A residual manganese detection method for manganese electrolytic plate based on multi-expert convolution
By employing a multi-expert convolution detection method, the problems of insufficient multi-scale feature extraction and dynamic adaptability in the detection of residual manganese in manganese electrolysis plates are solved, achieving high-precision, real-time detection of residual manganese, which is suitable for industrial production lines.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHANGSHA RES INST OF MINING & METALLURGY CO LTD
- Filing Date
- 2026-01-29
- Publication Date
- 2026-07-03
AI Technical Summary
Existing technologies for detecting residual manganese in manganese electrolysis plates suffer from problems such as insufficient multi-scale feature extraction, poor dynamic adaptability, incomplete detection results, low computational efficiency, and insufficient multi-task optimization. In particular, it is difficult to achieve high accuracy and high efficiency in the detection of small defects and in cases of sample imbalance.
A detection method based on multi-expert convolution is adopted. By combining depthwise separable convolutional blocks, hybrid convolutional blocks, and multi-scale fusion structures with multi-head attention and routing selection, the optimal convolutional kernel is dynamically selected to achieve multi-scale feature extraction and multi-task joint optimization. The output bounding boxes and segmentation masks are used to construct the detection model.
It improves the detection accuracy of long strips, small blocks and tiny residues, enhances the adaptability and robustness of the detection results, reduces computational complexity, and meets the real-time deployment requirements of industrial production lines.
Smart Images
Figure CN121599981B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of artificial intelligence and computer vision technology, and in particular to the detection and identification of residual manganese in the production process of manganese electrolysis plates. Background Technology
[0002] In the industrial production of manganese electrolytic plates, the detection and identification of residual manganese is a crucial step in ensuring product quality and production efficiency. If residual manganese is not detected and removed in a timely manner, it will not only affect the conductivity and surface uniformity of the electrolytic plate, but may also lead to equipment wear, short-circuit risks, and decreased product performance in subsequent processes. Therefore, how to achieve efficient and accurate detection of residual manganese has always been a critical problem that needs to be solved in the field of industrial testing.
[0003] With the development of artificial intelligence and deep learning, convolutional neural networks (CNNs) are increasingly being applied to industrial defect detection. CNNs have strong capabilities in feature extraction and can automatically learn spatial features in images. However, in the scenario of detecting residual manganese in manganese electrolytic plates, existing CNN-based methods still have shortcomings: their feature extraction usually relies on convolutional kernels of fixed size, making it difficult to simultaneously capture multi-scale defects such as long strips, blocks, and tiny particles; at the same time, existing methods mostly use fixed feature extraction paths, lacking dynamic adaptability, and are prone to false detections or missed detections in complex backgrounds.
[0004] Regarding the detection results, some methods only output bounding box location information, lacking fine segmentation of the target area. While the bounding box can provide the approximate location of the target, it cannot accurately depict the true shape of the residue, limiting subsequent quality assessment and process improvement. The single-task output leads to inconsistencies in the spatial location and morphological description of the detection results, making it difficult to meet the needs of industrial scenarios for defect morphology analysis and process optimization.
[0005] In recent years, the YOLO series (such as YOLOv5 and YOLOv8) has been widely used in industrial defect detection. They demonstrate high accuracy and real-time performance in target detection tasks, making them particularly suitable for rapid inspection on production lines. However, shortcomings remain in complex scenarios such as residual manganese detection in manganese electrolysis plates: YOLOv5 is prone to missed or false detections when detecting defects with small areas and blurred edges; while YOLOv8 has structural improvements, its detection accuracy is still insufficient when dealing with extremely small targets or low-contrast defects. Furthermore, industrial defect data often exhibits class imbalance, and the YOLO series tends to favor the dominant class in imbalanced sample conditions, leading to low recognition rates for minority class defects; its feature fusion structure does not adequately transfer information when processing complex industrial images, making it difficult to adapt to different forms of residue; simultaneously, the large number of parameters and high computational complexity in high-precision mode necessitate further optimization for lightweight deployment.
[0006] In summary, existing technologies have shortcomings in multi-scale feature extraction, dynamic adaptability, detection result integrity, computational efficiency, and multi-task joint optimization. Even the current mainstream YOLOv5 and YOLOv8 still have limitations in small defect detection, sample imbalance handling, feature fusion capabilities, and lightweight deployment, necessitating a new technical solution to address these issues. Summary of the Invention
[0007] This invention provides a method for detecting residual manganese in manganese electrolysis plates based on multi-expert convolution, in order to solve the problem that existing visual algorithms cannot accurately identify residual manganese in manganese electrolysis plates.
[0008] To achieve the above objectives, the present invention employs the following technical solution:
[0009] This invention provides a method for detecting residual manganese in a manganese electrolytic plate based on multi-expert convolution, comprising the following steps:
[0010] Step 1: Construct the first branch based on depthwise separable convolutional blocks combined with routing multilayer awareness and convolutional layers with different convolutional kernels; construct the second branch based on depthwise separable convolutional blocks combined with residual connections; construct the third branch based on depthwise separable convolutional blocks; and construct a hybrid convolutional block based on the first branch, the second branch, and the third branch combined with multi-head attention.
[0011] Step 2: Construct a detection backbone based on hybrid convolutional blocks combined with multi-layer convolution, construct a multi-scale fusion structure based on multi-scale feature extraction and routing selection, construct a single-input dual-output prediction head based on hybrid convolutional blocks, and construct a detection model based on the detection backbone, multi-scale fusion structure and prediction head;
[0012] Step 3: Obtain an image of the manganese electrolysis plate and input the image of the manganese electrolysis plate into the detection model to obtain the detection result of residual manganese on the manganese electrolysis plate.
[0013] Furthermore, the construction of the first branch based on depth-separable convolutional blocks combined with routing multi-layer perception and convolutional layers with different convolutional kernels includes: constructing a convolution set based on convolutional layers with different convolutional kernels, and constructing a first branch for selecting convolutional layers with different convolutional kernels to process features according to the connection structure of depth-separable convolutional blocks, routing multi-layer perception, and convolutional sets.
[0014] Furthermore, the convolutional layers with different kernels include convolutional layers of vertical kernel type, convolutional layers of horizontal kernel type, and depth-separable convolutional blocks of standard kernel type.
[0015] Among them, the vertical convolution kernel type convolutional layer is a convolution kernel type convolutional layer with a kernel height greater than the width convolution kernel type, the horizontal convolution kernel type convolutional layer is a convolution kernel type convolutional layer with a kernel height less than the width convolution kernel type, and the standard convolution kernel type convolutional layer is a convolution kernel type convolutional layer with a kernel height equal to the width convolution kernel type.
[0016] Furthermore, in the hybrid convolutional block, a query vector is obtained based on the first branch, a key vector is obtained based on the second branch, and a value vector is obtained based on the third branch. The processed features are output based on the query vector, key vector, and value vector combined with multi-head attention.
[0017] Furthermore, the construction of the detection backbone based on hybrid convolutional blocks combined with multi-layer convolution includes: constructing feature extraction units based on convolutional layers with 3 kernels and stride of 2 and hybrid convolutional blocks, and constructing the detection backbone based on convolutional layers with 3 kernels and stride of 1 combined with stacked multiple feature extraction units.
[0018] Furthermore, the multi-scale fusion structure extracts features from the hybrid convolutional blocks of each feature extraction unit in the detection backbone, splices the extracted features, determines several expert features in the expert set based on the spliced features and routing selection, and splices and fuses the several expert features again to obtain fused features.
[0019] The predictive head receives the fused features and obtains the detection results of residual manganese based on the fused features;
[0020] The expert set is constructed based on the extracted features.
[0021] Furthermore, the prediction head includes an input branch, a bounding box branch, and a mask branch. The input end of the input branch is connected to the output end of the multi-scale fusion structure, and the output end of the input branch is connected to the input ends of the bounding box branch and the mask branch, respectively.
[0022] The input branch includes a convolutional layer with a kernel size of 3 and a stride of 1;
[0023] The bounding box branch sequentially includes a hybrid convolutional block and a convolutional layer with a kernel size of 3 and a stride of 1.
[0024] The mask branch includes a hybrid convolutional block and a convolutional layer with a kernel size of 3 and a stride of 1.
[0025] Through the above design, both bounding box prediction and segmentation mask prediction are output simultaneously, realizing joint modeling of localization and segmentation, ensuring the consistency of spatial location and morphological information of the detection results, and providing more comprehensive data support for subsequent quality assessment and process optimization.
[0026] Furthermore, the loss function of the detection model is constructed based on the bounding box loss and the segmentation mask loss;
[0027] The bounding box loss is constructed based on the difference between the predicted bounding box and the true bounding box in the bounding box branch;
[0028] The segmentation mask loss is constructed based on the pixel-level difference between the predicted mask and the real mask in the mask branch and the degree of overlap between the predicted region and the real region.
[0029] Furthermore, the bounding box loss is expressed by the following formula:
[0030] ;
[0031] in, Indicates the bounding box loss; This represents the bounding box regression loss;
[0032] The segmentation mask loss is expressed by the following formula:
[0033] ;
[0034] in, Indicates the segmentation mask loss; This represents the binary cross-entropy loss; Indicates Dice loss; This represents the weighting coefficient.
[0035] Based on the above design, a weighted fusion objective function of bounding box loss and segmentation mask loss is proposed, and Dice Loss is introduced to improve the detection capability of small targets. This achieves multi-task joint optimization, improves detection accuracy and stability, and is particularly suitable for complex defect detection in industrial scenarios.
[0036] Beneficial effects:
[0037] This invention provides a method for detecting residual manganese in manganese electrolytic plates based on multi-expert convolution. It constructs a hybrid convolution block based on the concept of expert hybrid structure and introduces an industrial defect detection scenario to achieve dynamic selection of convolution kernels. It can dynamically select the optimal expert according to the differences in input features, thereby improving the detection accuracy of long strips, small blocks and tiny residues.
[0038] In the multi-scale fusion structure, multi-scale features are fused, and then routing is performed based on the multi-scale features to select the corresponding optimal expert. The expert's features are then fused to improve the model's adaptability and robustness, avoid information loss, and enhance the relevance of the detection results.
[0039] By using convolution refinement and depth-separable convolutional blocks, computational complexity is reduced, ensuring detection accuracy while improving computational efficiency, enabling the model to be deployed in real time on industrial production lines. Attached Figure Description
[0040] Figure 1 This is a schematic diagram of the network structure of the hybrid convolutional block according to an embodiment of the present invention;
[0041] Figure 2 This is a schematic diagram of the network structure of the detection model in an embodiment of the present invention;
[0042] Figure 3 This is a schematic diagram of the network structure of the prediction head in an embodiment of the present invention. Detailed Implementation
[0043] 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.
[0044] 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.
[0045] This application provides a method for detecting residual manganese in a manganese electrolysis plate based on multi-expert convolution, comprising the following steps:
[0046] Step 1: Construct the first branch based on depthwise separable convolutional blocks combined with routing multilayer perception and convolutional layers with different kernels; construct the second branch based on depthwise separable convolutional blocks combined with residual connections; construct the third branch based on depthwise separable convolutional blocks; and construct a hybrid convolutional block based on the first, second, and third branches combined with multi-head attention.
[0047] Please see Figure 1 A convolution set is constructed based on convolutional layers of vertical convolutional kernel type, convolutional layers of horizontal convolutional kernel type, and depth-separable convolutional blocks of standard convolutional kernel type. The first branch for selecting convolutional layers with different convolutional kernels to process features is constructed according to the connection structure of depth-separable convolutional blocks, routing multilayer perception, and convolutional sets.
[0048] In this embodiment, the routing multilayer perception adopts a multilayer perceptron structure, performs nonlinear mapping on the intermediate features after processing by depth-separable convolutional blocks, obtains expert weight vectors, and then obtains expert selection probability distribution based on expert weight vectors combined with softmax normalization. Based on the expert selection probability distribution, it determines which type of convolutional kernel to use for convolutional layer processing of features.
[0049] For vertical convolutional kernel types, 5*1 convolutional layers are used; for horizontal convolutional kernel types, 1*5 convolutional layers are used; and for standard convolutional kernel types, depth-separable convolutional blocks with 3*3 kernels are used. This design is specifically for the industrial scenario of detecting manganese electrolysis plate residues. In actual production environments, manganese electrolysis plate residues may appear as long, thin strips or irregular small pieces. This application achieves adaptive modeling of manganese residues of different shapes through differentiated expert convolutional kernels (5*1, 1*5, 3*3), thus accommodating the detection needs of both long, thin, and small targets within the same framework.
[0050] The second branch is constructed by combining two depthwise separable convolutional blocks with residual connections. Based on the two-layer depthwise separable convolutional blocks, deeper spatial features can be extracted step by step, so that the final expression can capture more complex structural information. The residual connections effectively alleviate the gradient vanishing problem and ensure the trainability and stability of the deep network.
[0051] The third branch is constructed by a depthwise separable convolutional block and serves as the output branch of the value vector. The value vector is the final weighted output part in the attention mechanism and directly carries the main information of the input features. It can be processed by a single-layer depthwise separable convolutional block, which reduces the computational complexity while ensuring the overall running efficiency.
[0052] After constructing the first, second, and third branches, a hybrid convolutional block is built by combining multi-head attention. The working process of the hybrid convolutional block is as follows: the query vector is obtained based on the first branch, the key vector is obtained based on the second branch, and the value vector is obtained based on the third branch. The processed features are then output based on the query vector, key vector, and value vector combined with multi-head attention.
[0053] The use of depthwise separable convolutional blocks in hybrid convolutional blocks reduces redundant computation and maintains feature integrity. The first branch also incorporates multi-layer perceptual routing, enabling adaptive and targeted feature extraction. It can dynamically select the optimal convolutional kernel based on the intermediate features processed by the depthwise separable convolutional blocks, avoiding the limitations of traditional fixed convolutional kernels. This takes into account the differences in different target shapes and improves the comprehensiveness of detection.
[0054] Step 2: Construct a detection backbone based on hybrid convolutional blocks combined with multi-layer convolution, construct a multi-scale fusion structure based on multi-scale feature extraction and routing selection, construct a single-input dual-output prediction head based on hybrid convolutional blocks, and construct a detection model based on the detection backbone, multi-scale fusion structure and prediction head;
[0055] Please see Figure 2 For the detection backbone, feature extraction units are constructed based on convolutional layers with 3 kernels and stride of 2 and mixed convolutional blocks. Then, the detection backbone is constructed based on convolutional layers with 3 kernels and stride of 1 combined with stacked feature extraction units.
[0056] In this embodiment, three feature extraction units are stacked in the detection backbone.
[0057] The multi-scale fusion structure extracts features from the hybrid convolutional blocks of each feature extraction unit in the detection backbone, and splices the extracted features. Based on the spliced features and route selection, several expert features are determined from the expert set, and the several expert features are spliced and fused again to obtain the fused features.
[0058] In this embodiment, an expert weight vector is generated based on the splicing features combined with the routing of the multilayer perceptron. After the expert weight vector is normalized by softmax, a probability distribution is obtained. Based on the probability distribution, two expert features are determined in the expert set, and the two expert features are spliced and fused to obtain a fused feature.
[0059] The expert set here is constructed by using multiple features extracted from the multi-scale fusion structure as multiple experts. Each expert expresses features at different levels. In this embodiment, three features are extracted, which can be the fine feature expression of smaller targets, the feature characterization of medium targets, and the feature for overall structure recognition of larger targets. After routing selection, two appropriate expert features (i.e., the two extracted features) are selected and spliced to obtain the fusion feature.
[0060] The routing of the multilayer perceptron is also used here to achieve dynamic selection of differences, avoiding limitations. The two expert features selected also improve the comprehensiveness and accuracy of detection. Furthermore, by combining layer-by-layer downsampling with hybrid convolutional blocks, the convolutional kernel is dynamically selected based on features of different spatial degrees at different levels. The synergistic effect here can more effectively capture the feature information of long strips and small pieces of residue.
[0061] Finally, the prediction head receives the fused features and obtains the detection results of residual manganese based on the fused features;
[0062] Please see Figure 3The prediction head includes an input branch, a bounding box branch, and a mask branch. The input end of the input branch is connected to the output end of the multi-scale fusion structure, and the output end of the input branch is connected to the input ends of the bounding box branch and the mask branch, respectively.
[0063] The input branch includes a convolutional layer with a kernel size of 3 and a stride of 1;
[0064] The bounding box branch consists of a hybrid convolutional block and a convolutional layer with a kernel size of 3 and a stride of 1.
[0065] Each mask branch consists of a hybrid convolutional block and a convolutional layer with a kernel size of 3 and a stride of 1.
[0066] The loss function for the detection model is constructed based on the bounding box loss and the segmentation mask loss, and is expressed by the following formula:
[0067] ;
[0068] in, Represents the loss function; Indicates the bounding box loss; Indicates the segmentation mask loss; These represent the weight coefficients of the bounding box loss and the segmentation mask loss, respectively, which are used to control the importance of the localization and segmentation tasks.
[0069] The bounding box loss is constructed based on the difference between the predicted bounding box and the true bounding box in the bounding box branch, and is expressed by the following formula:
[0070] ;
[0071] in, This represents the bounding box regression loss;
[0072] The segmentation mask loss is constructed based on the pixel-level difference between the predicted mask and the real mask in the mask branch and the degree of overlap between the predicted region and the real region, and is expressed by the following formula:
[0073] ;
[0074] in, This represents the binary cross-entropy loss, used to measure the pixel-level difference between the predicted mask and the true mask; This represents the Dice loss, used to measure the degree of overlap between the predicted region and the actual region; The weighting coefficients represent the Dice loss.
[0075] The loss function, constructed based on bounding box loss and segmentation mask loss, ensures the balance between positioning and segmentation task. Furthermore, the introduction of Dice loss into the segmentation mask loss improves the robustness of small target detection, making it more suitable for the complex shapes and large size differences of residues in industrial scenarios.
[0076] In this embodiment, 8000 images of the electrolytic plate surface under various angles, resolutions, batches, process conditions, and defects were collected as samples for the training process of the detection model. Regarding the learning rate, a phased adjustment strategy was adopted to ensure stable convergence in the early stages of training. As training progressed, cosine annealing or a learning rate decay mechanism was gradually applied to allow the detection model to better approximate the optimal solution in later stages, avoiding overfitting or oscillations. The optimizer for the detection model was either Adam or SGD to achieve efficient parameter updates. Weight decay parameters were used to suppress overfitting and improve the model's generalization ability. Momentum parameters were used to accelerate the gradient descent process and avoid getting trapped in local optima. Furthermore, an early stopping mechanism and validation set monitoring were employed during training to ensure that the model stopped training when it reached optimal performance, avoiding resource waste.
[0077] Step 3: Obtain an image of the manganese electrolysis plate and input the image of the manganese electrolysis plate into the detection model to obtain the detection results of residual manganese on the manganese electrolysis plate.
[0078] Finally, images of the manganese electrolysis plate are captured by an industrial camera, and these images are input into a detection model to detect residual manganese. A detection frame for residual manganese on the manganese electrolysis plate is obtained, and the removal of residual manganese from the manganese electrolysis plate can be completed based on the detection frame.
[0079] To verify the effectiveness of the proposed multi-expert convolution-based residual manganese detection method for manganese electrolytic plates and the performance of the detection model, U-Net, YOLOv5-Seg, and YOLOv8-Seg were selected as benchmarks. All methods were trained and tested on a uniformly constructed manganese electrolytic plate image dataset, and their performance was evaluated using the same preprocessing procedures and evaluation metrics. Evaluation metrics included mean intersection-over-union ratio (mIoU), mean Dice coefficient (mDice), boundary localization accuracy (Boundary F1-score), and inference speed (FPS). Experimental comparison results are shown in Table 1.
[0080] Table 1: Experimental comparison results.
[0081]
[0082] As shown in Table 1, the method of this invention outperforms the comparative methods in all three accuracy metrics: mIoU, mDice, and Boundary F1-score. Specifically, compared to YOLOv8-Seg, this method improves mIoU by approximately 2.9%, mDice by approximately 2.8%, and boundary localization accuracy by approximately 3.4%, demonstrating a significant advantage, particularly in the fine identification of minute residual manganese. This performance improvement is mainly attributed to the combination of an expert hybrid convolutional structure and a router dynamic selection mechanism, enabling the model to adaptively select the optimal expert in complex contexts, thereby achieving more accurate boundary characterization.
[0083] In terms of inference speed, this method significantly reduces computational complexity through lightweight design and optimization of depthwise separable convolutional blocks. Under the same hardware environment, the inference speed of this method reaches 78 FPS, higher than YOLOv8-Seg's 72 FPS, which can meet the needs of real-time detection in industrial production lines. Further analysis shows that this method has particularly outstanding advantages in small target detection. Traditional methods often suffer from missed detections or incomplete boundaries when dealing with tiny residual manganese, while this invention effectively improves the detection and segmentation performance of small targets through dual-task joint output and the introduction of Dice Loss.
[0084] Furthermore, in robustness tests under different working conditions, this method maintains high detection accuracy even under uneven lighting, scratches, or noise interference, demonstrating its stability and adaptability in real industrial environments. In summary, the method of this invention not only surpasses existing technologies in accuracy and speed but also exhibits unique advantages in small target detection and robustness in complex scenarios, fully demonstrating its innovation and practical value in industrial applications.
[0085] Regarding the residual manganese detection method and detection model for manganese electrolytic plates based on multi-expert convolution proposed in this application, it should be noted that it is not only applicable to the detection of residues in manganese electrolytic plates, but can also be extended to surface defect detection and intelligent quality control in other industrial scenarios, such as surface crack detection of metal materials, defect identification of electronic components, and anomaly detection in high-precision manufacturing processes.
[0086] 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. A method for detecting residual manganese in a manganese electrolytic plate based on multi-expert convolution, characterized in that, Includes the following steps: Step 1: Construct the first branch based on depthwise separable convolutional blocks, combined with routing multilayer perception and convolutional layers with different kernels; construct the second branch based on stacked depthwise separable convolutional blocks and combined with residual connections; construct the third branch based on depthwise separable convolutional blocks; and construct a hybrid convolutional block based on the first branch, the second branch, and the third branch, combined with multi-head attention. The routing multi-layer perception adopts a multi-layer perceptron structure, performs non-linear mapping on the intermediate features after processing by depth-separable convolutional blocks, obtains expert weight vectors, obtains expert selection probability distribution based on expert weight vectors combined with softmax normalization, and determines the convolutional layer for processing features based on expert selection probability distribution. In the hybrid convolutional block, the query vector is obtained based on the first branch, the key vector is obtained based on the second branch, and the value vector is obtained based on the third branch. The processed features are output based on the query vector, key vector, and value vector combined with multi-head attention. Step 2: Construct a detection backbone based on hybrid convolutional blocks combined with multi-layer convolution, construct a multi-scale fusion structure based on multi-scale feature extraction and routing selection, construct a single-input dual-output prediction head based on hybrid convolutional blocks, and construct a detection model based on the detection backbone, multi-scale fusion structure and prediction head; The multi-scale fusion structure extracts features from the hybrid convolutional blocks of each feature extraction unit in the detection backbone, splices the extracted features, determines several expert features in the expert set based on the spliced features and routing selection, and splices and fuses the several expert features again to obtain the fused features. The expert set is constructed based on the extracted features; The predictive head receives the fused features and obtains the detection results of residual manganese based on the fused features; Step 3: Obtain an image of the manganese electrolysis plate and input the image of the manganese electrolysis plate into the detection model to obtain the detection result of residual manganese on the manganese electrolysis plate.
2. The method for detecting residual manganese in a manganese electrolytic plate based on multi-expert convolution as described in claim 1, characterized in that, The construction of the first branch based on depth-separable convolutional blocks combined with routing multi-layer perception and convolutional layers with different convolutional kernels includes: constructing a convolution set based on convolutional layers with different convolutional kernels, and constructing a first branch for selecting convolutional layers with different convolutional kernels to process features according to the connection structure of depth-separable convolutional blocks, routing multi-layer perception, and convolutional sets.
3. The method for detecting residual manganese in a manganese electrolytic plate based on multi-expert convolution as described in claim 2, characterized in that, The convolutional layers with different kernels include convolutional layers with vertical kernel types, convolutional layers with horizontal kernel types, and depth-separable convolutional blocks with standard kernel types.
4. The method for detecting residual manganese in a manganese electrolytic plate based on multi-expert convolution as described in claim 1, characterized in that, The detection backbone based on hybrid convolutional blocks combined with multi-layer convolution includes: constructing feature extraction units based on convolutional layers with 3 kernels and stride of 2 and hybrid convolutional blocks; and constructing the detection backbone based on convolutional layers with 3 kernels and stride of 1 combined with stacked feature extraction units.
5. The method for detecting residual manganese in a manganese electrolytic plate based on multi-expert convolution as described in claim 1, characterized in that, The prediction head includes an input branch, a bounding box branch, and a mask branch. The input end of the input branch is connected to the output end of the multi-scale fusion structure, and the output end of the input branch is connected to the input ends of the bounding box branch and the mask branch, respectively. The input branch includes a convolutional layer with a kernel size of 3 and a stride of 1; The bounding box branch sequentially includes a hybrid convolutional block and a convolutional layer with a kernel size of 3 and a stride of 1. The mask branch includes a hybrid convolutional block and a convolutional layer with a kernel size of 3 and a stride of 1.
6. The method for detecting residual manganese in a manganese electrolytic plate based on multi-expert convolution according to any one of claims 1-5, characterized in that, The loss function of the detection model is constructed based on bounding box loss and segmentation mask loss; The bounding box loss is constructed based on the difference between the predicted bounding box and the true bounding box in the bounding box branch; The segmentation mask loss is constructed based on the pixel-level difference between the predicted mask and the real mask in the mask branch and the degree of overlap between the predicted region and the real region.
7. The method for detecting residual manganese in a manganese electrolytic plate based on multi-expert convolution as described in claim 6, characterized in that, The bounding box loss is expressed by the following formula: ； in, Indicates the bounding box loss; This represents the bounding box regression loss; The segmentation mask loss is expressed by the following formula: ； in, Indicates the segmentation mask loss; This represents the binary cross-entropy loss; Indicates Dice loss; This represents the weighting coefficient.