Graphite ore grade detection method based on improved YOLO11 model

Abstract

The invention belongs to the technical field of image processing, and particularly relates to a graphite ore grade detection method based on an improved YOLO11 model, C3k2-CAS and Detect-SEAM modules are introduced, the feature extraction capability is enhanced, the expression of different scales of ore textures is optimized, and the perception capability of the model for grade difference is improved, so that the detection precision is remarkably improved; the attention mechanism of the CAS module is adopted to replace the traditional multiplication operation, the calculation complexity and the model parameter quantity are greatly reduced while the detection performance is maintained, and the improved model is more efficient than a baseline model and is more suitable for edge device deployment; a plurality of data enhancement strategies are combined, so that the model can maintain high robustness and stable detection capability in a complex industrial environment; the trained optimization model can be efficiently deployed to graphite ore grade detection intelligent equipment, real-time and accurate industrial field detection is achieved, the mineral separation efficiency is greatly improved, and an efficient and reliable computer vision solution is provided for the intelligent mining industry.

Description

Technical Field

[0001] The present invention belongs to the field of image processing technology, and specifically relates to a graphite ore grade detection method based on an improved YOLO11 model. Background Art

[0002] As a crucial strategic resource for modern industrial development, graphite ore plays an irreplaceable, core role in new energy, energy storage batteries, aerospace, and other fields. Its mining efficiency and quality control are directly related to the stability and economic viability of the industrial chain. With the advancement of smart mine construction, mining companies urgently need to achieve precise resource development and green, low-carbon production through intelligent technologies. Ore grade identification, as a key step in optimizing mining processes, requires technological breakthroughs of paramount importance.

[0003] Traditional graphite ore grade detection relies heavily on manual sampling and laboratory analysis, which suffers from low efficiency, poor real-time performance, and susceptibility to subjective experience. This makes it difficult to rapidly determine ore quality in dynamic mining scenarios. Smart mines, by integrating the Internet of Things (IoT), multi-source data perception, and artificial intelligence (AI) technologies, construct a real-time online grade identification system. This system efficiently captures ore quality distribution information and guides intelligent decision-making in processes such as sorting, crushing, and transportation. Efficient grade identification technology helps mining companies obtain timely ore quality information, optimize production plans, reduce waste and losses during production, and improve mine production efficiency. It also reduces the risk of resource mismatch and environmental pollution caused by grade misjudgment, and is a core support for promoting the high-value utilization of graphite ore resources. Therefore, proposing an improved detection model for rapid graphite ore grade detection is a decision that meets the current needs of graphite resource mining. Summary of the Invention

[0004] In order to optimize the deployment on edge computing devices, improve the overall accuracy and efficiency of detection, and thus provide technical support for intelligent detection of graphite mines, the present invention proposes a graphite ore grade detection method based on an improved YOLO11 model.

[0005] According to a first aspect of the present invention, the present invention claims protection for a graphite ore grade detection method based on an improved YOLO11 model, comprising:

[0006] Collecting ore samples at a graphite mining site to be tested, pre-classifying the ore samples and performing image acquisition to obtain a graphite ore original image library;

[0007] performing data enhancement and expansion on the graphite ore original images in the graphite ore original image library, and marking and dividing the graphite ore original image set;

[0008] Construct an initial YOLO11 target detection network model, use a sample data set for pre-training, and obtain trainable parameters and weights of the initial YOLO11 target detection network model;

[0009] Constructing an improved YOLO11 target detection network model, and initializing parameters of the improved YOLO11 target detection network model based on the trainable parameters and weights of the initial YOLO11 target detection network model;

[0010] The improved YOLO11 target detection network model after training parameters are initialized based on the graphite mine original image set is validated and evaluated to obtain a feasible evaluation result of the improved YOLO11 target detection network model;

[0011] The improved YOLO11 target detection network model whose feasible evaluation results meet the requirements is deployed to intelligent equipment for graphite ore grade detection to process ore images at the industrial site in real time.

[0012] Furthermore, the method of collecting ore samples at the graphite mining site to be detected, pre-classifying the ore samples and performing image acquisition to obtain a graphite ore original image library further includes:

[0013] Collect multiple ore samples at the graphite mining site to be tested;

[0014] Conducting preliminary pre-classification of the plurality of ore samples using mineralogical identification methods and in accordance with industrial grade classification standards;

[0015] Using an imaging device to collect the plurality of pre-classified ore samples to obtain a graphite ore original image library containing a plurality of graphite ore original images;

[0016] The graphite ore original image includes the morphology, detailed texture and edge features of the multiple ore samples.

[0017] Furthermore, the data enhancement and expansion of the graphite ore original images in the graphite ore original image library, and the annotation and division of the graphite ore original image set, further include:

[0018] Performing data enhancement processing on the graphite mine original image in the graphite mine original image library by using dynamic blur, dust noise, contrast adjustment and Mosaic;

[0019] Expanding the graphite ore original image library based on the graphite ore original image after data enhancement processing;

[0020] The graphite ore original images in the expanded graphite ore original image library are manually annotated using a data annotation tool, marked according to the target detection data format, and divided into a training set, a validation set, and a test set according to a preset ratio.

[0021] Furthermore, the constructing of the initial YOLO11 target detection network model, using the sample data set for pre-training to obtain the trainable parameters and weights of the initial YOLO11 target detection network model, also includes:

[0022] The initial YOLO11 target detection network model includes a first Backbone module, a first Neck module and a first Head module;

[0023] The first Backbone module performs multi-layer convolution operations to extract and encode the basic features of the sample data set, including layers 0 to 10, including multiple convolution modules Conv and C3k2 modules, SPPF modules and C2PSA modules, after which the feature map enters the first Neck module;

[0024] The first Neck module enhances feature expression capabilities through multi-scale feature fusion, integrates feature information of different scales, including layers 11 to 22, through multiple feature fusion and upsampling module Upsample operations, and combines the feature splicing module Concat with the C3k2 module fusion. The extracted and optimized feature map is input to the first Head module;

[0025] The first Head module performs target regression and classification tasks, mapping the extracted features into specific detection results, including target location, category, and confidence score, to complete target recognition and positioning. It includes three detection heads, each of which calculates the regression and classification losses of feature maps at different scales.

[0026] The regression and classification losses include at least complete IoU loss, distribution focus loss and classification loss;

[0027] The initial YOLO11 target detection network model is trained using the PascalVOC dataset to obtain a pre-training result of the initial YOLO11 target detection network model;

[0028] The trainable parameters and weights in the pre-training results of the initial YOLO11 target detection network model are exported as structured storage objects using a parameter serialization mechanism, and a trainable parameter and weight file in a .pt format containing a complete description of the network state is generated;

[0029] The trainable parameters include at least a convolution kernel weight matrix and a sliding mean or variance of a batch normalization layer.

[0030] Furthermore, the constructing of the improved YOLO11 target detection network model, initializing the parameters of the improved YOLO11 target detection network model based on the trainable parameters and weights of the initial YOLO11 target detection network model, further includes:

[0031] The improved YOLO11 target detection network model includes a second Backbone module, a second Neck module and a second Head module;

[0032] The second Backbone module structure is the same as the first Backbone module structure;

[0033] The second Neck module includes layers 11 to 22, and through multiple feature fusion and upsampling module Upsample operations, combined with the feature splicing module Concat and the C3k2-CAS module fusion, the extracted and optimized feature map is input to the second Head module;

[0034] The second Head module includes three separation-enhanced attention detection heads Detect-SEAM, which performs multi-scale feature extraction and completes channel-space interaction modeling through depth-separable convolution and residual connection;

[0035] After weighted fusion of multi-scale features, channel enhancement weights are generated through the global attention mechanism and multiplied with the original features to strengthen the response of key areas;

[0036] Load the trainable parameters and weight files of the initial YOLO11 target detection network model as the initial weights of the improved YOLO11 target detection network model.

[0037] Furthermore, the C3k2-CAS module further includes:

[0038] Intergration module, CATM module and MLP unit;

[0039] The Intergration module enhances local feature interaction and establishes residual connections through 1×1 convolution stacking;

[0040] The CATM combines convolution with additive attention mechanism to simultaneously model local structure and global association;

[0041] The MLP unit realizes feature space transformation through linear projection and adopts dual residual connection.

[0042] Furthermore, the improved YOLO11 target detection network model after initialization of training parameters based on the graphite ore original image is validated and evaluated to obtain a feasible evaluation result of the improved YOLO11 target detection network model, further comprising:

[0043] Configuring training parameters, and training the improved YOLO11 target detection network model based on the training parameters using a training set divided by the original graphite mine image;

[0044] Based on the validation set, the performance of the improved YOLO11 target detection network model is monitored in real time. After each training cycle, the model evaluation effect is obtained based on the validation set to determine whether to continue or terminate the training;

[0045] The performance of the trained improved YOLO11 target detection network model is evaluated on the test set. By comparing it with the initial YOLO11 target detection network model, the differences in various experimental indicators are obtained, and a feasible evaluation result of the improved YOLO11 target detection network model is obtained.

[0046] The technical effects achieved by the present invention are as follows:

[0047] (1) The backbone of the network retains the original YOLO11 network as the model's initial feature extractor. Its core function is to efficiently extract multi-level and multi-scale semantic features from the input image through a deep convolutional structure. Using an optimized hierarchical design, the early convolutional layers capture local mineral details (such as edge textures). As the network depth increases, high-level semantic features (such as image textures and target contours) are gradually abstracted. Feature maps at different levels are fused through cross-stage connections or feature pyramid mechanisms, ensuring real-time performance while enhancing the perception of multi-scale targets.

[0048] (2) This improved model adds a C3k2-CAS module designed specifically for mineral detection to the neck. Through the synergistic effect of convolution and additive self-attention mechanisms, efficient feature expression is achieved. In scenarios with limited computational costs, it takes into account both local structural information and global dependencies, and is suitable for visual recognition and lightweight reasoning tasks.

[0049] (3) This model introduces the SEAM module based on the original YOLO11 detection head and proposes a new detection head Detect-SEAM, which helps improve the model's ability to capture details in feature-missing environments (such as occlusion, noise, and motion blur) through the mechanism of deep separable convolution and inter-channel information fusion.

[0050] (4) The improved YOLO11 network structure proposed in the present invention has been significantly optimized in terms of multi-scale feature extraction and fusion, application of attention mechanism, lightweight design and computational cost, making its performance in graphite ore composition detection more accurate and efficient, and suitable for deployment and real-time application in industrial sites. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 A flowchart of a graphite ore grade detection method based on an improved YOLO11 model as claimed in the present invention;

[0052] Figure 2 This is a schematic diagram of an example of graphite ore image data for a graphite ore grade detection method based on an improved YOLO11 model claimed in the present invention;

[0053] Figure 3 A data enhancement comparison diagram of a graphite ore grade detection method based on an improved YOLO11 model claimed in the present invention;

[0054] Figure 4 A schematic diagram of the structure of an initial YOLO11 target detection network model for a graphite ore grade detection method based on an improved YOLO11 model claimed in the present invention;

[0055] Figure 5 A schematic diagram of the structure of an improved YOLO11 target detection network model for a graphite ore grade detection method based on an improved YOLO11 model claimed in the present invention;

[0056] Figure 6 A schematic diagram of the structure of C3k2-CAS and CAS modules of a graphite ore grade detection method based on an improved YOLO11 model claimed in the present invention;

[0057] Figure 7 A schematic diagram of the Detect-SEAM and CSMM modules of a graphite ore grade detection method based on an improved YOLO11 model claimed in the present invention;

[0058] Figure 8 A PR curve diagram of a graphite ore grade detection method based on an improved YOLO11 model claimed in the present invention;

[0059] Figure 9 This is a detection effect diagram of a graphite ore grade detection method based on an improved YOLO11 model claimed in the present invention. DETAILED DESCRIPTION

[0060] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0061] The terms "first", "second" and "third" in this application are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, a feature defined as "first", "second" and "third" may explicitly or implicitly include at least one of such features. In the description of this application, "multiple" means at least two, for example, two, three, etc., unless otherwise clearly and specifically defined. All directional indications in the embodiments of this application (such as up, down, left, right, front, back...) are only used to explain the relative positional relationship, movement, etc. between the components under a certain specific posture (as shown in the accompanying drawings). If the specific posture changes, the directional indication will also change accordingly. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not limited to the listed steps or units, but may optionally also include steps or units that are not listed, or may optionally also include other steps or units inherent to these processes, methods, products or devices.

[0062] References to "embodiments" herein mean that a particular feature, structure, or characteristic described in connection with the embodiments may be included in at least one embodiment of the present application. The appearance of the phrase in various places in the specification does not necessarily refer to the same embodiment, nor does it constitute an independent or alternative embodiment that is mutually exclusive of other embodiments. It is understood, both explicitly and implicitly, by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0063] According to the first embodiment of the present invention, referring to Figure 1 The present invention claims protection for a graphite ore grade detection method based on an improved YOLO11 model, comprising:

[0064] Collecting ore samples at a graphite mining site to be tested, pre-classifying the ore samples and performing image acquisition to obtain a graphite ore original image library;

[0065] performing data enhancement and expansion on the graphite ore original images in the graphite ore original image library, and marking and dividing the graphite ore original image set;

[0066] Construct an initial YOLO11 target detection network model, use a sample data set for pre-training, and obtain trainable parameters and weights of the initial YOLO11 target detection network model;

[0067] Constructing an improved YOLO11 target detection network model, and initializing parameters of the improved YOLO11 target detection network model based on the trainable parameters and weights of the initial YOLO11 target detection network model;

[0068] The improved YOLO11 target detection network model after training parameters are initialized based on the graphite mine original image set is validated and evaluated to obtain a feasible evaluation result of the improved YOLO11 target detection network model;

[0069] The improved YOLO11 target detection network model whose feasible evaluation results meet the requirements is deployed to intelligent equipment for graphite ore grade detection to process ore images at the industrial site in real time.

[0070] Furthermore, the method of collecting ore samples at the graphite mining site to be detected, pre-classifying the ore samples and performing image acquisition to obtain a graphite ore original image library further includes:

[0071] Collect multiple ore samples at the graphite mining site to be tested;

[0072] Conducting preliminary pre-classification of the plurality of ore samples using mineralogical identification methods and in accordance with industrial grade classification standards;

[0073] Using an imaging device to collect the plurality of pre-classified ore samples to obtain a graphite ore original image library containing a plurality of graphite ore original images;

[0074] The graphite ore original image includes the morphology, detailed texture and edge features of the multiple ore samples.

[0075] In this embodiment, taking the detection of graphite ore grade as an example, this embodiment mainly targets primary graphite ore and oxidized graphite ore with three grades of carbon content: low (0-10%), medium (10-20%), and high (>20%). The method and steps adopted are as follows:

[0076] A Canon EOS 5D Mark II SLR camera was used to capture graphite ore images at the industrial site.

[0077] To ensure image quality and detail retention, the imaging resolution was set to 2784×1856, covering different ore grade morphologies to enhance data diversity.

[0078] Some representative images of graphite ores are shown below: Figure 2As shown, it can be clearly observed that its visual features show significant correlation:

[0079] High-grade ore has few impurities and is a typical dark graphite black color, and the color deepens as the grade increases; low-grade ore contains light-colored impurity minerals such as quartz, so the color is lighter or mottled.

[0080] Although oxidized associated minerals (such as hematite and pyrite) may cause color interference, the overall pattern still follows a positive correlation between carbon content and black concentration. This process generates a large number of high-resolution mineral image samples, providing high-quality raw data for subsequent data processing, annotation, and model training.

[0081] Furthermore, the data enhancement and expansion of the graphite ore original images in the graphite ore original image library, and the annotation and division of the graphite ore original image set, further include:

[0082] Performing data enhancement processing on the graphite mine original image in the graphite mine original image library by using dynamic blur, dust noise, contrast adjustment and Mosaic;

[0083] Expanding the graphite ore original image library based on the graphite ore original image after data enhancement processing;

[0084] The graphite ore original images in the expanded graphite ore original image library are manually annotated using a data annotation tool, marked according to the target detection data format, and divided into a training set, a validation set, and a test set according to a preset ratio.

[0085] In this embodiment, the following data enhancement and expansion preprocessing is performed:

[0086] (i) Dynamic blur: This simulates the blurring effect caused by vibrations of conveyor belts and mining equipment, or other dynamic factors (such as motion changes in signal processing) in a mining environment. This type of blurring effect is implemented using the OpenCV filter2D function.

[0087] (ii) Dust noise: Irregular elliptical spots are added to random locations in the image to simulate lens smudges caused by dust particles in industrial sites. This type of noise is achieved by using a random function to generate the spot mask area.

[0088] (iii) Contrast adjustment: This reduces overall brightness and enhances shadow areas to simulate the uneven lighting found in mining environments. This is achieved using the contrast and brightness enhancement factors in the convertTo function.

[0089] (iv) Mosaic data enhancement: Randomly cut the image into 4 pieces and splice them into the same training sample, so that multiple targets can be presented in the same training image, which improves the model's adaptability to different scales and background changes, effectively enhances the feature learning effect, and improves the model's generalization ability. The dataset after the above data enhancement and expansion processing will be used as the input for model training. The image before and after processing is compared. Figure 3 shown.

[0090] The initial YOLO11 object detection network model was built. The deep learning framework included: an operating system of Windows 11 Professional Edition, a programming environment based on the Python 3.8.19 language framework, a deep learning module using the PyTorch 1.12 architecture, GPU acceleration through CUDA 11.3, and the number of output categories set to 3, predicting three grades respectively.

[0091] Use the Labelme image annotation tool to manually annotate the dataset in step 2 in the YOLO target detection format (.txt) and randomly divide the dataset into training, validation, and test sets in an 8:1:1 ratio.

[0092] Furthermore, the constructing of the initial YOLO11 target detection network model, using the sample data set for pre-training to obtain the trainable parameters and weights of the initial YOLO11 target detection network model, also includes:

[0093] The initial YOLO11 target detection network model includes a first Backbone module, a first Neck module and a first Head module;

[0094] The first Backbone module performs multi-layer convolution operations to extract and encode the basic features of the sample data set, including layers 0 to 10, including multiple convolution modules Conv and C3k2 modules, SPPF modules and C2PSA modules, after which the feature map enters the first Neck module;

[0095] The first Neck module enhances feature expression capabilities through multi-scale feature fusion, integrates feature information of different scales, including layers 11 to 22, through multiple feature fusion and upsampling module Upsample operations, and combines the feature splicing module Concat with the C3k2 module fusion. The extracted and optimized feature map is input to the first Head module;

[0096] The first Head module performs target regression and classification tasks, mapping the extracted features into specific detection results, including target location, category, and confidence score, to complete target recognition and positioning. It includes three detection heads, each of which calculates the regression and classification losses of feature maps at different scales.

[0097] The regression and classification losses include at least complete IoU loss, distribution focus loss and classification loss;

[0098] The initial YOLO11 target detection network model is trained using the PascalVOC dataset to obtain a pre-training result of the initial YOLO11 target detection network model;

[0099] The trainable parameters and weights in the pre-training results of the initial YOLO11 target detection network model are exported as structured storage objects using a parameter serialization mechanism, and a trainable parameter and weight file in a .pt format containing a complete description of the network state is generated;

[0100] The trainable parameters include at least a convolution kernel weight matrix and a sliding mean or variance of a batch normalization layer.

[0101] In this embodiment, the structure of the initial YOLO11 target detection network model is as follows: Figure 4 As shown in the figure, this model boasts efficient real-time performance, precise feature extraction, and automated processing, effectively addressing the shortcomings of traditional graphite ore detection methods. Compared to two-stage object detection algorithms (such as FasterR-CNN), YOLO11 offers the advantages of a single-stage detector in terms of real-time performance and accuracy, making it more suitable for deployment on edge devices, enabling fast and accurate ore grade detection. The model consists of three core modules: Backbone, Neck, and Head.

[0102] Among them, Backbone is responsible for performing multi-layer convolution operations to extract and encode the basic features of the image and provide high-quality feature expression for subsequent processing.

[0103] Neck enhances feature expression capabilities through multi-scale feature fusion, ensuring that information at different scales is effectively integrated, thereby improving the model's ability to detect targets of different sizes.

[0104] The Head is responsible for the final target regression and classification tasks, mapping the extracted features into specific detection results, including target location, category, and confidence score, thereby achieving accurate target recognition and positioning.

[0105] Specifically, the image is first fed into the Backbone feature extraction network for deep processing. This architecture, comprised of a multi-layer convolutional architecture and C3k2 feature extraction modules, aims to gradually extract key information from the image and improve the model's ability to perceive objects of varying scales. The Backbone architecture spans 0-10 layers and includes multiple convolutional layers and C3k2 structures to effectively enhance feature extraction capabilities.

[0106] Subsequently, the SPPF (Spatial Pyramid Pooling-Fast) module fuses features from different receptive fields to enhance the model's multi-scale representation capabilities. The C2PSA module further optimizes feature allocation and weight adjustment, thereby improving detection accuracy. On this basis, the feature map enters the Neck component, where multiple feature fusion and upsampling operations are performed to enhance the detection capabilities of objects of different scales.

[0107] In this architecture, upsampling is first performed at layer 11, where it is further fused with C3k2 through feature concatenation. Layer 14 is then upsampled again and concatenated with the deep features transmitted by Backbone to form a more discriminative feature representation. This continuous processing by the C3k2 module ensures the effective integration of high- and low-level information, further improving the detection network's adaptability to small objects and complex backgrounds.

[0108] Finally, the feature map extracted and optimized by Neck is input to the Head end to complete the target detection task.

[0109] The head structure comprises multiple detection branches, performing target regression and classification at different scales to achieve high-precision target positioning and grade identification. Specifically, after multiple feature concatenation (Concat) and C3k2 modules, the detection head (Detect) generates final detection results at different scales to ensure efficient identification of ores of different sizes.

[0110] Specifically, the Head consists of three decoupled detection heads, each of which is responsible for calculating the regression and classification losses of feature maps of different scales to adapt to the detection requirements of targets of different sizes.

[0111] The regression loss mainly includes CIoULoss (Complete IoU Loss) and DFLLoss (Distribution Focal Loss), while the classification loss is calculated using Cls (Classification Loss). CIoULoss takes into account the consistency of IoU, center point distance, and aspect ratio, further improving the performance of bounding box regression. The specific calculation formula of CIoULoss is as follows:

[0112]

[0113] IoU (Intersection over Union) represents the intersection over union ratio between the predicted box and the real box. ρ(b,b gt) represents the Euclidean distance between the two center points, and d is the diagonal length of the minimum bounding box between the two. The parameter α represents the weight term, and v represents the correction factor.

[0114] DFLLoss (DistributionFocalLoss) is mainly used to discretize the coordinates of the modeling bounding box and optimize the prediction by distributing the weights. The specific calculation formula is as follows:

[0115]

[0116] Where K represents the number of discrete locations, Represents the predicted probability of the position after the Softmax function, and w k Represents the distance between the target value and the discrete location.

[0117] ClsLoss (ClassificationLoss) is used to measure the difference between the predicted category probability and the true category. Its calculation formula is as follows:

[0118]

[0119] Among them, N represents the total number of samples, C represents the total number of categories, and y i,c is the true category label, and is the class probability predicted by the model.

[0120] The structural design of this network structure fully utilizes the advantages of YOLO11 in real-time, computational efficiency and detection accuracy, making it highly applicable in graphite ore grade detection tasks. It is especially suitable for efficient and accurate automatic ore identification and evaluation on edge devices with limited computing resources.

[0121] The initial model was pre-trained on the large dataset PascalVOC to test the effectiveness of the initial YOLO11 target detection network model. The PascalVOC dataset covers 20 different categories of targets and contains nearly 20,000 information-rich images with precise annotations. It is widely used for training and performance evaluation of target detection algorithms. Selecting this dataset for pre-training not only provides high-quality initialization parameters for the model, which helps accelerate convergence, but also significantly improves its adaptability in complex environments. In particular, in the task of graphite ore grade detection, the surface texture features of the ore are complex and the category distinction is low. Pre-training can effectively enhance the model's ability to perceive mineral texture details, improve the accuracy and stability of feature extraction, thereby improving the robustness and accuracy of detection, and providing stronger performance guarantees for subsequent graphite ore grade identification.

[0122] After the initial YOLO11 object detection network model completes pre-training and converges, the system exports trainable parameters such as the convolution kernel weight matrix and the sliding mean or variance of the batch normalization layer as structured storage objects through the parameter serialization mechanism, and finally generates a .pt format file containing a complete description of the network status, laying the foundation for subsequent network improvement and formal training.

[0123] Furthermore, the constructing of the improved YOLO11 target detection network model, initializing the parameters of the improved YOLO11 target detection network model based on the trainable parameters and weights of the initial YOLO11 target detection network model, further includes:

[0124] The improved YOLO11 target detection network model includes a second Backbone module, a second Neck module and a second Head module;

[0125] The second Backbone module structure is the same as the first Backbone module structure;

[0126] The second Neck module includes layers 11 to 22, and through multiple feature fusion and upsampling module Upsample operations, combined with the feature splicing module Concat and the C3k2-CAS module fusion, the extracted and optimized feature map is input to the second Head module;

[0127] The second Head module includes three separation-enhanced attention detection heads Detect-SEAM, which performs multi-scale feature extraction and completes channel-space interaction modeling through depth-separable convolution and residual connection;

[0128] After weighted fusion of multi-scale features, channel enhancement weights are generated through the global attention mechanism and multiplied with the original features to strengthen the response of key areas;

[0129] Load the trainable parameters and weight files of the initial YOLO11 target detection network model as the initial weights of the improved YOLO11 target detection network model.

[0130] In this embodiment, the structure of the improved YOLO11 target detection network model is shown in Figure 5 The purpose is to better capture the texture details of graphite ore, enhance the model's ability to extract features of different scales in the image, effectively distinguish the differences between graphite ores of different grades, and ultimately improve detection accuracy and ensure the lightweight and real-time performance of the model. This invention has made the following improvements to the graphite ore detection task:

[0131] First, the feature extraction module of the Neck part in the initial YOLO11 target detection network model, namely C3k2 (corresponding to the 13th, 16th, 19th and 22nd layers), is replaced with the independently improved C3k2-CAS module, whose structure is as follows: Figure 6 shown.

[0132] Specifically, C3k2-CAS continues to use a branching structure to improve the diversity of feature extraction. After the input features are processed by the CBS module, they are divided into multiple subspaces through the Split operation and sent to multiple CASBlocks (Convolutional Additive Self-attention Blocks) in parallel for feature transformation.

[0133] Subsequently, the features of all sub-branches are aggregated through Concat and optimized again through CBS (Conv-BatchNorm-SiLU) to enhance feature consistency and expressiveness. Secondly, the CASBlock module of the C3k2-CAS core consists of an Intergration module, a CATM (Convolutional Additive Token Mixer) module, and an MLP unit: the Intergration module enhances local feature interactions and establishes residual connections through 1×1 convolution stacking; CATM integrates convolution with additive attention mechanisms to simultaneously model local structures and global associations with low computational overhead; the MLP unit achieves feature space transformation through linear projection, and dual residual connections are used as a whole to ensure gradient stability. This structure balances local and global feature representations under limited computing power and is suitable for lightweight visual tasks.

[0134] Finally, by integrating the Separated and Enhancement Attention Module (SEAM) into the detection head of YOLO11, the Detect-SEAM structure is further proposed. Figure 7 As shown in the figure, the input features of this module are subjected to multi-scale feature extraction by the multi-branch CSMM module, and each branch realizes channel-space interaction modeling through depth-wise separable convolution and residual connection.

[0135] After weighted fusion of multi-scale features, a global attention mechanism is used to generate channel enhancement weights, which are multiplied with the original features to enhance the response in key areas. To meet the needs of mineral texture recognition, Detect-SEAM effectively improves the ability to capture ore details in complex backgrounds through adaptive feature focusing.

[0136] Compared with the initial YOLO11 target detection network model, the improved modules C3k2-CAS and Detect-SEAM work together in feature extraction and efficient detection, aiming to enable the improved YOLO11 target detection network model to achieve more efficient feature expression, improve the accuracy and computational efficiency of detection tasks, and reduce computational costs, thereby providing a better computer vision solution for graphite ore grade detection.

[0137] Load the weight file of the initial YOLO11 object detection network model to provide optimized initial weights for the improved YOLO11 object detection network model. This operation can effectively accelerate the convergence process of the model. By using the trained weights as the initial conditions, you avoid training from scratch, thereby shortening the training time and improving the convergence speed. At the same time, the loaded initial weights also ensure that the model can more stably learn effective features in subsequent training, further improving the efficiency of the training process and the final performance of the model.

[0138] Furthermore, the C3k2-CAS module further includes:

[0139] Intergration module, CATM module and MLP unit;

[0140] The Intergration module enhances local feature interaction and establishes residual connections through 1×1 convolution stacking;

[0141] The CATM combines convolution with additive attention mechanism to simultaneously model local structure and global association;

[0142] The MLP unit realizes feature space transformation through linear projection and adopts dual residual connection.

[0143] Furthermore, the improved YOLO11 target detection network model after initialization of training parameters based on the graphite ore original image is validated and evaluated to obtain a feasible evaluation result of the improved YOLO11 target detection network model, further comprising:

[0144] Configuring training parameters, and training the improved YOLO11 target detection network model based on the training parameters using a training set divided by the original graphite mine image;

[0145] Based on the validation set, the performance of the improved YOLO11 target detection network model is monitored in real time. After each training cycle, the model evaluation effect is obtained based on the validation set to determine whether to continue or terminate the training;

[0146] The performance of the trained improved YOLO11 target detection network model is evaluated on the test set. By comparing it with the initial YOLO11 target detection network model, the differences in various experimental indicators are obtained, and a feasible evaluation result of the improved YOLO11 target detection network model is obtained.

[0147] In this example, the model training parameters were initialized as follows: batch size (batchsize) was 16, initial learning rate (lr) was 0.01, number of training epochs was set to 200, and SGD optimizer with momentum was used to accelerate convergence. For loss functions, CIoULoss and DFLLoss were used to calculate regression loss, and ClsLoss was used to calculate classification loss, to improve detection accuracy and feature expression capabilities.

[0148] The obtained pre-trained parameters were loaded into YOLO11_improved and trained using the enhanced dataset. During training, the weight parameters were continuously optimized to enhance the recognition ability of complex mineral textures. The accuracy and loss value changes were monitored on the validation set to ensure stable optimization and efficient convergence of the model.

[0149] The optimized YOLO11_improved model was applied to a pre-partitioned test set for a comprehensive performance evaluation. Compared to the baseline model, YOLO11_original, YOLO11_improved achieved a mAP@0.5 (mean average precision) of 97.6%, 7.2% higher than the baseline model. In terms of lightweighting, its parameter count was reduced to 1.9M and its model size was reduced to 4.0MB, 26% and 24% smaller than the baseline model, respectively. This validates the optimization effects of the improved modules C3k2-CAS and Detect-SEAM in feature extraction, bounding box regression, classification accuracy, and computational efficiency.

[0150] Draw the PR curve according to the model data. Figure 8 As shown, the detection effect is as follows Figure 9 As shown in the figure, it demonstrates the high-accuracy, real-time detection capability of graphite ore grade, which fully meets the deployment requirements of edge embedded devices.

[0151] In the several embodiments provided in this application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are merely schematic. For example, the division of units is only a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interfaces, devices or units, which can be electrical, mechanical or other forms.

[0152] In addition, the functional units in the various embodiments of the present application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The above-mentioned integrated units can be implemented in the form of hardware or in the form of software functional units. The above is only an implementation method of the present application and does not limit the patent scope of the present application. Any equivalent structure or equivalent process transformation made using the contents of the description and drawings of this application, or directly or indirectly used in other related technical fields, are also included in the patent protection scope of the present application.

[0153] The above detailed description of the specific embodiments of the invention is intended only as an example, and the present application is not limited to the specific embodiments described above. For those skilled in the art, any equivalent modifications or substitutions of the invention are also within the scope of the present application. Therefore, equivalent changes, modifications, and improvements made without departing from the spirit and scope of the present application should be included within the scope of the present application.

Claims

1. A graphite ore grade detection method based on an improved YOLO11 model, characterized in that: include: Collecting ore samples at a graphite mining site to be tested, pre-classifying the ore samples and performing image acquisition to obtain a graphite ore original image library; performing data enhancement and expansion on the graphite ore original images in the graphite ore original image library, and marking and dividing the graphite ore original image set; Construct an initial YOLO11 target detection network model, use a sample data set for pre-training, and obtain trainable parameters and weights of the initial YOLO11 target detection network model; Constructing an improved YOLO11 target detection network model, and initializing parameters of the improved YOLO11 target detection network model based on the trainable parameters and weights of the initial YOLO11 target detection network model; The improved YOLO11 target detection network model after training parameters are initialized based on the graphite mine original image set is validated and evaluated to obtain a feasible evaluation result of the improved YOLO11 target detection network model; The improved YOLO11 target detection network model whose feasible evaluation results meet the requirements is deployed to intelligent equipment for graphite ore grade detection to process ore images at the industrial site in real time.

2. The graphite ore grade detection method based on the improved YOLO11 model according to claim 1 is characterized in that: The method further comprises: collecting ore samples at a graphite mining site to be tested, pre-classifying the ore samples and performing image acquisition to obtain a graphite ore original image library; Collect multiple ore samples at the graphite mining site to be tested; Conducting preliminary pre-classification of the plurality of ore samples using mineralogical identification methods and in accordance with industrial grade classification standards; Using an imaging device to collect the plurality of pre-classified ore samples to obtain a graphite ore original image library containing a plurality of graphite ore original images; The graphite ore original image includes the morphology, detailed texture and edge features of the multiple ore samples.

3. The graphite ore grade detection method based on the improved YOLO11 model according to claim 1 is characterized in that: The data enhancement and expansion of the graphite ore original images in the graphite ore original image library, and the marking and dividing of the graphite ore original image set, further comprises: performing data enhancement processing on the graphite ore original images in the graphite ore original image library by using dynamic blur, dust noise, contrast adjustment, and Mosaic; The graphite ore original image library is expanded based on the graphite ore original image after data enhancement processing; the graphite ore original images in the expanded graphite ore original image library are manually annotated using a data annotation tool, marked according to the target detection data format, and the graphite ore original images are divided into a training set, a validation set and a test set according to a preset ratio.

4. The graphite ore grade detection method based on the improved YOLO11 model according to claim 1, characterized in that: The initial YOLO11 target detection network model is constructed, and a sample data set is used for pre-training to obtain trainable parameters and weights of the initial YOLO11 target detection network model, further comprising: The initial YOLO11 target detection network model includes a first Backbone module, a first Neck module and a first Head module; The first Backbone module performs multi-layer convolution operations to extract and encode the basic features of the sample data set, including layers 0 to 10, including multiple convolution modules Conv and C3k2 modules, SPPF modules and C2PSA modules, after which the feature map enters the first Neck module; The first Neck module enhances feature expression capabilities through multi-scale feature fusion, integrates feature information of different scales, including layers 11 to 22, through multiple feature fusion and upsampling module Upsample operations, and combines the feature splicing module Concat with the C3k2 module fusion. The extracted and optimized feature map is input to the first Head module; The first Head module performs target regression and classification tasks, mapping the extracted features into specific detection results, including target location, category, and confidence score, to complete target recognition and positioning. It includes three detection heads, each of which calculates the regression and classification losses of feature maps at different scales. The regression and classification losses include at least complete IoU loss, distribution focus loss and classification loss; The initial YOLO11 target detection network model is trained using the PascalVOC dataset to obtain a pre-training result of the initial YOLO11 target detection network model; The trainable parameters and weights in the pre-training results of the initial YOLO11 target detection network model are exported as structured storage objects using a parameter serialization mechanism, and a trainable parameter and weight file in a .pt format containing a complete description of the network state is generated; The trainable parameters include at least a convolution kernel weight matrix and a sliding mean or variance of a batch normalization layer.

5. The graphite ore grade detection method based on the improved YOLO11 model according to claim 4 is characterized in that: The step of constructing an improved YOLO11 target detection network model and initializing parameters of the improved YOLO11 target detection network model based on the trainable parameters and weights of the initial YOLO11 target detection network model further includes: The improved YOLO11 target detection network model includes a second Backbone module, a second Neck module and a second Head module; The second Backbone module has the same structure as the first Backbone module. The second Neck module includes layers 11 to 22. Through multiple feature fusion and upsampling module Upsample operations, combined with the feature concatenation module Concat and the C3k2-CAS module, the extracted and optimized feature map is input to the second Head module. The second Head module includes three separate enhanced attention detection heads Detect-SEAM, which perform multi-scale feature extraction and complete channel-space interaction modeling through deep separable convolution and residual connection. After weighted fusion of multi-scale features, channel enhancement weights are generated through the global attention mechanism and multiplied with the original features to strengthen the response of key areas; Load the trainable parameters and weight files of the initial YOLO11 target detection network model as the initial weights of the improved YOLO11 target detection network model.

6. The graphite ore grade detection method based on the improved YOLO11 model according to claim 5, characterized in that: The C3k2-CAS module further includes: Intergration module, CATM module and MLP unit; The Intergration module enhances local feature interaction and establishes residual connections through 1×1 convolution stacking; The CATM combines convolution with additive attention mechanism to simultaneously model local structure and global association; The MLP unit realizes feature space transformation through linear projection and adopts dual residual connection.

7. The method for detecting graphite ore grade based on the improved YOLO11 model according to claim 3, wherein: The improved YOLO11 target detection network model after initializing the training parameters based on the graphite mine original image and performing model verification and evaluation to obtain a feasible evaluation result of the improved YOLO11 target detection network model also includes: configuring training parameters, and training the improved YOLO11 target detection network model based on the training parameters using a training set divided by the graphite mine original image; Based on the validation set, the performance of the improved YOLO11 target detection network model is monitored in real time. After each training cycle, the model evaluation effect is obtained based on the validation set to determine whether to continue or terminate the training; The performance of the trained improved YOLO11 target detection network model is evaluated on the test set. By comparing it with the initial YOLO11 target detection network model, the differences in various experimental indicators are obtained, and a feasible evaluation result of the improved YOLO11 target detection network model is obtained.

Images