention to obtain the feature extraction and fusion feature.

7. The industrial-grade mineral belt boundary identification method based on multi-frame temporal fusion according to claim 1, characterized in that, The step of combining the first difference feature after boundary enhancement, the second difference feature, and the basic features of the current frame with multi-head attention to extract the fused features includes: The base features of the current frame after feature boundary enhancement are used as the query vector of multi-head attention. The first and second difference features after feature boundary enhancement are used as the key vector of multi-head attention. The first and second difference features after feature boundary enhancement, as well as the base features of the current frame, are used as the value vector. The multi-head attention output features are then combined with the features extracted and fused.

8. The industrial-grade mineral belt boundary identification method based on multi-frame temporal fusion according to claim 6 or 7, characterized in that, The feature extraction and fusion module includes a multi-branch extraction structure, a first cross-channel difference layer, a second cross-channel difference layer, multiple downsampling layers, multiple multi-branch hole boundary enhancement units, and multi-head attention; The multi-branch extraction structure includes a first extraction branch, a second extraction branch, and a third extraction branch. The first extraction branch, the second extraction branch, and the third extraction branch are all constructed based on an unbiased convolution with a kernel of 3. The input terminal of the first cross-channel difference layer is connected to the output terminals of the first extraction branch and the second extraction branch. The input terminal of the second cross-channel difference layer is connected to the output terminals of the second extraction branch and the third extraction branch. The output terminals of the second extraction branch, the first cross-channel difference layer, and the second cross-channel difference layer are sequentially connected to the downsampling layer, the multi-branch hole boundary enhancement unit, and the multi-head attention layer.

9. The industrial-grade mineral belt boundary identification method based on multi-frame temporal fusion according to claim 6 or 7, characterized in that, The feature enhancement module repeatedly downsamples and enhances the feature boundaries of the feature extraction and fusion module to obtain enhanced features at different scales. It then performs convolutional channel processing on the enhanced features at different scales and combines upsampling and residual connection to obtain the third fused feature after feature enhancement.

10. The industrial-grade mineral belt boundary identification method based on multi-frame temporal fusion according to claim 9, characterized in that, The feature enhancement module includes a downsampling backbone, main branches, auxiliary branches, upsampling, and residual connections; The downsampling backbone is constructed based on stacked downsampling enhancement submodules, which are constructed based on downsampling layers and multi-branch hole boundary enhancement units. The input ends of the main branch and auxiliary branch are connected to the multi-branch hole boundary enhancement unit in different downsampling enhancement submodules. Both the main branch and auxiliary branch are constructed based on channel convolution. The output end of the auxiliary branch is connected to the residual of the main branch after upsampling.

11. The industrial-grade mineral belt boundary identification method based on multi-frame temporal fusion according to claim 1, characterized in that, The prediction module includes a multi-branch cavity boundary enhancement unit, a first prediction branch, and a second prediction branch. Both the first and second prediction branches are constructed based on a multiple perceptron. The output of the multi-branch cavity boundary enhancement unit is connected to the input of the first and second prediction branches.

12. The industrial-grade mineral belt boundary identification method based on multi-frame temporal fusion according to claim 11, characterized in that, The mineral belt boundary identification results include anomaly region boxes, mineral belt boundary probability maps, and statistical time. The first prediction branch uses a multiple perceptron to acquire abnormal region boxes, and performs non-maximum suppression optimization on the abnormal region boxes output by the first prediction branch. The second prediction branch uses a convolutional layer to generate a probability map of ore zone boundaries and performs morphological closing operations to optimize the probability map of ore zone boundaries. The recognition model outputs a statistical time for inference time, and the model reduces model complexity by combining the statistical time with a statistical threshold and a lightweight switching mechanism.

13. The industrial-grade mineral belt boundary identification method based on multi-frame temporal fusion according to claim 11, characterized in that, The sorting based on the mineral belt boundary identification results includes: scoring the quality based on the mineral belt boundary probability map, calculating the center line position of the boundary line based on the mineral belt boundary probability map, obtaining the first control quantity, obtaining the PLC control quantity based on the first control quantity and the proportional mapping, and encapsulating and packaging the abnormal area box, the mineral belt boundary probability map, the quality score, and the PLC control quantity into a package and outputting it to the PLC. The PLC performs sorting based on the package.

Images