ElectroTrackNet electro-selection track recognition method and system

Abstract

The present invention discloses an ElectroTrackNet electrostatic track recognition method and system, the method comprising: collecting electrostatic images of mineral particles, inputting a trained ElectroTrackNet model, and outputting a track path of the electrostatic images; wherein the ElectroTrackNet model is trained by the following steps: collecting annotated data with electrostatic images as a data set; training a neural network model through the data set, passing the input electrostatic images through multiple convolutional layers in the ElectroTrackNet network structure; the multiple convolutional layers extract features and key information of the electrostatic images layer by layer through local receptive fields; and then performing multi-scale feature extraction, and generating a predicted track output through the fused feature images. The present invention has the advantages of high efficiency, accuracy, and robustness when processing complex electrostatic image track segmentation tasks.

Description

Technical Field

[0001] The present invention relates to the technical fields of electrostatic separation and image processing, and in particular to a method and system for electrostatic separation particle trajectory recognition and sorting optimization based on deep learning, and also to a method and system for electrostatic separation trajectory recognition based on ElectroTrackNet. Background Art

[0002] In the electrostatic separation industry, accurate separation of mineral particles is crucial, but the existing electrostatic separation equipment and processes have not yet been fully intelligentized and still rely on traditional manual operation and empirical judgment. This traditional method is not only labor-intensive, but also unstable in separation efficiency and accuracy, and cannot meet the requirements of modern industry for efficient production and resource utilization. In addition, the large amount of data generated during the electrostatic separation process has not been fully utilized, and there is a lack of effective intelligent decision-making support tools, which has led to bottlenecks in electrostatic separation process optimization and efficiency improvement.

[0003] Existing electrostatic separation image processing and particle trajectory recognition technologies are mostly based on traditional image processing algorithms or classical machine learning models, but these methods have poor adaptability to the complex and changeable scenes in the electrostatic separation process, especially in the case of multi-scale target segmentation and complex background interference, the recognition accuracy is low.

[0004] For example, the patent document with the publication number CN117315277A discloses a local feature extraction method and system based on weakly supervised learning, which uses a convolutional neural network (CNN) with an attention mechanism, and trains the convolutional neural network based on image data sets from various sources and categories. The convolutional neural network includes multiple convolutional layers and pooling layers for automatically learning features in the image. During the training process, by linking the input image with the labels or constraints of its local features, the network gradually learns how to extract local features that meet the local feature screening requirements from the image; and then the trained convolutional neural network is used to process the image data to obtain local features that meet the local feature screening requirements corresponding to each image data. An embodiment in the patent document points out that in the convolutional neural network, a multi-scale attention mechanism is introduced, wherein the multi-scale attention mechanism allows the convolutional neural network to consider local features of different scales at the same time, and the multi-scale attention is usually achieved by introducing an attention mechanism in different convolutional layers to ensure that the network can obtain multi-scale information at different levels. However, the multi-scale attention mechanism can usually extract more feature information from the image by dividing it into multiple scales. However, since the trajectory recognition of the electronically selected image is under complex background interference, the interference of the local image increases after the multi-scale segmentation of the traditional multi-scale extraction mechanism, and there is no effective measure to fuse and retain the results of these multi-scale feature extractions, resulting in low recognition accuracy.

[0005] And the patent document with publication number CN116485666A discloses a method for removing image moiré noise based on multi-scale dynamic feature aggregation. In this patent document, the multi-scale feature extraction module mainly aggregates information of different scales through residual dense blocks and scale feature fusion modules. However, this method shows obvious deficiencies in the electric selection scene. It lacks effective capture of dynamic features, cannot adapt to the continuous trajectory changes in the electric selection process, and the fusion method of multi-scale information is relatively simple, resulting in poor recognition accuracy and time consistency of the model when dealing with complex electric selection backgrounds.

[0006] In addition, the existing model has weak time series processing capabilities and cannot effectively capture the dynamic changes in the electrostatic separation process, resulting in poor timing consistency of trajectory recognition and difficulty in meeting the needs of real-time online monitoring and sorting optimization.

[0007] To address the above problems, the present invention proposes an electro-selection track segmentation and recognition method called Electro TrackNet. By introducing a multi-scale feature extraction mechanism and a timing enhancement module, the target segmentation accuracy and temporal consistency of the model in complex electro-selection images are greatly improved. Summary of the invention

[0008] The present invention aims to provide an ElectroTrackNet electric selection track recognition method and system to solve the technical problems of low recognition accuracy and poor temporal consistency of track recognition in the prior art under multi-scale target segmentation and complex background interference.

[0009] To achieve the above object, the present invention provides an ElectroTrackNet electro-selection track recognition method, comprising the following steps:

[0010] Collect the electrostatic images of mineral particles, input them into the trained ElectroTrackNet model, and output the trajectory path of the electrostatic images;

[0011] The ElectroTrackNet model is trained through the following steps:

[0012] Collect annotated data with electroselected images as a dataset, where each electroselected image contains a trajectory path;

[0013] The neural network model is trained by using the data set, and the input electro-selection image passes through multiple convolutional layers in the ElectroTrackNet network structure; multiple convolutional layers perform layer-by-layer feature extraction and key information extraction on the electro-selection image through local receptive fields; multi-scale feature extraction, time series information introduction and fusion are then performed, and the predicted trajectory output is generated through the fused feature image.

[0014] Preferably, the annotation data of the electro-selection images in the data set include annotation data of the electro-selection images under different electro-selection conditions, noise interference, camera angles and ambient brightness conditions.

[0015] Preferably, the training process of the ElectroTrackNet model also includes a back-propagation stage, which includes: calculating the prediction error between the trajectory output predicted by the model and the actual trajectory through a joint loss function, and based on the prediction error calculated by the joint loss function, calculating the gradient of each weight and bias in the network through a back-propagation algorithm, and passing the gradient of the loss function relative to the weight of each layer of the network layer by layer through the chain rule.

[0016] Preferably, the training process of the ElectroTrackNet model also includes an optimization phase, using a preset optimizer to iteratively adjust and update various weights and biases in the network according to the gradient obtained by the back-propagation algorithm.

[0017] Preferably, the ElectroTrackNet model processes the input electro-selection image, including:

[0018] Feature extraction process: The input electro-selected image is used as a feature map, and convolution operations are performed through two or more consecutive convolution layers to reduce the spatial resolution and increase the number of channels; then the MixScale Block module is used to extract and fuse multiple multi-scale features; then convolution and downsampling are performed;

[0019] Feature fusion process: Perform multi-layer upsampling operations through multi-step convolution transposition to restore the spatial resolution of the feature map; interlace the multiple sets of feature maps obtained by multi-layer upsampling operations and multiple multi-scale feature extraction operations, and then use the MixScale Block module to extract and fuse multi-scale features on the interlaced splicing results; then adjust the number of channels and spatial resolution of the obtained feature map through convolution operations and output the fused feature map;

[0020] Output process: The fused feature map is segmented through the Mix Segment module to generate the final segmentation result and segmentation features.

[0021] Preferably, the processing of the input electro-selection image by the ElectroTrackNet model further includes:

[0022] Temporal dimension enhancement process: After combining the feature map of the current moment of the segmentation feature with the feature map of the previous moment of the previous segmentation feature, the multi-scale feature extraction and fusion operation is performed through the MixScale Block module. The operation result is spliced ​​with the output result of the feature extraction process and fused into the feature fusion process. The multi-scale feature extraction and fusion operation is performed together with the result of the interlaced splicing through the MixScale Block module.

[0023] Preferably, the fused feature map is segmented by a Mix Segment module, including:

[0024] The fused feature map is transformed through a convolutional layer and multiple MixScale Block modules are used to extract and fuse multi-scale features. The processing result of each MixScale Block module is residually connected with the processing result of the previous step; then, after multiple convolutional layers, the segmentation result is generated and output as a single category trajectory.

[0025] Preferably, the multi-scale feature fusion operation performed by the multi-scale feature fusion module includes: performing multiple multi-scale feature extraction and multi-scale convolution operations through the MixScale Block module to extract small-, medium- and large-scale features, and splicing and fusing feature maps of different scales.

[0026] Preferably, the MixScale Block module includes:

[0027] The number of channels of the input feature map is compressed from the first dimension to the second dimension while keeping the spatial resolution of the feature map unchanged; the feature map is divided into three parallel sub-branches through a shunt operation, and the feature map is subjected to dilated convolution operations with different dilation rates corresponding to small scale, medium scale and large scale through the three sub-branches respectively, and the dilated convolution operation results of the previous scale are spliced ​​before the dilated convolution operations of the medium scale and large scale respectively, and the results are fused and spliced ​​into a feature map with a second dimension with three times the number of channels, and then channel compression is performed to restore the number of channels to the first dimension of the original input.

[0028] Preferably, the joint loss function is:

[0029] ;

[0030] in, Is a weight coefficient used to adjust the proportion of Dice loss and Tversky loss in the total loss;

[0031] The Dice loss is:

[0032] ;

[0033] in, Represents the index of the sample, that is, the pixel or voxel currently being processed; Indicates the total number of samples; Indicates The true label value of each sample; Indicates The predicted value of samples;

[0034] The Tversky loss is:

[0035] ;

[0036] Tversky loss is set by and Penalize false positives and false negatives differently, and .

[0037] The present invention also provides an ElectroTrackNet electric selection track identification system, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of any of the above methods when executing the computer program.

[0038] The present invention has the following beneficial effects:

[0039] 1. The ElectroTrackNet electrostatic separation trajectory recognition method and system of the present invention intelligently analyzes and identifies the particle motion trajectory during the electrostatic separation process through deep learning, realizes real-time monitoring of the electrostatic separation equipment and intelligent optimization of the separation process, thereby improving the separation efficiency and accuracy, and promoting the intelligent development of the electrostatic separation industry.

[0040] 2. In the preferred embodiment, the electro-selection track recognition method and system of the ElectroTrackNet of the present invention improves the dilated convolution module and adaptive loss function that adapt to the characteristics of the electro-selection process, so that the model can still maintain efficient and accurate performance when processing complex backgrounds and variable targets. It can greatly improve the accuracy and real-time performance of electro-selection image processing.

[0041] In addition to the above-described purposes, features and advantages, the present invention has other purposes, features and advantages. The present invention will be further described in detail below with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] The drawings constituting a part of this application are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0043] Figure 1 It is a flow chart of the electro-selection track identification method of ElectroTrackNet according to a preferred embodiment of the present invention;

[0044] Figure 2 is a flow chart of the MixScale Block module processing process of the preferred embodiment of the present invention;

[0045] Figure 3 It is a flow chart of the MixSegment module processing process of the preferred embodiment of the present invention. DETAILED DESCRIPTION

[0046] The embodiments of the present invention are described in detail below with reference to the accompanying drawings, but the present invention can be implemented in many different ways as defined and covered by the claims.

[0047] Figure 1 : is a flow chart of the electrotrack identification method of ElectroTrackNet in this embodiment. The electrotrack identification method of ElectroTrackNet of the present invention comprises the following steps:

[0048] Collect the electrostatic images of mineral particles, input them into the trained ElectroTrackNet model, and output the trajectory path of the electrostatic images.

[0049] The ElectroTrackNet model is trained through the following steps:

[0050] The labeled data with electroselection images are collected as a data set, wherein each electroselection image includes a trajectory path; during implementation, the labeled data of the electroselection images in the data set include the labeled data of the electroselection images under different electroselection conditions, noise interference, camera angles and ambient brightness conditions.

[0051] The neural network model is trained through the data set, and the input electro-selection image is passed through multiple convolutional layers in the ElectroTrackNet network structure; multiple convolutional layers extract features and key information of the electro-selection image layer by layer through the local receptive field; then multi-scale feature extraction, time series information introduction and fusion are performed, and the predicted trajectory output is generated through the fused feature image. Multi-scale feature extraction, time series information introduction and fusion can refine the segmentation task. The ElectroTrackNet network achieves efficient and accurate feature extraction and enhancement of the time dimension by applying the MixScale Block module multiple times and combining convolution, upsampling and feature splicing operations.

[0052] When implemented, the ElectroTrackNet model processes the input electro-selection image mainly including:

[0053] Feature extraction process: The input electro-selected image is used as a feature map, and a convolution operation is performed through two or more consecutive convolution layers to reduce the spatial resolution and increase the number of channels; then multiple multi-scale feature extraction and fusion operations are performed through the MixScale Block module; then convolution and downsampling are performed. During implementation, a multi-scale feature fusion operation is performed through a multi-scale feature fusion module, including: multiple multi-scale feature extraction and multi-scale convolution operations are performed through the MixScale Block module to extract small, medium, and large scale features, and splice and fuse feature maps of different scales. Among them, the MixScale Block module includes:

[0054] The number of channels of the input feature map is compressed from the first dimension to the second dimension while keeping the spatial resolution of the feature map unchanged; the feature map is divided into three parallel sub-branches through a shunt operation, and the feature map is subjected to dilated convolution operations with different dilation rates corresponding to small scale, medium scale and large scale through the three sub-branches, and then fused and spliced ​​into a feature map with three times the number of channels in the second dimension, and then channel compression is performed to restore the number of channels to the first dimension of the original input.

[0055] Feature fusion process: Perform multi-layer upsampling operations through multi-step convolution transposition to restore the spatial resolution of the feature map; interleave the multiple sets of feature maps obtained by multi-layer upsampling operations and multiple multi-scale feature extraction operations, and then use the MixScale Block module to extract and fuse multi-scale features on the results of the interleaved splicing; then adjust the number of channels and spatial resolution of the obtained feature map through the convolution operation and output the fused feature map.

[0056] When implemented, the ElectroTrackNet model processes the input electro-selection image and also includes:

[0057] Temporal dimension enhancement process: After combining the feature map of the current moment of the segmentation feature with the feature map of the previous moment of the previous segmentation feature, the multi-scale feature extraction and fusion operation is performed through the MixScale Block module. The operation result is spliced ​​with the output result of the feature extraction process and fused into the feature fusion process. The multi-scale feature extraction and fusion operation is performed together with the result of the interlaced splicing through the MixScale Block module.

[0058] Output process: The fused feature map is segmented through the Mix Segment module to generate the final segmentation result and segmentation features.

[0059] The following is a detailed description with reference to the accompanying drawings:

[0060] 1. Feature extraction module (nodes 1-0 to 1-9).

[0061] The input features first pass through two consecutive 3×3 convolutional layers (nodes 1-0 and 1-1), where the stride of the first convolution is 1 and the stride of the second convolution is 2, to reduce the spatial resolution of the feature map and increase the number of channels. Then, after passing through the first MixScale Block module (node ​​1-2), the input features are subjected to multi-scale convolution operations to extract small, medium, and large scale features, and feature maps of different scales are fused by splicing. Subsequently, the feature map passes through the third 3×3 convolution (node ​​1-3) and is downsampled again with a stride of 2.

[0062] The feature extraction module includes a multi-scale feature fusion module (nodes 1-4 to 1-9).

[0063] After downsampling, the feature map is sequentially passed through the MixScale Block module (nodes 1-4 and nodes 1-6) for multiple multi-scale feature extraction. Each MixScale Block module expands the receptive field through dilated convolution while maintaining computational efficiency. At nodes 1-5 and 1-7, further downsampling is performed through 3×3 convolution with a stride of 2 to gradually reduce the spatial resolution of the feature map. Nodes 1-9 use the SPPF module (spatial pyramid pooling) to compress and fuse the spatial features of the feature map while retaining the global context information.

[0064] 2. Upsampling and feature fusion module (nodes 1-10 to 1-20).

[0065] After feature extraction, the network performs upsampling through a series of convolution transposes (ConvTranspose, nodes 1-10, nodes 1-12, nodes 1-13) to restore the spatial resolution of the feature map. Node 1-11 concatenates the feature map of node 1-6 with the upsampling result of node 1-10 through the Concat operation, enhancing the detail information of the feature map. Subsequently, the network concatenates the feature maps from nodes 1-4 and 1-13 through another Concat operation at node 1-14, and further fuses multi-scale features through MixScale Block (node ​​1-15). The 3×3 convolution of node 1-16 is used to adjust the number of channels and spatial resolution of the concatenated feature map.

[0066] In the subsequent upsampling process, the Concat operation of node 1-17 concatenates the features of node 1-12 with node 1-16, and the MixScale Block of node 1-18 further performs multi-scale fusion on the concatenated feature map. The 3×3 convolution of node 1-19 is responsible for the final fusion and refinement of the feature map. The Concat operation of node 1-20 concatenates the features of node 1-26 with node 1-19, and the MixScale Block of node 1-21 further performs multi-scale fusion on the concatenated feature map.

[0067] The above steps include cross-layer splicing of multiple sets of feature maps obtained by multi-layer upsampling operations and multiple multi-scale feature extraction operations through staggered splicing to fuse cross-layer features. The Concat operation is used in this network structure to fuse features between different layers, ensuring that the model does not lose important detail information during the downsampling process, and can pass detail information in the high-resolution layer to the deep features of the low-resolution layer, thereby realizing multi-scale fusion of features. Each Concat splices feature maps from different resolutions to integrate information of different scales. This splicing operation can enhance the network's expressiveness and improve its perception of details and global information.

[0068] Cross-layer feature fusion in this embodiment:

[0069] Whenever the resolution of a feature map decreases through convolutional layers and downsampling operations in the network, some layers (such as 1-4, 1-6, 1-8) will save these low-resolution feature maps until higher layers (such as 1-10, 1-12) perform upsampling and then concatenate these low-resolution features with the upsampled features through the Concat operation.

[0070] For example: After the two upsampling operations of 1-12 and 1-13, they are concatenated with the high-resolution features from 1-4 to retain the detail information and enhance the expression of deep features.

[0071] Multi-scale feature concatenation after upsampling:

[0072] After the network obtains a larger receptive field through the downsampling layer, the low-resolution feature map is restored to a higher resolution through upsampling (such as ConvTranspose layer, 1-12, 1-13, 1-10, etc.), and is spliced ​​with the feature map from the shallower layer (high-resolution path). For example: 1-12 or 1-13 upsample the upper layer features respectively, and then splice with the high-resolution feature map from the shallower layer (1-4) at 1-11 or 1-14.

[0073] These operations combine local high-resolution features with global low-resolution deep features, allowing the model to enhance spatial information without losing deep semantic information.

[0074] Deep and shallow combination of features:

[0075] The network forms a feature expression that combines shallow and deep features by concatenating high-resolution features from shallow layers (such as 1-4, 1-6) with low-resolution features from deep layers (such as 1-10, 1-13).

[0076] For example, at 1-20, features from the deep layers are concatenated with upsampled high-resolution features, allowing the network to take into account both global semantic information (deep layers) and local detail information (shallow layers). This ensures that the network can process both details and global context while gradually generating higher resolution outputs.

[0077] Repeat the Concat connection multiple times:

[0078] This Concat operation does not only occur once, but is repeated many times at different stages of the network structure. Almost every upsampling stage is accompanied by a high-resolution splicing operation. For example, at levels 1-10 and 1-14, deep features are upsampled back to high resolution and spliced ​​with shallow features. 1-20 is spliced ​​again to ensure that feature information can be gradually integrated and strengthened.

[0079] This Concat staggered splicing design enables the network to simultaneously integrate high-level semantic information and low-level detail information by splicing features between feature maps at different resolution levels, thus enhancing the expressiveness of the model. This mechanism is particularly suitable for tasks that require precise positioning and detail processing, such as image segmentation. By retaining multi-scale information, the network can accurately capture local and global feature information while ensuring resolution.

[0080] 3. Time dimension enhancement module (nodes 1-24 to 1-26).

[0081] Node 1-24 introduces time scale information by combining the segmentation features of the previous moment (that is, the additional feature information representing the output result of the last output, output by node 1-23) with the feature map of the current moment, thereby enhancing the network's ability to capture time series features. This design combines historical information with the features of the current frame to help the model model the changes of the target in the time dimension, thereby enhancing the temporal consistency and accuracy of the segmentation results.

[0082] In addition, the historical segmentation information introduced by nodes 1-24 can help the model better process long time series data, especially in scenarios where the target changes dynamically, the shape is blurred or temporarily occluded. The model can make more robust judgments based on the correlation of the time dimension, significantly improving the accuracy and stability of the segmentation. By introducing time scale information through nodes 1-24, this network can effectively use historical time series information to enhance the continuity and accuracy of segmentation while maintaining the advantages of spatial feature extraction. Especially in dynamic scenes such as video streaming, this fusion design of time features can significantly improve the temporal consistency of the model and reduce jitter or mis-segmentation. In addition, the MixScale Block module is applied multiple times, which enables the model to perform well in processing multi-scale targets. Combined with upsampling and feature splicing operations, efficient feature extraction and fusion are achieved.

[0083] After combining the temporal information, nodes 1-25 further optimize the joint features through a 3×3 convolution. The generated feature map is further processed by MixScale Block (nodes 1-26) and then undergoes another convolution operation to generate a refined feature map at the current moment.

[0084] The time dimension enhancement module mainly introduces and integrates time series information to enhance the time series features. The time series features are expressed in the feature graph as follows:

[0085] Dynamically changing trajectory information: The model captures the trajectory continuity and motion changes in the electrical selection image by splicing and fusing feature maps at different times. This feature reflects the motion trend and morphological changes of the target in the time dimension.

[0086] Temporal-dependent contextual information: Temporal union not only enhances the segmentation effect of a single frame, but also brings richer spatiotemporal context, enabling the model to correct the segmentation result at the current moment based on the prediction at the previous moment, thus avoiding mis-segmentation or missed segmentation.

[0087] ElectroTrackNet's temporal joint operation enhances the model's ability to capture dynamic trajectories by fusing features at different times. At the same time, it extracts multi-scale features through the MixScale Block module, enabling the model to handle complex changes in electro-selection trajectories and generate accurate segmentation results.

[0088] 4. Output module (nodes 1-21 to 1-23).

[0089] After the feature map completes the fusion of multi-level features, it is processed by MixScale Block (node ​​1-21) and sent to MixSegment module (node ​​1-22) to generate the final segmentation result (outSeg). At the same time, the segmentation feature (outF) generated by node 1-23 provides additional feature information for the model, further enhancing the output performance of the model.

[0090] Final feature fusion output: In the final stage, the network passes the fused features to the output module (outF, outSeg) through a series of Concat and MixScaleBlock operations. This means that all features processed at different scales are completely preserved and fused before the final output, so that the output result has multi-scale feature information.

[0091] Figure 2 is the structure of the MixScale Block module. Figure 2 The MixScale Block shown ( Figure 1 1-2, 1-4, 1-6, 1-8, 1-15, 1-26, 1-18 or 1-21) in is one of the core components of the present invention. It adopts a multi-branch and multi-scale dilated convolution (hole convolution) design, which significantly improves the model's ability to capture and express features of different scales. This module first performs preliminary processing on the input feature map, using a 1×1 convolution (node ​​2-0) to compress the number of channels to the required dimension while keeping the spatial resolution of the feature map unchanged. Assume that the size of the input feature map is , after 1×1 convolution, the size of the output feature map becomes ,in is the number of channels after compression.

[0092] After the initial processing of 1×1 convolution, the feature map is divided into three parallel sub-branches, namely Sub1, Sub2 and Sub3, through a split operation (node ​​2-1). These three branches process different scale information of the feature map respectively.

[0093] Branch 1 (Sub1, node 2-2) first further compresses the number of channels and fuses features through a 1×1 convolution (node ​​2-5), and then undergoes a 5×5 dilated convolution (node ​​2-6) with a dilation rate of 1. This convolution operation can expand the receptive field, allowing the model to enhance the capture of small-scale local features while maintaining a high resolution. Subsequently, the feature map is further fused through another 1×1 convolution (node ​​2-7), and the output size is , at this time, the features have been extracted and optimized on a small scale.

[0094] The processing of branch 2 (Sub2, nodes 2-3) is similar to that of branch 1. It first undergoes a 1×1 convolution (nodes 2-8) for preliminary channel compression. Subsequently, the feature map undergoes a 7×7 asymmetric dilated convolution (nodes 2-9) with a dilation rate of 2. This convolution significantly expands the receptive field through convolution kernels of different sizes in the horizontal and vertical directions, effectively capturing medium-scale feature information, especially fine-grained medium-sized objects. The feature map is then fused through a 1×1 convolution (nodes 2-10), and the output also maintains size.

[0095] Branch 3 (Sub3, nodes 2-4) performs channel compression on the input feature map through a 1×1 convolution (node ​​2-11), and then extracts features through a 9×9 dilated convolution (node ​​2-12) with a dilation rate of 3. This convolution kernel is large and suitable for capturing large-scale global feature information, especially when processing large objects and global context information. After the dilated convolution, the feature map is fused again through a 1×1 convolution (node ​​2-13), and the output size is .

[0096] After the three branches complete their respective multi-scale feature extraction, all feature maps are concatenated in the feature connection operation (nodes 2-14) to form a channel number of The feature map combines multi-scale information from different branches. At this point, the model can effectively utilize small, medium, and large-scale feature information to achieve comprehensive perception of complex targets.

[0097] Finally, the fused feature map is compressed through a 1×1 convolution to restore the number of channels to the number of channels C of the original input, generating a size of This operation not only completes the effective fusion of multi-scale features, but also further refines the representation of the feature map through convolution operations, providing a more recognizable feature expression for the processing of subsequent network layers.

[0098] Through the clever combination of multi-scale dilated convolution and 1×1 convolution, MixScale Block can greatly expand the receptive field of the model while maintaining computational efficiency, achieving efficient detection and representation of objects of different scales. This design is particularly effective in multi-scale object detection tasks and significantly improves the performance of the model.

[0099] During implementation, the fused feature map is segmented through the Mix Segment module, including:

[0100] The fused feature map is transformed through a convolutional layer and multiple MixScale Block modules are used to extract and fuse multi-scale features. The processing result of each MixScale Block module is residually connected with the processing result of the previous step; then, after multiple convolutional layers, the segmentation result is generated and output as a single category trajectory. Figure 3 This is the structural diagram of the Mix Segment module.

[0101] like Figure 3 The MixSegment network structure shown in the figure is designed specifically for processing single-category trajectory tracking of electrical selection images. The network first performs preliminary feature conversion through a 1×1 convolution layer, and then enters the MixScale Block module to extract and fuse multi-scale features. The feature map is processed by multiple MixScale Blocks, and the features are effectively transferred and enhanced through residual connections (addition operations). At the end of the network, after a series of 1×1 convolution layers, a refined single-category trajectory output is generated to meet the precise trajectory tracking requirements of electrical selection images.

[0102] The loss function of the ElectroTrackNet model in this example is:

[0103] In order to further improve the performance of the track segmentation of the electrical image, the present invention proposes to combine the Dice loss with the Tversky loss and design a joint loss function to better balance the conflict between the complexity of the background and the small proportion of the target area. The combination of the two can give full play to the robustness of the Dice loss in the overall segmentation area and the flexibility of the Tversky loss in processing category imbalanced data, thereby enhancing the model's ability to detect small targets.

[0104] The Dice loss aims to maximize the overlap between the model prediction results and the true label, and improves the model's segmentation effect on small targets by focusing on the target area. The Tversky loss introduces a weight parameter and , differently penalize false positives and false negatives, and further optimize the problem of small objects and background imbalance.

[0105] In order to combine the advantages of both, the joint loss function proposed in the present invention can be expressed as:

[0106] ;

[0107] in, is a weight coefficient used to adjust the proportion of Dice loss and Tversky loss in the total loss. The value can flexibly adjust the impact of the loss function according to the characteristics of the target and background in the data set to achieve the optimal segmentation effect.

[0108] Dice loss:

[0109] ;

[0110] in, Represents the index of the sample, that is, the pixel or voxel currently being processed; Indicates the total number of samples; Indicates The true label value of each sample; Indicates This part of the loss focuses on optimizing the segmentation quality of the model for the target area and maximizing the overlap between the true trajectory pixels and the predicted trajectory pixels.

[0111] Tversky losses:

[0112] ;

[0113] Tversky loss is set by and Different penalties are imposed on false positives and false negatives, which is particularly suitable for solving the problem of small targets and confusion with complex backgrounds. For track segmentation of electrostatic images, we usually choose To reduce the impact of false positives (background misdetection) and enhance the detection of small targets.

[0114] The advantages of the joint loss function are as follows:

[0115] Dealing with category imbalance: Dice loss focuses on optimizing the overlap of target areas, while the false positive and false negative adjustment mechanism of Tversky loss can better balance the ratio of target and background areas, which is especially suitable for situations where the track ratio in electrostatic images is small and the background is complex.

[0116] Improve small target detection performance: The target tracks in the electrical selection images are usually small, so the dynamic adjustment mechanism of Tversky loss can more effectively suppress background noise and improve the segmentation accuracy of the track area.

[0117] Enhanced model convergence and robustness: Through the design of joint loss, the model can converge more quickly during training and maintain good robustness in unbalanced scenarios, avoiding overfitting of background information.

[0118] By combining Dice loss and Tversky loss, the joint loss function proposed in the present invention not only effectively addresses the class imbalance problem between small targets and complex backgrounds, but also enhances the model's ability to capture trajectory areas, providing a strong guarantee for improving segmentation accuracy.

[0119] During implementation, the training steps of the ElectroTrackNet model of the present invention are as follows. Based on its task of tracking electrostatic image trajectories, it is assumed that the input data is electrostatic image trajectories with a single category.

[0120] Data preparation phase: includes data collection and data preprocessing.

[0121] Data collection: Collect a large number of annotated datasets with electrostatic images to ensure that each image contains a clear trajectory path. The dataset should cover different electrostatic conditions, noise interference, different camera angles, ambient brightness, etc. to ensure the generalization ability of the model.

[0122] Data preprocessing: Perform uniform resizing and normalization on the images (for example, scaling pixel values ​​to between [0,1]), and divide them into training sets, validation sets, and test sets. To increase data diversity, data augmentation operations can be performed, such as random cropping, rotation, flipping, etc.

[0123] Network initialization phase:

[0124] Use the ElectroTrackNet model architecture.

[0125] Choose a suitable weight initialization method, such as Xavier or He initialization, to ensure that the initial distribution of network weights is conducive to rapid convergence of the model.

[0126] Set the hyperparameters in the network:

[0127] Learning rate (set according to the task, such as 0.001 or 0.0001);

[0128] Optimizer (commonly used Adam or SGD optimizer);

[0129] The number of iterations (epochs) is usually set to 50~100 times, which is adjusted according to the model performance.

[0130] The following methods are used in model training:

[0131] Forward propagation: During the forward propagation process, the input electro-selected image is first fed into the network model. Inside the model, the electro-selected image passes through multiple convolutional layers in the ElectroTrackNet network structure. These convolutional layers extract features from the image layer by layer through local receptive fields and gradually refine the key information in the image. Next, the image features are further multi-scale fused through the MixScaleBlock module to capture the detailed features at different scales in the image. Finally, these fused features are input to the last layer of the network, and after further processing, the predicted trajectory output is generated.

[0132] Loss calculation: In order to further improve the performance of trajectory segmentation of electrical selection images, this paper proposes a method that combines Dice loss and Tversky loss. This strategy aims to better balance the conflict between background complexity and the small proportion of the target area. Specifically, Dice loss is mainly used to evaluate the robustness of the model in the overall segmentation area. It can effectively measure the degree of overlap between the predicted area and the true labeled area, thereby improving the accuracy of segmentation overall. However, when faced with data with unbalanced categories, Dice loss may not be able to fully handle the problem of a small proportion of the target area. For this reason, Tversky loss is introduced as a supplement. It has high flexibility and can adjust the penalty weights for false positives and false negatives according to actual needs, thereby effectively dealing with the problem of unbalanced categories.

[0133] By combining Dice loss with Tversky loss, the model can fully utilize the robustness of Dice loss in the overall segmentation area, while leveraging Tversky loss's ability to handle imbalanced data. This combination strategy not only improves the model's ability to detect small targets, but also enhances the overall segmentation performance, making the model more accurate and reliable when processing electrical image trajectory segmentation tasks.

[0134] Back propagation: In the back propagation phase, based on the prediction error calculated by the loss function, the back propagation algorithm is used to calculate the gradients of each weight and bias in the network (wx+b, commonly used in neural networks, can be used, where w and b are weights and biases). Specifically, the gradient of the loss function relative to the weights of each layer of the network is passed layer by layer through the chain rule. These gradients reflect the sensitivity of the loss function to the weights, which can guide how to adjust the network weights to reduce the prediction error.

[0135] Optimization: During the optimization phase, a preset optimizer (such as the Adam optimizer) is used to adjust the network weights based on the gradients calculated by the back-propagation algorithm. The optimizer gradually updates the weight values ​​to optimize the performance of the model by combining gradient information and strategies such as momentum. The Adam optimizer uses an adaptive learning rate to adjust the pace of weight updates, thereby effectively improving the convergence speed and prediction accuracy of the model. The goal of the optimization process is to gradually reduce the model prediction error and improve the model's performance in the task of predicting the trajectory of electrical images.

[0136] Verification and parameter adjustment.

[0137] Validation set evaluation: After each training cycle (epoch), it is crucial to evaluate the model using the validation set data. This process includes calculating and monitoring the model's loss value on the validation set and several evaluation metrics, such as accuracy and intersection over union (IoU). Validation loss can provide an indication of the model's generalization ability on unseen data, while metrics such as accuracy and IoU further reflect the model's performance on object detection and segmentation tasks. This evaluation process helps to promptly identify potential problems in the model during training and ensure that the model performs well in actual application scenarios.

[0138] Early stopping strategy: To avoid overfitting of the model, the early stopping strategy is an effective method. When the loss on the validation set no longer decreases in several consecutive cycles, it is considered that the learning of the model has reached a bottleneck, and further training may lead to overfitting. The early stopping strategy sets a tolerance (such as the size of the monitoring window) and stops training if the validation loss no longer improves within the window. This method not only helps save computing resources, but also ensures the generalization ability of the model and avoids performance degradation caused by overtraining.

[0139] Adjusting hyperparameters: During the training process, dynamically adjusting hyperparameters based on the performance of the model is the key to optimizing model performance. Adjusting hyperparameters such as learning rate, batch size, or optimizer parameters based on the performance of the model on the validation set can significantly affect the training results. For example, if the model converges too slowly or the loss drop is unstable, you may need to adjust the learning rate; if the training process is too unstable, you may need to adjust the batch size. In addition, if necessary, you can redesign the model structure or introduce regularization methods (such as L1 or L2 regularization, dropout, etc.) to further improve the performance and generalization ability of the model. Dynamically adjusting these hyperparameters and structures can help the model better adapt to the data characteristics and achieve optimal performance.

[0140] Test model:

[0141] Final test evaluation: After the model training is completed, it is a crucial step to use the test set data to perform a final evaluation of the model. This evaluation is designed to verify the generalization ability of the model and ensure that its performance on unseen data is consistent with that on the training and validation sets. By evaluating on the test set, an objective indicator of the true performance of the model can be obtained.

[0142] Depending on the needs of specific tasks, the main evaluation indicators may include but are not limited to:

[0143] Accuracy: This metric measures the proportion of samples correctly predicted by the model to the total number of samples. In classification tasks, accuracy can reflect the overall accuracy of the model.

[0144] Intersection over Union (IoU): This metric is used to measure the degree of overlap between the predicted area and the true annotated area. Especially in object detection and segmentation tasks, IoU is an important criterion for evaluating the quality of model predictions, and can provide feedback on the model's local details and regional consistency.

[0145] Through these evaluation indicators, we can fully understand the actual performance of the model and judge whether it meets the requirements and expectations of the task. The results of the test set evaluation not only help to confirm the effectiveness of the model in practical applications, but also provide data support for further model optimization and adjustment.

[0146] Model optimization and deployment.

[0147] The trained model is deployed in the actual electric selection control system to ensure that it can process images in real time and perform accurate trajectory tracking.

[0148] The present invention also provides an ElectroTrackNet electric selection track identification system, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of any of the above embodiments when executing the computer program.

[0149] In summary, the electro-selection track recognition method and system using the ElectroTrackNet of the present invention have the following beneficial effects:

[0150] Multi-scale feature extraction and fusion: ElectroTrackNet achieves efficient extraction and fusion of small, medium and large scale features by applying MixScale Block modules multiple times. In particular, the combination of dilated convolution expands the receptive field and improves the model's ability to capture multi-scale targets in complex electro-selection images.

[0151] Temporal dimension enhancement: By introducing the segmentation results of historical moments and combining them with the characteristics of the current moment, the model's ability to capture time series features is enhanced, effectively improving the temporal consistency of electrical trajectory recognition, which is particularly suitable for processing long time series data and dynamically changing targets.

[0152] Accurate feature extraction and upsampling: The convolution, upsampling and feature concatenation operations enable the model to gradually restore the spatial resolution of the feature map while reducing the computational cost, thus ensuring segmentation accuracy and output quality.

[0153] Efficient trajectory segmentation and tracking: With the help of the MixSegment module, the present invention can accurately track the electrical selection trajectory for the single category trajectory segmentation task of the electrical selection image, ensure the refinement and accuracy of trajectory recognition, and adapt to the needs of small target detection in complex backgrounds.

[0154] Adaptive loss function design: Combining Dice loss and Tversky loss, it balances the contradiction between complex background and target area, improves the model's ability to detect small targets, and makes the segmentation effect more robust, especially in the case of large background noise.

[0155] Improve the convergence and robustness of the model: Through joint loss design, the model converges faster during training and can adapt to unbalanced data scenarios, reduce mis-segmentation and jitter, and enhance the stability and reliability of the model.

[0156] These features make the present invention highly efficient, accurate, and robust when processing complex electrostatic image trajectory segmentation tasks, and can maintain excellent temporal consistency and segmentation accuracy in dynamic scenes.

[0157] The present invention solves the urgent need for automation and intelligent technology upgrades in the field of electrostatic separation, as well as the shortcomings of existing technologies in sorting efficiency and intelligent decision support. The innovative solution proposed by the present invention can greatly improve the accuracy and real-time performance of electrostatic separation image processing, provide intelligent decision support for the electrostatic separation process, and promote the intelligent and automated upgrade of the electrostatic separation industry. This not only helps to improve resource utilization and reduce production costs, but also ensures product quality, helping industrial production move towards a more intelligent and efficient direction.

[0158] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. An ElectroTrackNet electro-selection track recognition method, characterized in that: The following steps are involved: Collect the electrostatic images of mineral particles, input them into the trained ElectroTrackNet model, and output the trajectory path of the electrostatic images; The ElectroTrackNet model is trained through the following steps: Collect annotated data with electroselected images as a dataset, where each electroselected image contains a trajectory path; The neural network model is trained by the data set, and the input electro-selection image is passed through multiple convolutional layers in the ElectroTrackNet network structure; the multiple convolutional layers extract features and key information of the electro-selection image layer by layer through the local receptive field; then multi-scale feature extraction, time series information introduction and fusion are performed, and then the fused feature map is segmented to generate the final segmentation result and segmentation feature, and the segmentation result is output as the predicted trajectory; The ElectroTrackNet model processes the input electro-selection image, and further includes: Temporal dimension enhancement process: after combining the feature map of the current moment of the segmentation feature with the feature map of the previous moment of the last segmentation feature, the multi-scale feature extraction and fusion operation is performed through the MixScale Block module, and the operation result is spliced ​​with the output result of the feature extraction process, and fused into the feature fusion process, and the multi-scale feature extraction and fusion operation is performed together with the result of the interlaced splicing through the MixScale Block module; The MixScale Block module includes: The number of channels of the input feature map is compressed from the first dimension to the second dimension while keeping the spatial resolution of the feature map unchanged; the feature map is divided into three parallel sub-branches through a shunt operation, and the feature map is subjected to dilated convolution operations with different dilation rates corresponding to small scale, medium scale and large scale through the three sub-branches respectively, and the dilated convolution operation results of the previous scale are spliced ​​before the dilated convolution operations of the medium scale and large scale respectively, and the results are fused and spliced ​​into a feature map with a second dimension with three times the number of channels, and then channel compression is performed to restore the number of channels to the first dimension of the original input.

2. The method for identifying the electrotrack of ElectroTrackNet according to claim 1, characterized in that: The annotated data of the electro-selected images in the data set include the annotated data of the electro-selected images under different electro-selected conditions, noise interference, camera angles and environmental brightness conditions.

3. The method for identifying the electrotrack track of ElectroTrackNet according to claim 1 or 2, characterized in that: The training process of the ElectroTrackNet model also includes a back-propagation stage, which includes: calculating the prediction error between the trajectory output predicted by the model and the actual trajectory through a joint loss function, and based on the prediction error calculated by the joint loss function, calculating the gradient of each weight and bias in the network through a back-propagation algorithm, and passing the gradient of the loss function relative to the weight of each layer of the network layer by layer through the chain rule.

4. The method for identifying the electrotrack of ElectroTrackNet according to claim 3, characterized in that: The training process of the ElectroTrackNet model also includes an optimization phase, which uses a preset optimizer to iteratively adjust and update the various weights and biases in the network according to the gradient obtained by the back-propagation algorithm.

5. The method for identifying the electrotrack track of ElectroTrackNet according to claim 1, characterized in that: The ElectroTrackNet model processes the input electro-selection image, including: Feature extraction process: The input electro-selected image is used as a feature map, and convolution operations are performed through two or more consecutive convolution layers to reduce the spatial resolution and increase the number of channels; then the MixScale Block module is used to extract and fuse multiple multi-scale features; then convolution and downsampling are performed; Feature fusion process: Perform multi-layer upsampling operations through multi-step convolution transposition to restore the spatial resolution of the feature map; interlace the multiple sets of feature maps obtained by multi-layer upsampling operations and multiple multi-scale feature extraction operations, and then use the MixScale Block module to extract and fuse multi-scale features on the interlaced splicing results; then adjust the number of channels and spatial resolution of the obtained feature map through convolution operations and output the fused feature map; Output process: The fused feature map is segmented through the Mix Segment module to generate the final segmentation result and segmentation features.

6. The method for identifying the electrotrack of ElectroTrackNet according to claim 5, characterized in that: The fused feature map is segmented through the Mix Segment module, including: The fused feature map is subjected to feature conversion through a convolutional layer and multi-scale feature extraction and fusion through multiple MixScale Block modules. The processing result of each MixScale Block module is subjected to a residual connection operation with the processing result of the previous step. Then, after passing through multiple convolutional layers, a segmentation result is generated and output as a single category trajectory.

7. The method for identifying the electrotrack of ElectroTrackNet according to claim 3, characterized in that: The joint loss function is: ； in, Is a weight coefficient used to adjust the proportion of Dice loss and Tversky loss in the total loss; The Dice loss is: ； in, Represents the index of the sample, that is, the pixel or voxel currently being processed; Indicates the total number of samples; Indicates The true label value of each sample; Indicates The predicted value of samples; The Tversky loss is: ； Tversky loss is set by and Penalize false positives and false negatives differently, and .

8. An ElectroTrackNet electric selection track recognition system, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the method described in any one of claims 1 to 7 when executing the computer program.

Images