Semantic segmentation coal piling identification method and device based on large and small model linkage and medium

Abstract

The invention belongs to the technical field of image semantic segmentation, and particularly relates to a semantic segmentation coal piling identification method and device based on large and small model linkage and a medium, and the method comprises the steps: carrying out the rough semantic segmentation through a visual small model and a semantic segmentation method of an SAM large model, and carrying out the rough semantic segmentation through a YOLOV11 small model, and generating an initial mask; performing fine segmentation by using a fine-tuned SAM large model, converting a coarse mask into a cue word through a mask encoder, and inputting the cue word into a mask decoder in combination with image features to output a final segmentation result; the efficiency of the small model and the precision of the large model are combined through linkage of the large model and the small model, and the real-time performance and generalization performance of coal piling recognition are achieved. In addition, SAM fine adjustment adopts a general parameter fine adjustment method, generalization ability is reserved, and meanwhile, the method adapts to a coal piling scene; alarm is triggered by setting a coal piling area threshold value, and the method is suitable for complex underground coal mine environments. According to the invention, image input, coarse segmentation, fine segmentation and automatic alarm based on the area threshold are realized, and a complete automatic closed-loop system without manual intervention is formed.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image semantic segmentation, in particular to a semantic segmentation coal piling recognition method and device based on a large-small model linkage and a medium. BACKGROUND

[0002] In the process of coal transportation and production, the conveyor belt is the key equipment for material transportation. However, due to uneven feeding, equipment failure or belt slip, etc., abnormal accumulation of coal on the conveyor belt often occurs, i.e. "coal piling". If it is not detected and handled in time, coal piling will cause a series of serious consequences, including but not limited to belt wear, deviation, motor overload, and even fire caused by friction heat, which seriously threatens the safety of mine production.

[0003] Traditional coal piling detection relies on manual inspection, which is inefficient. With the progress of technology, current underground coal piling detection mainly relies on several technical solutions: (1) Physical sensor-based solutions: This type of solution includes mechanical travel switches, laser radars, and depth cameras. Mechanical switches have problems such as easy wear, response lag, and the need for contact installation, and have low reliability. Laser radars can perform non-contact detection, but their equipment cost is high, and in the high-dust environment of coal mines, the laser beam is easily scattered and absorbed, resulting in poor point cloud data quality, inaccurate measurement, and high false alarm rate. Depth cameras usually require a long exposure time due to their working principle, and in the case of rapid coal piling, there is significant detection delay, which cannot meet the needs of real-time early warning.

[0004] (2) Traditional computer vision-based solutions: With the development of industrial intelligence, visual monitoring solutions based on visible light cameras and image recognition algorithms have emerged. Among them, semantic segmentation methods based on lightweight convolutional neural networks have attracted attention because they can achieve pixel-level recognition. However, such small models have limited model capacity and generalization ability in pursuit of inference speed. In the face of complex and variable conditions in coal mines, such as dramatic changes in lighting, large amounts of dust and water vapor interference, and differences in equipment layout and coal texture between different mines, trained small models are prone to poor performance, false detection and missed detection, poor generalization ability, and difficulty in direct promotion in different mining areas.

[0005] (3) Visual large-scale model-based solutions: In recent years, the emergence of visual large-scale models such as SAM (Segment Anything Model) has brought revolutionary progress to image segmentation. These models are trained on massive amounts of data and have powerful zero-shot generalization ability and fine segmentation accuracy. Related existing technologies, such as Chinese patent CN119863623A, propose a self-prompting semantic segmentation method based on SAM, which automatically generates prompt boxes by integrating target detectors, reducing manual intervention and realizing end-to-end industrial defect detection. However, directly transferring this technology to the coal pile detection scenario still faces challenges: First, the SAM model itself has a huge number of parameters, and direct deployment requires high computing resources for edge devices, making it difficult to meet the low latency requirements of real-time monitoring in underground mines; Second, although its generalization ability is strong, it still needs to be effectively fine-tuned to make it accurately adapt to the subtle features of the specific scenario of coal piles, and full parameter fine-tuning of SAM is extremely costly and has the risk of catastrophic forgetting, which may weaken its inherent general advantages.

[0006] Therefore, there is an urgent need in this field for a solution that can integrate the advantages of both large and small visual models, and take into account all-weather, non-contact, and high-precision visual recognition technology for coal piles. Summary of the Invention

[0007] In view of this, the present invention aims to solve the problems of high cost and performance degradation of lidar solutions in dusty environments, detection delay of depth cameras, poor generalization ability and insufficient adaptability of semantic segmentation methods based on pure vision small models in different coal mine environments, and high cost of large-scale semantic segmentation models when directly applied to coal pile scenarios due to large computational load, poor real-time performance and high cost of direct full parameter fine-tuning. The present invention provides a semantic segmentation coal pile recognition method, device and medium based on the linkage of large and small models.

[0008] In a first aspect, the present invention provides a semantic segmentation coal pile identification method based on the linkage of size models.

[0009] Includes the following steps: S1 Image Acquisition: Real-time acquisition of images of the conveyor belt's operating status using cameras deployed underground in the coal mine; S2 Small Model Coarse Segmentation: The acquired image is input into a pre-trained semantic segmentation small model, which quickly generates the initial segmentation result of the coal pile area, i.e., the coarse segmentation mask; S3 Large Model Fine Segmentation: The original image is input into an image encoder of a specially fine-tuned semantic segmentation large model to obtain high-dimensional image features; the coarse segmentation mask is input into the mask encoder of the fine-tuned semantic segmentation large model to encode and generate mask prompt features containing the shape and location information of the coal pile area; the above encoded features and mask prompt features are input together into the mask decoder of the fine-tuned semantic segmentation large model for fine pixel-level segmentation and output fine segmentation mask; S4 Decision Alarm: Calculates the pixel area of ​​the coal pile region in the fine segmentation mask. When the pixel area exceeds the preset threshold, the coal pile alarm signal is automatically triggered. The fine-tuned semantic segmentation model is a fine-tuned SAM model, which is fine-tuned using a general parameter fine-tuning method to adapt to the coal pile scenario.

[0010] Furthermore, the semantic segmentation mini-model is preferably a YOLO series model. During training, its network structure integrates a dual-scale attention mechanism in the feature fusion layer to capture local detail features and global context features in parallel, and performs feature enhancement through channel-space joint attention.

[0011] Furthermore, the training process of the semantic segmentation mini-model includes: A coal pile image dataset was constructed, and data augmentation techniques such as dynamic scaling, random cropping, and adaptive MixUp were employed. In terms of model structure, a dual-scale attention mechanism is introduced into the feature fusion layer to enhance feature representation capabilities; The model is trained using the SGD optimizer and cosine annealing learning rate scheduling, and finally lightweighted by knowledge distillation and model pruning.

[0012] Furthermore, the fine-tuning of the semantic segmentation large model employs a general parameter fine-tuning method. The core of this method is to fine-tune each Transformer attention module in the SAM large model image encoder and mask decoder. and Original parameters of the matrix During the fine-tuning process, a fine-tuning parameter is added to each fine-tuning dataset. : ;in, It is a two-dimensional matrix representing the fine-tuning parameters used in the fine-tuning process of the u-th dataset.

[0013] Next, fine-tune the parameters of all datasets. Merge into a three-dimensional matrix and to Perform low-rank decomposition and recombination: ;in, , , All are low-rank matrices, with matrix dimensions of respectively , and ; , express and The input and output dimensions of the matrix. , This represents the rank of a low-rank matrix and can be adjusted as a hyperparameter during fine-tuning. This indicates the number of datasets used in the multi-dataset fine-tuning process.

[0014] and The parameter matrix for the general part is pre-trained using a large-scale general segmentation dataset and kept frozen to ensure that the model retains strong generalization ability. The parameter matrix for the dedicated part is fine-tuned and updated using only the underground coal pile dataset, so that the model can be accurately adapted to the coal pile scenario.

[0015] The process of fine-tuning the parameters is divided into two stages.

[0016] Phase 1: [Regarding] and The matrix is ​​fine-tuned. Training is performed using a general object segmentation dataset, with supervision using binary cross-entropy loss and Dice loss as segmentation loss functions.

[0017] Phase Two: Fine-tuning a single dataset given multiple datasets, fixing... and Parameter matrix, only for The parameter matrix is ​​fine-tuned. Fine-tuning is supervised using the same loss function as the first stage.

[0018] This method effectively avoids catastrophic forgetting and achieves efficient and high-performance adaptation of large models to specific scenarios with a very small number of parameters.

[0019] Secondly, the present invention provides a semantic segmentation coal pile identification device based on the linkage of size models.

[0020] The apparatus corresponds to the method described in the first aspect, comprising: (1) Image acquisition module, used to acquire images of the coal mine conveyor belt to be detected; (2) Coarse segmentation module, configured with a semantic segmentation mini-model, used to process the image and output a coarse segmentation mask for the coal pile area; (3) A prompting encoding module, used to encode the coarse segmentation mask and generate mask prompting features; (4) Image encoding module, configured with an image encoder of a finely tuned semantic segmentation large model, used to extract features from the image and obtain image encoding; (5) Fine segmentation module, equipped with a mask decoder of a finely tuned semantic segmentation large model, used to fuse the mask prompt features and the image features to generate a fine segmentation mask for the coal pile area; (6) Decision alarm module, which determines the coal pile status based on the fine segmentation mask and outputs early warning information; The semantic segmentation model is fine-tuned using a general parameter fine-tuning method.

[0021] Furthermore, the semantic segmentation mini-model is a YOLOv11 model, and its network structure integrates a dual-scale attention mechanism in the feature fusion layer.

[0022] Thirdly, the present invention provides an electronic device and a readable storage medium.

[0023] An electronic device includes a memory, a processor, and a computer program stored in the memory. When the processor executes the program, it is able to implement all the steps of the semantic segmentation coal pile identification method based on size model linkage described in this invention.

[0024] A computer-readable storage medium, such as a USB flash drive, hard disk, or server storage space, stores a computer program that, when executed by a processor, can implement all the steps of the semantic segmentation coal pile identification method based on size model linkage described in this invention.

[0025] The beneficial effects of this invention are as follows: (1) By using the small visual model for coarse localization and the large SAM model for fine segmentation, combined with the fine-tuning of the general parameters of the large SAM model, the efficiency of the small model is utilized, and the segmentation accuracy and cross-scene generalization ability of the large model are fully utilized. It can stably cope with the complex environment of different mines and has high precision and strong generalization ability.

[0026] (2) Using ordinary industrial cameras, the hardware cost is lower than that of lidar. Through model lightweighting and efficient parameter fine-tuning, the requirements for deployment hardware computing power are reduced, making it easy to run on resource-constrained edge devices, thus reducing costs and facilitating deployment.

[0027] (3) Two-level model verification and area-based threshold judgment reduce false alarm and false alarm rates and improve system reliability; realize end-to-end automated process from image acquisition to alarm decision-making, and reduce manual intervention.

[0028] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0029] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein: Figure 1 This is a diagram of the linkage coal pile segmentation architecture of the large and small models of this invention; Figure 2 This is a diagram of the YOLOv11 small model network architecture; Figure 3 This is a schematic diagram of the parameter matrix decomposition of the SAM large model. Detailed Implementation

[0030] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0031] The accompanying drawings are for illustrative purposes only and are schematic diagrams, not actual pictures. They should not be construed as limiting the invention. To better illustrate the embodiments of the invention, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0032] In the accompanying drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components. In the description of the present invention, it should be understood that if terms such as "upper," "lower," "left," "right," "front," and "rear" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the terms used to describe positional relationships in the drawings are only for illustrative purposes and should not be construed as limiting the present invention. For those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.

[0033] This invention provides a semantic segmentation method, device, and medium for coal pile identification based on the linkage of large and small models. The invention will be further described in detail below with reference to the accompanying drawings.

[0034] Example 1: A semantic segmentation coal pile identification method based on the linkage of size models.

[0035] This embodiment details the implementation process of the coal pile identification method based on the linkage of large and small models. Figure 1 The diagram shows the coal stacking and segmentation architecture of the large and small models.

[0036] 1. System Construction and Data Preparation Industrial cameras are installed at key locations such as the head and tail of conveyor belts in underground coal mines, and connected to industrial control computers or edge computing devices with certain computing capabilities.

[0037] Construct a coal pile dataset, obtain a training set containing N images, each image is labeled with M instances, and create a directory structure containing a training set, a validation set and a test set, where the ratio of the training set to the validation set is 8:2.

[0038] 2. Training and Deployment of Small Semantic Segmentation Models (1) Data preprocessing: The training images are dynamically scaled while maintaining the aspect ratio, and randomly cropped to a resolution of 640×640. Adaptive MixUp enhancement is applied, and the mixing ratio is dynamically adjusted between 0.1 and 0.3.

[0039] (2) Model initialization: Load the pre-trained YOLOv11n model weights, retain the feature extraction part, and reset the segmentation head part. Optimize the network structure according to the segmentation task, and introduce the DSAM attention mechanism in the feature fusion layer. The specific implementation is as follows: adopt a multi-scale feature fusion strategy, spatially align the feature maps of different levels to design a dual-scale attention module, extract local detail features and global context features through parallel paths, and fuse them through channel and spatial dimension attention weighting.

[0040] (3) Enhanced feature representation training: The SGD optimizer was used with an initial learning rate of 0.01. Cosine annealing was used to schedule the learning rate and the batch size was set to 16. The training lasted for 200 epochs. In the later stages of training (epochs > 150), a progressive disabling of Mosaic enhancement was adopted to gradually reduce the intensity of Mosaic enhancement. Gradient clipping was applied with a threshold of 1.0.

[0041] (4) Model Evaluation and Optimization: When the Mask mAP50 metric of the model is less than 0.68 on the validation set, the learning rate and data augmentation parameters are adjusted based on the validation set performance, and the model is retrained. During model deployment, model compression is performed first. Knowledge distillation technology is used to use the trained YOLOv11n-seg model as the teacher model to generate soft labels and train a smaller model. Channel pruning technology is applied to remove redundant channels and retain important feature channels. The trained final model is converted to ONNX format and quantized using TensorRT to convert the floating-point model to an INT8 quantized model to further improve inference speed.

[0042] 3. Fine-tuning of general and specific parameters of the large-scale semantic segmentation SAM model To address the unique characteristics of underground coal storage scenarios, a method for fine-tuning general parameters is employed. The specific process is as follows: Figure 2 As shown.

[0043] The SAM large model is decomposed into parameter matrices containing general and specific parts, where: the parameter matrix of the general part is... and The SAM model is trained using a large amount of general image data, enabling it to retain strong generalization ability and segment various objects; the parameter matrix of the specialized part is... The training uses coal pile data collected underground, and only updates are performed during the training of the coal pile scenario. Partially. Provided the model retains general knowledge, it can exhibit strong recognition capabilities for coal piles.

[0044] For each Transformer attention module in the SAM image encoder and mask decoder and Matrix, let its original parameters be... During the fine-tuning process, a fine-tuning parameter is added to each fine-tuning dataset. :

[0045] in, It is a two-dimensional matrix representing the fine-tuning parameters used in the fine-tuning process of the u-th dataset.

[0046] Next, fine-tune the parameters of all datasets. Merge into a three-dimensional matrix and to Perform low-rank decomposition. Figure 3 Schematic diagram of matrix decomposition:

[0047] in, , , All are low-rank matrices, with matrix dimensions of respectively , and ; in, , express and The input and output dimensions of the matrix. , This represents the rank of a low-rank matrix and can be adjusted as a hyperparameter during fine-tuning. This indicates the number of datasets used in the multi-dataset fine-tuning process.

[0048] The fine-tuning process is divided into two stages.

[0049] Phase 1: [Regarding] and The matrix is ​​fine-tuned. Training is performed using the general object segmentation dataset COCO, with binary cross-entropy loss and Dice loss used as segmentation loss functions for supervision.

[0050] Phase Two: Fine-tuning a single dataset given multiple datasets, fixing... and Parameter matrix, only for The parameter matrix is ​​fine-tuned. Fine-tuning is supervised using the same loss function as the first stage.

[0051] The binary cross-entropy loss function supervises each segmented pixel, while the Dice loss function supervises the intersection-union ratio (IoU) between the predicted and ground truth values. Combining these two loss functions allows the model to focus on both local and global segmentation, improving its convergence.

[0052] 4. Coal pile identification and alarm linked by large and small models First, the cameras at the coal pile identification point collect video streams in real time and decode them into image frames.

[0053] Then, each frame of image is input into the deployed YOLOv11 mini-model for preliminary semantic segmentation to obtain a coarse segmentation mask.

[0054] Secondly, each frame of image is input into the image encoder of the fine-tuned SAM large model to obtain the image code; the coarse segmentation mask is input into the mask encoder of the fine-tuned SAM large model to obtain the mask cue features; the image code and the mask cue features are input into the mask decoder of the fine-tuned SAM large model to obtain the final fine coal pile segmentation mask output.

[0055] Finally, the system calculates the total pixel area S of the coal pile in the mask, setting a coal pile area threshold of 1600. This threshold can be adjusted according to the camera's installation position and angle, as well as the image resolution. Under normal circumstances, coal piles will not occur, and S=0. As coal piles occur and the process progresses, the coal pile area S will gradually increase. When S≥1600, an audible and visual alarm signal is immediately triggered, and the alarm information is uploaded to the monitoring center to remind on-site personnel and managers to take immediate action to ensure that no serious accidents occur during the coal transportation process.

[0056] Example 2: A semantic segmentation coal pile identification device based on size model linkage.

[0057] This embodiment provides an apparatus for implementing a semantic segmentation coal pile identification method based on size model linkage. It includes: (1) Image acquisition module: It consists of an industrial camera and a video decoding unit, which is responsible for acquiring video streams and decoding them into image frames.

[0058] (2) Coarse segmentation module: integrated into industrial control computer or edge computing device, with built-in lightweight YOLOv11 small model for outputting coarse segmentation mask.

[0059] (3) Prompt Encoding Module: This is a software module that calls the mask encoder of the SAM large model after fine-tuning the parameters to encode the coarse segmentation mask into feature vectors, i.e., mask prompt words.

[0060] (4) Image encoding module: The built-in SAM image encoder with fine-tuned parameters is used to extract image depth features and obtain image encoding.

[0061] (5) Fine segmentation module: It has a built-in SAM mask decoder with fine-tuned parameters, which is responsible for fusing features and outputting fine coal pile segmentation mask.

[0062] (6) Decision alarm module: includes area statistics unit and threshold comparator. When it is determined that the coal pile area exceeds the threshold, the alarm is driven to issue an alarm.

[0063] Example 3: A semantic segmentation storage medium for coal pile identification based on size model linkage.

[0064] A USB flash drive, hard disk, or server storage space storing a computer program that, when executed by a processor, can perform the steps of the method described in Embodiment 1.

[0065] In summary, this invention employs a semantic segmentation method that combines a small model with a large model. After initial segmentation using a YOLOv11 small model, a finely tuned large model is used to obtain accurate coal pile segmentation results, combining the efficiency of the small model with the accuracy of the large model. By introducing the SAM large model, the generalization performance of the overall system is improved, and the pure vision method overcomes the problems of high cost of LiDAR and long exposure time of depth camera.

[0066] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A size model linkage-based semantic segmentation coal piling recognition method, characterized in that, The method comprises the following steps: S1 image acquisition: acquiring an image of a coal mine conveying belt to be detected; S2 small model rough segmentation: inputting the image into a pre-trained semantic segmentation small model to obtain a rough segmentation mask of the coal stacking area; S3 large model fine segmentation: inputting the image into an image encoder of a fine-tuned semantic segmentation large model to obtain image features; inputting the rough segmentation mask into a mask encoder of the fine-tuned semantic segmentation large model to obtain mask prompt features; and inputting the image features and the mask prompt features into a mask decoder of the fine-tuned semantic segmentation large model to obtain a fine segmentation mask of the coal stacking area; S4 decision alarm: determining a coal stacking state based on the fine segmentation mask and outputting early warning information; wherein the semantic segmentation large model is a SAM large model, and a general-special parameter fine-tuning method is used for fine-tuning to adapt to the coal stacking scene.

2. The method of claim 1, wherein, The semantic segmentation small model is a YOLO series model, and in the training process, a double-scale attention mechanism is integrated into the network structure at a feature fusion layer to capture local detail features and global context features in parallel, and feature enhancement is performed through channel-spatial joint attention.

3. The method of claim 1, wherein, The steps of the general-special parameter fine-tuning method comprise: performing low-rank decomposition on a parameter matrix in the semantic segmentation large model to obtain general part and special part parameter matrices; pre-training and keeping frozen the general part parameters using a general target segmentation dataset; only fine-tuning and updating the special part parameters using a coal stacking scene dataset; in the fine-tuning stage, binary cross-entropy loss and Dice loss are used as segmentation loss functions for supervision.

4. The method of claim 3, wherein, The specific process of the low-rank decomposition is: The parameter matrix W is decomposed into wherein, each Transformer attention module in the SAM large model image encoder and the mask decoder and the original parameters of the matrix, the fine-tuning parameters added in the u-th data set fine-tuning process; Fine-tuning parameters for all datasets Merging into a three-dimensional matrix , for Low-rank decomposition and reorganization: where, , , are low-rank matrices with matrix dimensions , and ; , denote the input and output dimensions of and matrices, , denote the rank of low-rank matrices, which can be adjusted as hyperparameters during fine-tuning, denotes the number of datasets in the multi-dataset fine-tuning process.

5. The method according to claim 3 or 4, characterized in that, The general-special parameter fine-tuning method adopts a two-stage training strategy: In the first stage, the general part parameters and the special part parameters are trained simultaneously using a general image segmentation dataset; In the second stage, the general part parameters are fixed, and only the special part parameters are fine-tuned using a coal stacking scene dataset.

6. A size model linkage-based semantic segmentation coal piling recognition device, characterized in that, comprise: (1) an image acquisition module for acquiring an image of a coal mine conveying belt to be detected; (2) a rough segmentation module configured with a semantic segmentation small model for processing the image and outputting a rough segmentation mask of a coal stacking area; (3) a prompt encoding module for encoding the rough segmentation mask to generate mask prompt features; (4) an image encoding module configured with an image encoder of a fine-tuned semantic segmentation large model for feature extraction of the image to obtain image encoding; (5) a fine segmentation module configured with a mask decoder of the fine-tuned semantic segmentation large model for fusing the mask prompt features and the image features to generate a fine segmentation mask of the coal stacking area; (6) a decision alarm module for determining a coal stacking state based on the fine segmentation mask and outputting early warning information; wherein the semantic segmentation large model is a SAM large model, and a general-special parameter fine-tuning method is used for fine-tuning to adapt to the coal stacking scene.

7. The apparatus of claim 6, wherein, The training process of the semantic segmentation small model comprises: Data preprocessing, dynamic scaling and random cropping of training images to a preset resolution, adaptive MixUp data enhancement, with a mixing rate dynamically adjusted between 0.1 and 0.3; Model initialization, load pre-trained model weights, retain the feature extraction part, reset the segmentation head part, optimize the network structure according to the segmentation task, and introduce an attention mechanism at the feature fusion layer; Enhance feature representation capability training, use the SGD optimizer, and use the progressive closing Mosaic enhancement strategy in the later training stage; Model evaluation and optimization, evaluate the model performance on the validation set, calculate the target indicators, and determine whether to retrain according to the target achievement.

8. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the program to implement the method of any one of claims 1 to 5.

9. A computer readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the method of any one of claims 1 to 5.

Images