Graphite ore image segmentation method based on improved YOLO11-seg model

By improving the YOLO11-seg model and combining it with the GSConv and C3k2-Faster modules to optimize feature extraction, the high computational overhead of the graphite ore segmentation model was solved, and lightweight and high-precision ore segmentation was achieved, which is suitable for real-time applications on edge devices.

CN120782783APending Publication Date: 2025-10-14JIANGXI UNIV OF SCI & TECH

Patent Information

Application Number
CN202510777772.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-06-11
Publication Date
2025-10-14

AI Technical Summary

Technical Problem

The existing graphite ore segmentation model has high computational overhead, making it difficult to achieve real-time and accurate low-grade mineral segmentation on edge devices, affecting the accuracy of ore grade assessment and resource utilization efficiency.

Method used

The improved YOLO11-seg model was used to construct a lightweight graphite ore image segmentation model through data enhancement and expansion, parameter initialization and model verification. The GSConv and C3k2-Faster modules were used to optimize feature extraction, reduce computational complexity and improve segmentation accuracy.

Benefits of technology

It significantly reduces the model parameter scale and computational complexity, improves adaptability and segmentation accuracy in resource-constrained environments, and meets the real-time processing needs of industrial sites.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN120782783A_ABST
    Figure CN120782783A_ABST
Patent Text Reader

Abstract

The invention belongs to the technical field of image processing, and particularly relates to a graphite ore image segmentation method based on an improved YOLO11-seg model, which adopts a balanced design of precision and light weight, optimizes redundancy in a network structure, introduces GSConv and C3k2-Faster modules, effectively reduces calculation overhead and memory occupation on the premise of keeping strong feature representation capability, and improves the image segmentation efficiency. The real-time reasoning efficiency of the model is obviously improved; a Mosai c data enhancement method is fully utilized, the data diversity is improved while the number of samples is increased, and the robustness and generalization ability of the model in a complex environment are enhanced; the trained optimization model can be deployed on intelligent edge equipment of graphite ores, supports real-time and accurate grade estimation on site, remarkably improves the mineral separation efficiency, and provides an efficient and reliable computer vision solution for intelligent mining scenes.
Need to check novelty before this filing date? Find Prior Art

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of image processing, and in particular relates to a graphite ore image segmentation method based on an improved YOLO11-seg model. BACKGROUND

[0002] In the process of graphite ore mining, dressing and processing, accurately segmenting and identifying low-grade minerals has important practical significance and technical value. This process not only relates to the control of the overall quality of the ore, but also plays a key supporting role in the efficient use of resources and ecological environment protection.

[0003] Building a graphite ore segmentation model with low computational overhead, real-time inference capability and high segmentation accuracy is of great significance for realizing automatic evaluation of ore grade. The model should be deployable on edge devices and capable of quickly and accurately segmenting mineral areas of different grades. This method provides strong technical support and a feasible implementation path for building an intelligent, green and efficient mine system. SUMMARY

[0004] In order to optimize the beneficiation process from the source, effectively avoid low-grade ore entering the subsequent processing link, and thus reduce resource waste, energy consumption and carbon emissions, the present application proposes a graphite ore image segmentation method based on an improved YOLO11-seg model.

[0005] According to the first aspect of the present application, the present application claims a graphite ore image segmentation method based on an improved YOLO11-seg model, comprising:

[0006] Collecting ore samples in the graphite mining field to be detected, pre-classifying the ore samples and collecting images to obtain a graphite ore original image library;

[0007] Data augmentation and expansion are performed on the graphite ore original images of the graphite ore original image library, and the graphite ore original image set is labeled and divided;

[0008] An initial YOLO11-seg instance segmentation model is constructed, and sample data sets are used for pre-training to obtain trainable parameters and weights of the initial YOLO11-seg instance segmentation model;

[0009] An improved YOLO11-seg instance segmentation model is constructed, and the trainable parameters and weights of the initial YOLO11-seg instance segmentation model are used to initialize the parameters of the improved YOLO11-seg instance segmentation model;

[0010] training the improved YOLO11-seg instance segmentation model based on the parameter initialization and the original image set of the graphite ore, and performing model verification and evaluation to obtain a feasible evaluation result of the improved YOLO11-seg instance segmentation model;

[0011] The improved YOLO11-seg instance segmentation model meeting the requirements in the feasible evaluation result is deployed to graphite ore grade detection intelligent equipment to process ore images in an industrial field in real time.

[0012] Further, the ore samples are collected in the graphite mining field to be detected, the ore samples are pre-classified and image collection is performed to obtain an original image library of graphite ore, and the method further comprises:

[0013] A plurality of ore samples are collected in the graphite mining field to be detected.

[0014] The plurality of ore samples are pre-classified by using a mineralogical identification method and according to an industrial grade classification standard;

[0015] The plurality of pre-classified ore samples are collected by using an imaging device to obtain an original image library of graphite ore containing a plurality of original images of graphite ore;

[0016] The original images of graphite ore in the original image library of graphite ore include the morphology, detailed texture and edge features of the plurality of ore samples.

[0017] Further, the original images of graphite ore in the original image library of graphite ore are subjected to data enhancement and expansion, and the original images of graphite ore are labeled and divided, and the method further comprises:

[0018] The original images of graphite ore in the original image library of graphite ore are subjected to data enhancement processing by using Mosaic;

[0019] The original image library of graphite ore is expanded based on the original images of graphite ore subjected to data enhancement processing;

[0020] The original images of graphite ore in the expanded original image library of graphite ore are manually labeled by using a data labeling tool, are marked according to a target detection data format, and are divided into a training set, a verification set and a test set according to a preset proportion.

[0021] Further, the initial YOLO11-seg instance segmentation model is constructed, and a sample data set is used for pre-training to obtain trainable parameters and weights of the initial YOLO11-seg instance segmentation model, and the method further comprises:

[0022] The initial YOLO11-seg instance segmentation model comprises a first Backbone module, a first Neck module and a first Head module.

[0023] The first Backbone module extracts low-level and high-level features of the sample dataset, provides semantic information for downstream tasks, including layers 0 to 10, including multiple convolution modules Conv and C3k2 modules, SPPF modules and C2PSA modules, and then the feature mapping enters the first Neck module;

[0024] The first Neck module enhances the feature expression ability through multi-scale feature fusion and integrates different scale feature information, including layers 11 to 22, through multiple feature fusion and upsample module Upsample operations, combined with feature splicing module Concat and C3k2 module fusion, and the extracted and optimized features are mapped to the first Head module;

[0025] The first Head module performs target positioning and classification work, and generates corresponding masks combined with the segmentation branch, outputs target categories and corresponding confidence, including three segmentation modules Segment;

[0026] The ImageNet or COCO dataset is used to train the initial YOLO11-seg instance segmentation model, and the pre-training result of the initial YOLO11-seg instance segmentation model is obtained;

[0027] The trainable parameters and weights in the pre-training result of the initial YOLO11-seg instance segmentation model are serialized using the parameter serialization mechanism;

[0028] The trainable parameters at least include edge, texture and shape distribution.

[0029] Further, the improved YOLO11-seg instance segmentation model is constructed, and the trainable parameters and weights of the initial YOLO11-seg instance segmentation model are used for parameter initialization of the improved YOLO11-seg instance segmentation model, and further comprising:

[0030] The improved YOLO11-seg instance segmentation model keeps the overall channel width of the initial YOLO11-seg instance segmentation model unchanged, and proportionally reduces the global depth and maximum channel number, including a second Backbone module, a second Neck module and a second Head module;

[0031] The second Backbone module structure includes layers 0 to 10, including multiple grouped spatial convolution modules GSConv and C3k2-Faster modules, SPPF modules and C2PSA modules, and then the feature mapping enters the first Neck module;

[0032] The second Neck module includes the 11th layer to the 22nd layer, and the extracted and optimized feature maps are input into the second Head module through multiple feature fusion and up-sampling module Upsample operations, combined with feature splicing module Concat and C3k2-Faster module fusion;

[0033] The second Head module includes three segmentation modules Segment;

[0034] The trainable parameters and weight data of the initial YOLO11-seg instance segmentation model are loaded as the initial weight of the improved YOLO11-seg instance segmentation model.

[0035] Further, the group spatial convolution module GSConv further includes:

[0036] When performing group convolution, for the input tensor The convolution process of the standard convolution SC branch is represented as:

[0037]

[0038] Wherein, "*" represents convolution operation, and c2 represents the number of output channels;

[0039] The depth separable convolution DSC branch is represented as:

[0040]

[0041] The outputs of the two branches are spliced Concat, and the effective fusion of the standard convolution SC feature and the depth separable convolution DSC output is realized through the channel shuffling operation Shuffle, to generate an optimized feature map;

[0042] The overall process of GSConv is represented as:

[0043]

[0044] Further, the improved YOLO11-seg instance segmentation model based on the training parameter initialized according to the original image of the graphite ore is subjected to model verification and evaluation, to obtain a feasible evaluation result of the improved YOLO11-seg instance segmentation model, and the method further includes:

[0045] The training parameters are configured, and the improved YOLO11-seg instance segmentation model is trained based on the training set divided according to the original image of the graphite ore according to the training parameters;

[0046] Real-time monitoring performance of the improved YOLO11-seg instance segmentation model based on the validation set, obtaining model evaluation effect based on the validation set after each training cycle, and determining whether to continue or terminate training;

[0047] Performance evaluation of the trained improved YOLO11-seg instance segmentation model on the test set, obtaining the difference in each experimental index by comparing with the initial YOLO11-seg instance segmentation model, and obtaining the feasible evaluation result of the improved YOLO11-seg instance segmentation model.

[0048] The technical effects achieved by the application are as follows:

[0049] (1) By reducing the global depth factor by equal proportion, the parameter size and the calculation complexity of the model are effectively reduced, and the deployment adaptability in the resource-limited environment is significantly improved.

[0050] (2) Introducing an efficient GSConv to replace the traditional down-sampling convolution operation, the module integrates a deep separable convolution and a channel shuffle mechanism, which significantly reduces the calculation overhead and memory usage while strengthening the feature extraction capability.

[0051] (3) On the basis of the original C3k2 feature extraction module, a lightweight bottleneck structure Faster Block is integrated to form a new C3k2-Faster module. The module adopts a partial convolution (PConv) strategy for selective processing of input channels, significantly reducing the floating point operations per second and memory access requirements, while improving memory utilization efficiency, effectively enhancing the expression and modeling ability of the model in complex scenes.

[0052] (4) Through multiple optimization measures, the application considers segmentation accuracy, calculation efficiency and real-time performance, which will significantly improve the segmentation adaptability and practicality of the model in industrial fields for low-grade graphite ore. BRIEF DESCRIPTION OF DRAWINGS

[0053] Figure 1 A work flow chart of a graphite ore image segmentation method based on an improved YOLO11-seg model as claimed by the application;

[0054] Figure 2 A partial ore image example schematic diagram of a graphite ore image segmentation method based on an improved YOLO11-seg model as claimed by the application;

[0055] Figure 3 A data labeling example schematic diagram of a graphite ore image segmentation method based on an improved YOLO11-seg model as claimed by the application;

[0056] Figure 4 A initial YOLO11-seg instance segmentation model structure schematic diagram of a graphite ore image segmentation method based on an improved YOLO11-seg model;

[0057] Figure 5 An improved YOLO11-seg instance segmentation model structure schematic diagram of a graphite ore image segmentation method based on an improved YOLO11-seg model;

[0058] Figure 6 A GSConv convolution flow schematic diagram of a graphite ore image segmentation method based on an improved YOLO11-seg model;

[0059] Figure 7 A C3k2-Faster and Faster Block module structure diagram of a graphite ore image segmentation method based on an improved YOLO11-seg model;

[0060] Figure 8 A PConv convolution schematic diagram of a graphite ore image segmentation method based on an improved YOLO11-seg model;

[0061] Figure 9 A P-R curve diagram of a graphite ore image segmentation method based on an improved YOLO11-seg model;

[0062] Figure 10 A P-R mask curve diagram of a graphite ore image segmentation method based on an improved YOLO11-seg model;

[0063] Figure 11 A detection effect diagram of a graphite ore image segmentation method based on an improved YOLO11-seg model. DETAILED DESCRIPTION

[0064] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of the present application.

[0065] The terms "first", "second", "third", etc. in the present application are only for descriptive purpose and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the technical features indicated. Thus, the features defined with "first", "second", "third" can explicitly or implicitly include at least one of the features. In the description of the present application, the meaning of "a plurality of" is at least two, such as two, three, etc., unless otherwise explicitly and specifically limited. All directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of the present application are only used to explain the relative position relationship, movement condition, etc. between components in a certain posture (as shown in the drawings), and if the certain posture changes, the directional indications also change accordingly. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device including a series of steps or units is not limited to the listed steps or units, but can optionally include steps or units not listed, or can optionally include other steps or units inherent to the process, method, product or device.

[0066] Reference herein to "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment can be included in at least one embodiment of the application. The appearance of the phrase in various places in the specification does not necessarily all refer to the same embodiment, nor is it necessarily mutually exclusive of other embodiments. It is explicitly and implicitly understood that the embodiments described herein can be combined with other embodiments.

[0067] According to a first embodiment of the present application, with reference to Figure 1 The present application claims a graphite ore image segmentation method based on an improved YOLO11-seg model, comprising:

[0068] Collecting ore samples in a graphite mining site to be detected, pre-classifying the ore samples and collecting images to obtain a graphite ore original image library;

[0069] Performing data enhancement and expansion on the graphite ore original images of the graphite ore original image library, and labeling and dividing the graphite ore original image set;

[0070] Constructing an initial YOLO11-seg instance segmentation model, pre-training with a sample data set to obtain trainable parameters and weights of the initial YOLO11-seg instance segmentation model;

[0071] Constructing an improved YOLO11-seg instance segmentation model, initializing parameters of the improved YOLO11-seg instance segmentation model based on the trainable parameters and weights of the initial YOLO11-seg instance segmentation model;

[0072] training the improved YOLO11-seg instance segmentation model based on the parameter initialization and the original image set of the graphite ore, and performing model verification and evaluation to obtain a feasible evaluation result of the improved YOLO11-seg instance segmentation model;

[0073] deploying the improved YOLO11-seg instance segmentation model meeting the requirements in the feasible evaluation result to graphite ore grade detection intelligent equipment to process ore images in an industrial field in real time.

[0074] Further, the graphite ore sample is collected in the graphite mining field to be detected, the graphite ore sample is pre-classified and image collection is performed to obtain a graphite ore original image library, and the method further comprises:

[0075] a plurality of ore samples are collected in the graphite mining field to be detected;

[0076] The plurality of ore samples are pre-classified by using a mineralogical identification method and according to an industrial grade classification standard;

[0077] The plurality of pre-classified ore samples are collected by using an imaging device to obtain a graphite ore original image library containing a plurality of graphite ore original images;

[0078] The graphite ore original image contains the morphology, detailed texture and edge features of the plurality of ore samples.

[0079] In this embodiment, a Canon EOS 5D Mark II single-lens reflex camera is used to collect graphite ore images in an industrial field. To ensure image quality and detail retention, the imaging resolution is set to 5616x3744, and then batch adjusted to 3744x3744 pixels (based on short side cropping) by using Lightroom software to adapt to the model input size, while ensuring uniform distribution of ore particle size and covering various grade forms, thereby improving the representativeness and adaptability of the data. To further enhance the diversity of the data, the camera exposure time is adjusted to 400μs and 600μs to simulate dark light conditions and standard light, respectively, to enrich the feature performance of the samples under different light environments. Some of the images of the graphite ore obtained by shooting are shown in FIG. 5, and the visual features of different graphite ore blocks can be clearly observed to have significant correlation: high-grade ore has less impurities and is typically dark graphite; low-grade and low-medium grade ore has a light color or mixed color due to the presence of quartz and other light-colored impurity minerals. This process obtains a large number of high-resolution mineral images, which provides high-quality original data for subsequent data processing, labeling and model training. Figure 2

[0080] Further, the data augmentation and expansion of the graphite ore original images in the graphite ore original image library, and the labeling and division of the graphite ore original image set further comprise: ​

[0081] The original image of the graphite ore in the original image library of the graphite ore is subjected to data enhancement processing by using Mosaic;

[0082] The original image of the graphite ore in the original image library of the graphite ore is subjected to data enhancement processing by using Mosaic;

[0083] The original image of the graphite ore in the original image library of the graphite ore is subjected to data enhancement processing by using Mosaic;

[0084] In this embodiment, the existing graphite ore image is subjected to data enhancement and expansion, mainly using Mosaic data enhancement strategy, aiming to fully exploit limited sample resources and improve the generalization ability and mineral recognition accuracy of the model. This method selects a center point (x c ,y c ) as the splicing reference, scales four different images and embeds them in the four quadrants of the new image, and synchronously transforms and matches the boundary boxes, so as to ensure the consistency of the labels and the accuracy of the detection. This process not only aims to expand the representation of small-size ore blocks in the image, but also significantly enriches the texture details and spatial structure features that the model can learn. By introducing random rotation, cropping and recombination of images, Mosaic enhancement can effectively improve the diversity of training data, alleviate overfitting, enhance the learning ability of texture granularity and spatial distribution, and is especially suitable for complex visual tasks such as mineral segmentation and recognition.

[0085] The data set is manually annotated using a data annotation tool, and an irregular polygon is used to draw a segmentation mask according to the edge contour of the ore block to segment and mark the target ore block. Then, the training set, the validation set and the test set are automatically divided according to a proper ratio to ensure the reasonable allocation of the data set, thereby improving the generalization ability of the model.

[0086] The data set in step 2 is manually annotated in json format using Labelme image annotation tool, and the medium and low grade is annotated as "below" and the low grade is annotated as "low". For example, Figure 3 , then convert it to txt format with the help of label conversion code, and randomly divide the training set, the validation set and the test set according to the ratio of 8:1:1

[0087] Further, the initial YOLO11-seg instance segmentation model is constructed, pre-trained using sample data sets to obtain trainable parameters and weights of the initial YOLO11-seg instance segmentation model, and further comprising:

[0088] The initial YOLO11-seg instance segmentation model comprises a first Backbone module, a first Neck module and a first Head module;

[0089] The first Backbone module extracts low-level and high-level features of the sample data set to provide semantic information for downstream tasks, and comprises 0th to 10th layers, a plurality of convolution modules Conv and C3k2 modules, an SPPF module and a C2PSA module, and then the features are mapped into the first Neck module;

[0090] The first Neck module enhances the feature expression capability through multi-scale feature fusion and integrates different scale feature information, and comprises 11th to 22nd layers, a plurality of feature fusion and up-sampling module Upsample operations, and a combination of feature splicing module Concat and C3k2 module fusion, and the extracted and optimized features are mapped into the first Head module;

[0091] The first Head module performs target positioning and classification work, and generates a corresponding mask in combination with a segmentation branch, and outputs a target category and a corresponding confidence, and comprises three segmentation modules Segment;

[0092] The initial YOLO11-seg instance segmentation model is trained by using an ImageNet or COCO data set to obtain a pre-training result of the initial YOLO11-seg instance segmentation model;

[0093] The trainable parameters and weights in the pre-training result of the initial YOLO11-seg instance segmentation model are serialized by using a parameter serialization mechanism.

[0094] The trainable parameters at least comprise edge, texture and shape distribution.

[0095] In this embodiment, the initial YOLO11-seg instance segmentation model is denoted as YOLO11-seg-original, and the architecture thereof is as shown in Figure 4YOLO11-seg has the advantages of both accuracy and real-time compared to existing segmentation networks (such as UNet, Mask R-CNN, etc.), which can effectively make up for the shortcomings of traditional mineral segmentation methods and is more suitable for deployment on edge devices to achieve fast and accurate ore image segmentation algorithms. The model consists of three parts: Backbone, Neck, and Head. Backbone is responsible for extracting low-level and high-level features of the image, providing rich semantic information for downstream tasks. Neck enhances the model's expression ability for targets of different scales using a multi-scale feature fusion mechanism, improving robustness in complex scenes. Head is responsible for target positioning and classification, and generates the corresponding mask in combination with the segmentation branch, outputting the target class and corresponding confidence, achieving integrated processing of target detection and instance segmentation, and thus achieving accurate target positioning and segmentation. Specifically, the image is first input to the Backbone feature extraction network for layer-by-layer depth processing. This part consists of multiple convolution units (Conv) and C3k2 feature extraction modules, covering layers 0 to 10 of the model structure, aiming to gradually extract key structural information from the image and enhance feature perception ability at different scales. Then, the feature map enters the Spatial Pyramid Pooling-Fast (SPPF) module to fuse features at multiple receptive fields to improve the model's multi-scale expression ability. Then, the C2PSA module further optimizes the feature distribution strategy and weight adjustment mechanism, effectively improving the overall detection accuracy. Then, on this basis, the feature mapping is sent to the Neck part, which realizes multi-scale feature fusion by multiple upsample, concatenation, and C3k2 module processing to enhance the model's detection adaptability for targets of different sizes. Specifically, layer 11 is upsampled and concatenated with the shallow layer features, and then processed by the C3k2 module for fusion; layer 14 is upsampled again and concatenated with the deep layer features output by Backbone for further processing, finally forming a multi-scale feature map with high fusion of semantic and spatial information. Finally, the feature map optimized by the Neck module is input to the Head end to perform target detection and segmentation tasks, achieving accurate positioning, classification, and mask generation of mineral targets, fully ensuring the stability and accuracy of the model in complex backgrounds and fine-grained textures.

[0096] Obtaining the segmentation pre-training weight. The pre-training weight is usually derived from a model trained on a large-scale general image dataset (such as ImageNet or COCO), and the weight has effectively learned low-level and medium-level features commonly seen in images, such as edges, textures, and shape distributions. Using these pre-training weights as model initialization parameters in specific tasks not only improves the feature extraction capability of the model, but also significantly speeds up the convergence, significantly reduces the training time compared to re-training, and improves the performance of the segmentation model in specific task scenarios.

[0097] Further, the improved YOLO11-seg instance segmentation model is parameterized and weighted based on the initial YOLO11-seg instance segmentation model, and further comprises:

[0098] The improved YOLO11-seg instance segmentation model keeps the overall channel width of the initial YOLO11-seg instance segmentation model unchanged, and proportionally reduces the global depth and maximum channel number, including a second Backbone module, a second Neck module and a second Head module.

[0099] The second Backbone module structure includes the 0th layer to the 10th layer, including a plurality of group spatial convolution modules GSConv and C3k2-Faster modules, SPPF modules and C2PSA modules, and then the feature mapping enters the first Neck module;

[0100] The second Neck module includes the 11th layer to the 22nd layer, which is operated by multiple feature fusion and up-sampling modules Upsample, and is combined with feature splicing modules Concat and C3k2-Faster modules for fusion. The extracted and optimized feature mapping is input to the second Head module;

[0101] The second Head module includes three segmentation modules Segment.

[0102] Load the trainable parameters and weight data of the initial YOLO11-seg instance segmentation model as the initial weight of the improved YOLO11-seg instance segmentation model.

[0103] Further, the group spatial convolution module GSConv further comprises:

[0104] When performing group convolution, for an input tensor The channel number is c1, the height is h, the width is w, and the convolution process of the standard convolution SC branch is represented as:

[0105]

[0106] wherein "*" represents a convolution operation, and c2 represents the number of output channels;

[0107] The deep separable convolution DSC branch is represented as:

[0108]

[0109] The outputs of the two branches are spliced Concat, and the effective fusion of the standard convolution SC feature and the deep separable convolution DSC output is realized through a channel shuffling operation Shuffle, to generate an optimized feature map;

[0110] The overall process of GSConv is represented as:

[0111]

[0112] In this embodiment, an improved YOLO11-seg instance segmentation model is built, named YOLO11-seg-improved, and its structure is as shown in Figure 5 The model aims to further enhance the capture ability of texture details in the graphite ore image, improve the extraction efficiency of multi-scale target features, and especially optimize the segmentation accuracy of low-grade ore blocks. Under the premise of ensuring the overall lightweight and real-time response capability of the model, the segmentation performance is comprehensively improved through structural improvement of the module. Based on the original YOLO11-seg-original model, the present application mainly makes the following improvements around the segmentation task of different low-grade graphite ores:

[0113] First, to optimize the network depth redundancy and realize a more efficient network architecture, the global depth and maximum channel number are reduced in proportion while keeping the overall channel width of the YOLO11-seg baseline model unchanged, so that the improved model is more lightweight and compact than the baseline model, which is the first improvement point.

[0114] Secondly, a lightweight grouped spatial convolution (GSConv) is used to replace the conventional convolution as the downsampling operation of the improved model, which is the second improvement point, Figure 6 which directly represents the process of its grouped convolution.

[0115] The core advantage of GSConv is that it skillfully combines the advantages of standard convolution and deep separable convolution through channel shuffling operation, while preserving the implicit connection between channels and significantly reducing the computational complexity. This downsampling strategy effectively maintains the prediction accuracy of the model while avoiding additional computational overhead.

[0116] Finally, the Faster Block bottleneck structure is integrated into the feature extraction module C3k2 to construct the improved module C3k2-Faster, which is the third improvement point, and its detailed structure is shown in Figure 7 . This module continues the traditional branch structure to enhance the diversity of feature extraction: the input features are first processed by the CBS module (Conv-Batch Norm-SiLU), and then divided into multiple subspaces by the Split operation and sent into multiple Faster Block bottlenecks for feature transformation. Subsequently, the outputs of each sub-branch are aggregated by Concat and optimized again by the CBS module before output to improve feature consistency and expression ability. The Faster Block bottleneck uses PConv 3x3 to extract spatial features, combines CBS and Conv to realize feature expression and channel adjustment, and finally integrates input and output through residual connection to realize efficient information extraction and integration. This module uses partial convolution to significantly accelerate the training and inference process while ensuring performance, especially suitable for practical scenarios with limited computing resources.

[0117] Further explanation, the partial convolution (Partial Convolution, PConv) in Faster Block is to perform convolution operation on only part of the input channels, and the rest are directly mapped to output. This operation helps to optimize memory access and maintain a low floating-point operations per second (Floating-point Operations Per second, FLOPs). The workflow of PConv is shown in Figure 8 . The theoretical calculation formula of FLOPs and memory access (Memory access) of PConv can be expressed as follows:

[0118]

[0119] where h and w represent the width and height of the feature map, k is the size of the convolution kernel, and c p is the number of channels operated by the standard convolution in PConv. When the partial ratio , since the remaining (c-c p ) channels do not participate in convolution, there is no need to access memory, and the memory access of PConv is times that of the regular convolution. The FLOPs of PConv is times that of the regular convolution, which is much less than the computing resources required by the standard convolution.

[0120] Compared with YOLO11-seg-original, the improved module GSConv and C3k2-Faster form a synergistic effect in efficient downsampling and feature extraction, effectively reducing the computational cost while ensuring the accuracy of the segmentation task, providing a more optimal computer vision solution for image segmentation of different grades of graphite ore.

[0121] The improved YOLO11-seg-improved model is initialized for various configuration parameters, including determining the total number of training rounds (epoch), the number of samples processed per batch (batch size), the patience threshold (patience), and the initial learning rate (learning rate). At the same time, an appropriate gradient descent-based optimization algorithm is selected and the related hyperparameters are configured. In addition, quantitative indicators for evaluating the accuracy, segmentation accuracy, and efficiency of the model are defined to monitor and adjust in real time during the training process, thereby further optimizing the overall performance of the model.

[0122] Further, the improved YOLO11-seg instance segmentation model based on the graphite mine original image training parameter initialization is further trained and verified, and the feasible evaluation results of the improved YOLO11-seg instance segmentation model are obtained, which further includes:

[0123] Configure training parameters, and train the improved YOLO11-seg instance segmentation model based on the training set divided from the graphite mine original image based on the training parameters;

[0124] Real-time monitoring of the performance of the improved YOLO11-seg instance segmentation model based on the validation set, after each training cycle, the model evaluation effect is obtained based on the validation set, and the training is determined to continue or terminate;

[0125] The trained improved YOLO11-seg instance segmentation model is evaluated on the test set, and the differences in various experimental indicators are obtained by comparing with the initial YOLO11-seg instance segmentation model, and the feasible evaluation results of the improved YOLO11-seg instance segmentation model are obtained.

[0126] In this embodiment, the YOLO11-seg-improved model is trained using the divided training set, and the model performance is monitored in real time through the validation set. After each training cycle, the performance of the model on the validation set is evaluated immediately, focusing on the change of accuracy and the trend of loss value. If the accuracy continues to improve and the loss value gradually decreases, the training continues; if the accuracy decreases or the loss value increases, the training should be terminated in time to avoid overfitting and save computing resources.

[0127] The trained YOLO11-seg-improved model is obtained, and its performance is evaluated on the pre-divided test set. By comparing with the baseline model YOLO11-seg-original, the differences between the two in various experimental indicators are analyzed, the improvement of the improved modules (GSConv and C3k2-Faster) is summarized, and the potential in practical application is discussed.

[0128] The existing YOLO11-seg-original network is improved, and the specific operation is as follows:

[0129] (i) Keep the width factor unchanged, adjust the depth factor and the maximum number of channels to 1 / 2 of the original network to optimize the global redundancy, and Table 1 shows the detailed comparison of the related parameters of the baseline model;

[0130] Table 1 Comparison of part of parameters before and after network depth scaling

[0131] Model \ Parameter Depth factor Width factor Max channels Model size (MB) YOLO11-seg-orginal 0.5 0.25 1024 5.75 YOLO11-seg-improved 0.25 0.25 512 2.53

[0132] (ii) Introduce a lightweight GSConv downsampling module to replace the Conv modules at layers 0, 1, 3, 5, 7, 17, and 20;

[0133] (iii) Introduce Faster Block bottleneck structure in the original C3k2 module of the network to build an efficient C3k2-Faster feature extraction module, and replace the feature extraction modules at layers 2, 4, 6, 8, 13, 16, 19, and 22. Through the above steps, the improved YOLO11-seg-improved model is built.

[0134] The parameter initialization settings for model training are as follows: the batch size is 8, the total number of training epochs is 150, the initial learning rate is set to 0.01, the momentum coefficient is 0.937, the weight decay factor is 5×10 -4 , the patience threshold is set to 50 rounds, that is, the training is terminated in advance when the performance on the validation set has no significant improvement for 50 consecutive rounds. The optimizer uses the Stochastic Gradient Descent (SGD) algorithm with momentum to improve the convergence speed and reduce the risk of falling into local minimum or saddle point. In terms of model evaluation, the precision indicators include the average precision of detection frame mAP box and the average precision of segmentation mAP maskThe lightweight performance is quantified by the parameters (Params), floating-point operations per second (FLOPs), and frames per second (FPS), which represent the spatial complexity, computational complexity, and temporal complexity of the model, respectively.

[0135] The obtained pre-trained weights are loaded into the YOLO11-seg-improved model, and the dataset processed by Mosaic enhancement is used for training. During the training process, the weight parameters of the model are continuously optimized to improve its recognition ability and robustness for different low-grade graphite ore textures. To ensure the training effect of the model, the accuracy and loss value changes on the validation set are monitored in real time, and the training strategy is adjusted in time to avoid overfitting or underfitting.

[0136] The trained YOLO11-seg-improved model is comprehensively evaluated on the pre-divided test set. Compared with the baseline model YOLO11-seg-original, the mAP box and mAP mask of YOLO11-seg-improved reach 92.7% and 86.4%, respectively, which are slightly better than the baseline model. In terms of lightweight, the parameter quantity is reduced to 1.16M, and the FLOPs is reduced to 7.1G, which are significantly reduced by 59% and 30% compared with the baseline model, respectively. At the same time, the FPS is increased by 8%, which verifies the optimization effect of the improved modules GSConv and C3k2-Faster in terms of segmentation accuracy, feature extraction, and computational efficiency. For specific comparison, see Table 2.

[0137] Table 2 Comprehensive comparison table of YOLO11-seg-improved and baseline model

[0138]

[0139] According to the model data, the Precision-Recall (P-R) curves of the generated detection frame and segmentation mask are drawn as shown in Figure 9 , Figure 10 The segmentation effect is shown in Figure 11 . It shows that the improved model has high accuracy and real-time processing capability, fully meeting the deployment requirements of edge embedded devices.

[0140] In several embodiments provided in the present application, it should be understood that the disclosed system, device and method can be implemented in other manners. For example, the division of the units is only a logical function division, and there can be another division manner for the actual implementation, for example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the displayed or discussed mutual couplings or direct couplings or communication connections between different units, or the among different units, can be indirect couplings or communication connections through some interfaces, devices or units, and can be in electrical, mechanical or other forms.

[0141] In addition, each function unit in the various embodiments of the present application can be integrated into a processing unit, or each unit can exist alone physically, or two or more units can be integrated into one unit. The integrated unit can be implemented in the form of hardware, or in the form of a software functional unit. The above is only an embodiment of the present application, and does not limit the patent scope of the present application. Any equivalent structure or equivalent process transformation using the content of the present application specification and drawings, or direct or indirect application in other related technical fields, are also included in the patent protection scope of the present application.

[0142] The specific embodiments of the application are described in detail above, but they are only examples. The present application is not limited to the specific embodiments described above. Any equivalent modification or substitution to the application by those skilled in the art is also within the scope of the present application. Therefore, any equivalent transformation and modification, improvement, etc. made without departing from the spirit and principle range of the present application should be included in the scope of the present application.

Claims

1. A graphite ore image segmentation method based on an improved YOLO11-seg 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-seg instance segmentation model, use the sample data set for pre-training, and obtain the trainable parameters and weights of the initial YOLO11-seg instance segmentation model; Constructing an improved YOLO11-seg instance segmentation model, and initializing parameters of the improved YOLO11-seg instance segmentation model based on the trainable parameters and weights of the initial YOLO11-seg instance segmentation model; The improved YOLO11-seg instance segmentation model after initialization of training parameters based on the graphite ore original image set is validated and evaluated to obtain a feasible evaluation result of the improved YOLO11-seg instance segmentation model; The improved YOLO11-seg instance segmentation model whose feasible evaluation results meet the requirements is deployed to intelligent equipment for graphite ore grade detection to process ore images in industrial sites in real time.

2. The graphite ore image segmentation method based on the improved YOLO11-seg model according to claim 1, 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 image segmentation method based on the improved YOLO11-seg 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 division of the graphite ore original image set, further include: Using Mosaic to perform data enhancement processing on the graphite ore original image in the graphite ore original image library; 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 image segmentation method based on the improved YOLO11-seg model according to claim 1, characterized in that: The initial YOLO11-seg instance segmentation model is constructed, and pre-trained using a sample data set to obtain trainable parameters and weights of the initial YOLO11-seg instance segmentation model, further comprising: The initial YOLO11-seg instance segmentation model includes a first Backbone module, a first Neck module and a first Head module; The first Backbone module extracts low-level and high-level features of the sample data set to provide semantic information for downstream tasks, 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 positioning and classification, and generates corresponding masks in combination with the segmentation branch, outputs the target category and corresponding confidence, and includes three segmentation modules Segment; the initial YOLO11-seg instance segmentation model is trained using the ImageNet or COCO dataset to obtain the pre-training results of the initial YOLO11-seg instance segmentation model; The trainable parameters and weights in the pre-training results of the initial YOLO11-seg instance segmentation model are serialized using a parameter serialization mechanism; The trainable parameters include at least edge, texture and shape distribution.

5. The graphite ore image segmentation method based on the improved YOLO11-seg model according to claim 4 is characterized in that: The step of constructing an improved YOLO11-seg instance segmentation model and initializing parameters of the improved YOLO11-seg instance segmentation model based on the trainable parameters and weights of the initial YOLO11-seg instance segmentation model further includes: The improved YOLO11-seg instance segmentation model maintains the overall channel width of the initial YOLO11-seg instance segmentation model unchanged, reduces the global depth and the maximum number of channels in proportion, and includes a second Backbone module, a second Neck module, and a second Head module; The second Backbone module structure includes layers 0 to 10, including multiple grouped spatial convolution modules GSConv and C3k2-Faster modules, SPPF modules and C2PSA modules, after which the feature map enters the first Neck module; 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-Faster module fusion, the extracted and optimized feature map is input to the second Head module; the second Head module includes three segmentation modules Segment; Load the trainable parameters and weight data of the initial YOLO11-seg instance segmentation model as the initial weights of the improved YOLO11-seg instance segmentation model.

6. The graphite ore image segmentation method based on the improved YOLO11-seg model according to claim 5, characterized in that: The grouped spatial convolution module GSConv further includes: In grouped convolution, for the input tensor X∈R c1×h×w , the number of channels is c1, the height is h, and the width is w. The convolution process of the standard convolution SC branch is expressed as: Among them, "*" represents the convolution operation, and c2 represents the number of output channels; The depthwise separable convolution DSC branch is expressed as: The outputs of the two branches are concatenated and the channel shuffle operation is used to effectively fuse the standard convolution SC features with the depthwise separable convolution DSC output to generate an optimized feature map. The overall process of GSConv is summarized as follows:

7. The graphite ore image segmentation method based on the improved YOLO11-seg model according to claim 3, characterized in that: The improved YOLO11-seg instance segmentation model after initialization of training parameters based on the graphite ore original image and model verification and evaluation are performed to obtain a feasible evaluation result of the improved YOLO11-seg instance segmentation model, further comprising: Configuring training parameters, and training the improved YOLO11-seg instance segmentation model based on the training parameters using a training set divided by the original graphite mine image; Based on the validation set, the performance of the improved YOLO11-seg instance segmentation 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-seg instance segmentation model is evaluated on the test set. By comparing it with the initial YOLO11-seg instance segmentation model, the differences in various experimental indicators are obtained, and a feasible evaluation result of the improved YOLO11-seg instance segmentation model is obtained.

Citation Information

Patent Citations

  • Graphite ore grade identification method

    CN117975032A

  • Lepidolite ore image segmentation method based on improved SAM model

    CN119478398A

  • Lithium mineral component detection method based on improved YOLOv8

    CN119495372A

  • Dynamic environment dense point cloud SLAM method and system based on YOLOv11 and ORB-SLAM3

    CN119540942A

  • DST-YOLO-based unmanned aerial vehicle aerial photography small target detection method

    CN119992049A

Cited By

  • Ore XRT image instance segmentation optimization method and system based on YOLOv11-seg

    CN121527434A