Lightweight detection model for surface defects of cage guide and detection method thereof

Abstract

The invention discloses a cage guide surface defect lightweight detection model, which comprises a trunk feature extraction network, a neck feature fusion network and a prediction output network, and is characterized in that the trunk feature extraction network is composed of a dynamic space adaptive fusion module DSAF, five CBSs, three C2f and a SPPF module; the neck feature fusion network is composed of four dynamic efficient local attention fusion modules DELA-C2f, two CBSs and two up-sampling and splicing flexible operations; the prediction output network adopts a brand-new GH-Detect detection head module, and adopts a sharing idea to construct a GH-Block module to form a main path. According to the invention, high-precision detection of the surface defects of the cage guide is facilitated, the industrial application requirement of defect detection of the underground coal mine cage guide can be met, and the system can be conveniently deployed on embedded equipment with limited computer resources.

Description

Technical Field

[0001] The present invention belongs to the field of target detection and relates to a lightweight detection model for tank channel surface defects and a detection method thereof. Background Art

[0002] Coal is China's primary energy source, and the safe and efficient extraction of coal resources is crucial to economic stability and sustainable development. The vertical shaft, the primary access point between miners and the surface, houses a canal that guides the elevator cage. Over time, canal defects such as wear, cracks, and rust can develop. If not addressed promptly, minor defects can affect the safe and stable operation of the elevator cage; severe ones can lead to accidents, resulting in significant economic losses. Therefore, accurate detection of canal defects in coal mines is crucial.

[0003] At present, the detection method of surface defects in vertical shaft tank tunnels mainly relies on manual inspections, which is not only time-consuming and labor-intensive, but also limited by the experience and subjective judgment of inspectors, making it difficult to ensure the accuracy and efficiency of detection. In order to solve the problems of low detection efficiency and accuracy, researchers in related fields have proposed an automated detection method for tank tunnel defects. Hu Wenbin et al. proposed to perform data algorithm processing on the tank tunnel scanning data extracted by a two-dimensional laser scanner, and write the algorithm into the host computer software to detect defects, thereby improving the detection efficiency. For details, see "Hu Wenbin, Wu Feng, Gan Weibing, et al. Tank tunnel detection algorithm based on two-dimensional laser scanning technology [J]. Infrared and Laser Engineering, 2021, 50(10): 227-233.". However, the above method can improve the detection accuracy, but the underground environment of the mine is complex, and the data obtained by the instrument is easily interfered by cables and noise, resulting in insufficient defect detection accuracy, which is difficult to meet the actual needs of the industry. With the development of deep learning, deep learning has gradually been applied to the industrial field of defect detection.

[0004] With the development of computer vision and deep learning, deep learning methods based on convolutional neural networks have demonstrated strong feature extraction and representation capabilities, as well as advantages such as fast detection speed, high accuracy, and low cost. Consequently, deep learning has been increasingly applied to the defect detection industry. Current object detection algorithms are primarily categorized as single-stage and two-stage. Typical two-stage algorithms include Faster R-CNN and Mask R-CNN. They first acquire candidate regions for the object to be detected, then generate candidate bounding boxes within these regions for regression prediction to obtain the final predicted boxes. Single-stage algorithms, such as SSD and YOLO, directly generate multiple candidate regions within the input image and classify them based on the object's type and location. Due to their simplified image processing steps, single-stage object detection algorithms are faster than two-stage algorithms and better meet the needs of real-time industrial detection. Xiang Kuan et al. proposed an improved Faster R-CNN deep learning network to detect aluminum surface defects. They introduced a feature pyramid network (FPN) structure into the backbone network to enhance the feature extraction capability of small defects, and then added a K-means algorithm to cluster the defect data to obtain an anchor frame that is more suitable for aluminum defects. For details, see "Xiang Kuan, Li Songsong, Luan Minghui, et al. Aluminum surface defect detection method based on improved Faster RCNN [J]. Chinese Journal of Scientific Instruments, 2021, 42(1): 191-198." Zhang et al. proposed a new detection system based on YOLOv3. By replacing the original ordinary convolution with a multi-scale convolution module, they can obtain spatial feature information at different scales in the image and improve the detection accuracy. For details, see "ZHANG JF, HU YW, JIS J. Angle steel tower bolt defect detection based on YOLO-V3 [J]. ITM Web of Conferences, 2022, 45." Li Xin et al. proposed a steel surface defect detection algorithm based on improved YOLOv5. First, the C3 module in Backbone was replaced with the GhostBottleneck structure, and then the SE attention mechanism was introduced to strengthen important feature channels; a feature fusion structure was added to the Neck part, and DW convolution was used to replace some standard convolutions to reduce the amount of calculation. For details, see "Li Xin, Wang Cheng, Li Bin, et al. Improved YOLOv5 steel surface defect detection algorithm [J]. Journal of Air Force Engineering University: Natural Science Edition, 2022, 23(2): 26-33."Du et al. proposed a rigid tank guide fault detection algorithm based on improved YOLOv7. They added an attention mechanism at different positions in the network and replaced the loss function, which improved the detection accuracy. However, the model complexity is relatively high and difficult to deploy on edge computing devices. For details, see "Du F, Mo DD, Ma TB, et al. Rigid tankguide fault detection algorithm based on improved YOLOv7[J]. Journal of Real-Time Image Processing, 2025, 22(1): 1-14.". Although the above methods can improve detection accuracy, it is difficult to balance the problems of detection speed, accuracy and lightweight.

[0005] The above method is difficult to find the optimal balance between lightweight, detection accuracy and multi-scale recognition. It cannot meet the industrial needs of tank road defect detection in complex mine environments and is difficult to deploy on low-computing-power embedded devices. Summary of the Invention

[0006] In order to solve the above problems, the present invention proposes a lightweight detection model and detection method for tankway surface defects, which can increase the accuracy of the model in detecting tankway surface defects while reducing the number of model parameters, and at the same time meet the industrial application needs of tankway defects in underground coal mines.

[0007] The technical solution adopted by the present invention is a lightweight detection model for tank surface defects, which includes three parts: a backbone feature extraction network, a neck feature fusion network and a prediction output network;

[0008] The backbone feature extraction network consists of an adaptive fusion module DSAF, 5 CBS, 3 C2f modules, and 1 SPPF module;

[0009] The neck feature fusion network consists of four dynamic and efficient local attention fusion modules DELA-C2f, two CBS, two UpSample upsampling layers, and two Concat layers;

[0010] The prediction output network consists of three GH-Detect detection heads, each of which extracts image features through two GH-Block modules in cascade and two Conv2d modules in parallel;

[0011] Furthermore, the described lightweight detection model for tankway surface defects is characterized in that a DSAF module is designed in the backbone network, the backbone branch is embedded with dynamic snake convolution, and the other branch extracts feature map context information through global average pooling of the feature map, obtains a global receptive field, improves adaptability to the complex environment under the mine, and solves the problem of insufficient feature extraction of the backbone network; a DELA-C2f module is designed in the neck network to improve the network's ability to focus on defect information during the feature extraction process, effectively reduce the impact of irrelevant information, and thus improve detection accuracy; a GH-Detect detection module is constructed, and a shared GH-Block module is used to reduce the model complexity, thereby achieving model lightweighting while maintaining the accuracy as much as possible, and facilitating deployment to embedded devices.

[0012] The DSAF module first uses 1×1 standard convolution for feature extraction, then uses dynamic snake convolution DSConv to capture slender crack features, uses residual branches to perform adaptive average pooling operations on the feature image in the H and W dimensions, compresses the H×W×C input feature map into a 1×1×C feature vector, then performs an excitation operation using a full connection, and then uses an upsampling operation to make the feature map the same size as the feature map from the main branch. The Concat operation is used to fuse features at two different levels, and then another standard convolution is used to extract detailed features.

[0013] The DELA-C2f module improves the Bottleneck module by fusing 3×3 standard convolution and improved efficient local attention (ELA) to obtain the DE-Block module structure. It first uses 1×1 standard convolution for feature extraction, then enhances local feature extraction through 3×3 standard convolution and D-ELA in parallel, and then performs a Concat fusion operation. Finally, it extracts more detailed information through 1×1 standard convolution and then adds it to the input feature map to obtain the output, effectively reducing feature redundancy and protecting multi-scale contextual information. The improved efficient local attention mechanism D-ELA module structure introduces an intermediate branch based on ELA, which includes point-by-point convolution, normalization and activation function. The module first performs average pooling on the height and width dimensions of the input feature map X, and then uses Conv1d for faster and lighter local processing. It then refines the feature map through group normalization and nonlinear activation function, and generates the final output feature maps Y from two directions. h and Y wThe intermediate branch uses PWConv1 1×1 point-by-point convolution to reduce the number of channels in the input feature map X to 1 / 4 of its original number. PWConv2 then maintains the number of channels consistent with the original input feature map. The output is calculated as Y, and X' is the final output feature map. This improved module effectively enhances the interactive capabilities of embedded positioning information, accurately localizing the target region of interest, and making the model more lightweight.

[0014] Y h ==σ s (G n (F h (z h ))) (5)

[0015] Y w ==σ s (G n (F w (z w ))) (6)

[0016] Where: σ s Represents the Sigmoid activation function, G h represents group normalization, F h Represents the feature map after the H dimension Conv1d operation, z h Represents the feature map after H-dimensional pooling operation, F w Represents the feature map after the W dimension Conv1d operation, z w Represents the feature map after the W-dimensional pooling operation.

[0017] Y=σ s {B{PWConv2{σ r [B(PWConv1(X))]}}} (7)

[0018] Where: σ s represents the Sigmoid activation function, PWConv1 and PWConv2 represent point-by-point convolution operations, B represents batch normalization, σ r Represents the ReLU activation function, and X represents the input feature map.

[0019] X'=X×Y h ×Y w ×Y (8)

[0020] Where: Y h Represents the H-dimensional output feature map, Y w represents the W-dimensional output feature map, and Y represents the output feature map of the intermediate branch.

[0021] The GH-Detect detection module consists of two cascaded GH-Block modules and two parallel Conv2d modules. The GH-Block module structure introduces GhostConv to lightweight the detection head and then improves detection performance through two layers of convolution and a skip connection. GhostConv performs simple linear operations on the feature map obtained by conventional convolution to increase the number of channels. This linear operation is highly flexible and can be performed simultaneously with the feature map, preserving the integrity of the original feature map, avoiding feature information loss, and reducing the number of model parameters while maintaining detection accuracy as much as possible.

[0022] The beneficial effects of the present invention are:

[0023] 1. Compared with the traditional backbone networks in the YOLOv3, YOLOv5s, YOLOv7-tiny, and YOLOv8n models, the DSAF module proposed in this paper can focus more on the key features of the target object, enhance the perception of slender cracks, facilitate comprehensive detection of tank defects, improve detection accuracy, and effectively reduce redundant calculations.

[0024] 2. Compared with YOLOv11s, YOLOv10n and YOLOv8n-GAM, the E-C2f module proposed in this paper is designed to integrate the different scale features of tank defects, enhance the neck network feature fusion capability, enable the entire network to focus more accurately on the object detail features, optimize the model's detection performance, and improve the accuracy of model detection.

[0025] 3. Compared with RT-DETR, Faster-RCN, and SSD, which have high parameter counts and high computational complexity, the GH-Detect detection module proposed in this invention and the GH-Detect detection head can reduce the interference of redundant gradient information, enhance the learning ability of the network, reduce the number of parameters and computational complexity, compress the model size, make the model easier to deploy on mobile terminals, and achieve comprehensive detection of tank defects.

[0026] 4. The lightweight detection method for tankway surface defects proposed in this invention only requires placing the image to be detected into the proposed target detection model to quickly and accurately obtain the type and location of tankway surface defects, thereby improving the efficiency of automated inspections in underground coal mines. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0028] Figure 1 It is a structural schematic diagram of the tank surface defect detection model of the present invention.

[0029] Figure 2 It is a schematic structural diagram of the DSAF module of the present invention.

[0030] Figure 3 It is a schematic diagram of the DELA-C2f module structure of the present invention.

[0031] Figure 4 It is a schematic diagram of the structure of the GH-Detect detection module of the present invention.

[0032] Figure 5 This is a comparison curve of the mAP@50 indicators of the DT-YOLOv8 and YOLOv8n algorithms of the present invention.

[0033] Figure 6 This is a comparison chart of the detection results of tank surface defects using the DT-YOLOv8 detection algorithm of the present invention and other detection algorithms.

[0034] Figure 7 This is a physical picture of a low-computing-power embedded platform for a lightweight detection model for tank surface defects according to the present invention.

[0035] Figure 8 This is a real-time detection effect diagram of a lightweight detection model for tank surface defects on a low-computing-power embedded platform according to the present invention. DETAILED DESCRIPTION

[0036] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0037] A lightweight detection model for tank surface defects, the structure of which is as follows Figure 1 As shown, it includes three parts: feature extraction backbone network, neck feature fusion network and prediction output network;

[0038] The backbone network consists of one adaptive fusion module DSAF, five CBSs, three C2f modules, and one SPPF module;

[0039] The neck feature fusion network consists of four dynamic and efficient local attention fusion modules DELA-C2f, two CBS, two UpSample upsampling layers, and two Concat layers;

[0040] The prediction output network consists of three GH-Detect detection heads, each of which extracts image features through two GH-Block modules in cascade and two Conv2d modules in parallel;

[0041] like Figure 2 As shown in the figure, the backbone network DSAF module first uses 1×1 standard convolution for feature extraction, then uses dynamic snake convolution (DSConv) to capture the features of slender cracks, uses residual branches to perform adaptive average pooling operations on the feature image in the H and W dimensions, compresses the H×W×C input feature map into a 1×1×C feature vector, and then performs an excitation operation using a full connection. Then, an upsampling operation is performed to make the feature map the same size as the feature map from the main branch, and a Concat operation is used to fuse the features of two different levels. Then, another standard convolution is used to extract detail features.

[0042] like Figure 3 As shown, the neck feature fusion DELA-C2f module improves the Bottleneck module by fusing 3×3 standard convolution and improved efficient local attention (ELA) to obtain the DE-Block module structure. First, 1×1 standard convolution is used for feature extraction, and then 3×3 standard convolution and D-ELA are used in parallel to enhance local feature extraction. Then, the Concat fusion operation is performed, and finally, more detailed information is extracted by 1×1 standard convolution and then added to the input feature map to obtain the output, which effectively reduces feature redundancy and can also protect multi-scale contextual information. The improved efficient local attention mechanism D-ELA module structure introduces an intermediate branch based on ELA, which includes point-by-point convolution, normalization and activation function. This module first performs average pooling on the height and width dimensions of the input feature map X, and then uses Conv1d for faster and lighter local processing. The feature map is then refined through group normalization and nonlinear activation function, and the final output feature maps are generated from two directions, namely Y h and Y w The intermediate branch uses PWConv1 1×1 point-by-point convolution to reduce the number of channels in the input feature map X to 1 / 4 of its original number. PWConv2 then maintains the number of channels consistent with the original input feature map. The output is calculated as Y, and X' is the final output feature map, calculated as shown in the following formula. The improved module effectively enhances the interactive capabilities of embedded positioning information, accurately locates the target area of ​​interest, and makes the model more lightweight.

[0043] like Figure 4As shown in the figure, the GH-Detect detection module in the prediction output network consists of two cascaded GH-Block modules and two parallel Conv2d modules. The GH-Block module structure introduces GhostConv to lightweight the detection head and then improves detection performance through two layers of convolution and a skip connection. GhostConv increases the number of channels by performing simple linear operations on the feature map obtained by conventional convolution. This linear operation is highly flexible and can be performed simultaneously with the feature map, preserving the integrity of the original feature map, avoiding feature information loss, and reducing the number of model parameters while maintaining detection accuracy as much as possible.

[0044] A method for detecting a lightweight detection model for tank channel surface defects, characterized in that the tank channel surface defect detection is performed according to the following steps:

[0045] S1. Preprocess the tank channel surface defect image to convert the image size to 640×640, send the processed tank channel defect image to the backbone network Backbone, complete the tank channel surface feature extraction, and send the shallow and deep features to the neck feature fusion Neck in three ways;

[0046] S2, the neck feature fusion module Neck outputs the fused features after passing through four DELA-C2 modules and sends them to the prediction output network;

[0047] S3. The prediction output network completes the surface detection of tank defects and outputs the detection results of the tank surface defects.

[0048] To verify the detection effectiveness of the DT-YOLOv8 proposed in this paper, a comparative experiment was conducted. Relying on the State Key Laboratory of Mining Response and Disaster Prevention in Deep Coal Mines, this paper established a mine tunnel simulation platform and constructed a tank tunnel surface defect dataset. An industrial camera was used to collect 900 images of tank tunnel defect data under different light intensities, and the defective parts were annotated using Labelimg annotation software. To avoid the adverse effects of class imbalance on the algorithm, the dataset was expanded through methods such as rotation, mirroring, and cropping, ultimately expanding the dataset to 3,000 images, and dividing the training and validation sets in a ratio of 8:2.

[0049] like Figure 5 As shown in the figure, the mAP@50 indicator comparison curves of the DT-YOLOv8 algorithm and the YOLOv8n algorithm are shown. After 200 rounds of iteration, the mAP@50 curve of the model of the present invention begins to stabilize, which can intuitively show the superiority and effectiveness of the algorithm.

[0050] This invention compares YOLOv3 (see "RED MON J,FARHADI A.YOLOv3:An inc-rementalimprovement[J].arXivpreprint arXiv:1804.02767,2018."), YOLOv5s (see "JocherGlenn.YOLOv5release v6.1.https: / / github.com / ultralytics / yolov5 / releases / tag / v6.1,2022"), YOLOv7-tiny (see "Wang C Y,Bochkovskiy A,Liao H Y M.YOLOv7:Trainable bag-of-freebies sets new state-of-the-art for real-time objectdetectors[J].arXiv preprint arXi v:2207.02696,2022"), YOLOv10n (see "WANG A,CHEN H,Liu L,et al.YOLOv10:Real-Time End-to-End Object Detection[J].2024."), YOLOv11s (see "KHANAM R,HUSSAIN M.YOLOv11:An Overview of the Key ArchitecturalEnhancements[J].2024."), RT-DETR (see "YU C S,YOAN S.MCG-RTDETR:Multi-Convolution and Context-Guided Network with Cascaded Group Attention forObject Detection in Unma nned Aerial Vehicle Imagery[J].Remote Sensing,2024,16(17):3169."), Faster-RCNN (see "REN S,HE K,GIRSHICK R,et al.FasterR-CNN:Towards real-time object detection wi th re-gion proposal networks[J].Advances in Neural Information Processing Systems,2015,28:91-99.”), SSD (for details, see “LIU W, ANGUELOV D, ERHAN D, et al. Ssd: Single shot multi box detector [C]. European Conference on Computer Vision. Springer, Cham, 2016: 21-37.”), and YOLOv8n-GAM (for details, see “XU HW, XIAO XY, ZHAO ZW, et al. YOLOv8n-GAM: an improved surface defect detection network for hot-rolled strip steel [J]. Engineering Research Express, 2024, 6: 025557.”) classic target detection networks. The experimental results are shown in Table 1, where “↑” indicates that the performance is better as the index increases, and “↓” indicates that the performance is better as the index decreases. The data in bold indicates the best performance, and the data with a red wavy line indicates the worst performance.

[0051] Table 1 Comparative experimental results of different detection models

[0052]

[0053]

[0054] As shown in Table 1, the DT-YOLOv8 algorithm in this paper is superior to the current popular detection algorithms. The Faster-RCNN model has a large number of parameters and high computational complexity, and is not suitable for deployment on hardware devices. The SSD model is relatively weak in multi-scale feature fusion and detail feature extraction, and its accuracy and mAP@50 are significantly lower than those of the model in this paper, and its weight is relatively large. RT-DETR is an algorithm developed by Baidu. This algorithm is a real-time end-to-end detector based on the DETR architecture. The model has large parameters and high computational cost. In the single-stage detection model, YOLOv3 and YOLOv5s are used in the backbone network. The network design is relatively simple, and its feature fusion and accuracy do not meet the actual industrial requirements for tankway defect detection. YOLOv7-tiny and YOLOv10n are lightweight models with relatively fast inference speed, but their accuracy may not fully meet the detection requirements for high precision and lightweightness. The YOLOv11s model's detection performance is similar to that of this algorithm, but the number of parameters and floating-point operations is approximately 379% and 371% of that of this model. Compared with YOLOv8n-GAM, DT-YOLOv8's computational complexity and model weight are reduced by 70% and 29.5%, and by 26.1% and 17.9%, respectively. Table 1 shows that DT-YOLOv8 has good detection performance and lightweight capabilities, meeting the requirements of underground coal mine tankway defect detection and subsequent deployment on lightweight equipment.

[0055] In order to more comprehensively demonstrate the performance of the DT-YOLOv8 network model in steel surface defect detection, this paper Figure 6 A detailed comparative analysis was conducted. To more intuitively evaluate the detection performance of the improved algorithm, the tank channel test set was tested using the YOLOv5s, YOLOv7-tiny, YOLOv10n, YOLOv8n, and DT-YOLOv8 models. Group (a) shows that the baseline model, YOLOv8n, misses small cracks, while the improved model performs better than the baseline model in detecting small cracks under poor lighting conditions. Group (b) shows that YOLOv5s, YOLOv7-tiny, YOLOv10n, and YOLOv8n all miss wear and crack defects. The proposed model performs better than the other models in detecting defects on rough tank channel surfaces. Group (c) shows that when multiple defects are present simultaneously, YOLOv7-tiny, YOLOv10n, YOLOv8n, and DT-YOLOv8 can all detect them, while only YOLOv5s misses them. This shows that the DT-YOLOv8 model has good detection performance in tank channel defect identification while being lightweight, further illustrating the effectiveness of the improved model.

[0056] like Figure 7 As shown, the low-computing-power embedded platform selected for the network of the present invention is the Jetson development board produced by NVIDIA, which is powerful and very suitable as a development platform for edge devices, on which various image processing tasks can be deployed.

[0057] like Figure 8 As shown, the network proposed by the present invention can complete normal detection functions on a low computing power platform and still have high inspection accuracy.

[0058] If the image classification method described in the embodiment of the present invention is implemented in the form of a software functional module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the existing technology, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server or network device, etc.) to perform all or part of the steps of the image classification method described in the embodiment of the present invention. The aforementioned storage medium includes various media that can store program codes, such as a USB flash drive, a mobile hard disk, a ROM, a RAM, a magnetic disk or an optical disk.

[0059] The above description is only a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention are included in the scope of protection of the present invention.

Claims

1. A lightweight detection model for tank surface defects, characterized by: It consists of three parts: backbone feature extraction network, neck feature fusion network and prediction output network; The backbone feature extraction network consists of an adaptive fusion module DSAF, 5 CBS, 3 C2f modules and 1 SPPF module; The neck feature fusion network consists of four dynamic and efficient local attention fusion modules DELA-C2f, two CBS, two UpSample upsampling layers and two Concat layers; The prediction output network consists of three GH-Detect detection heads, each of which extracts image features by connecting two GH-Block modules in cascade and two Conv2d modules in parallel.

2. A lightweight detection model for tank surface defects according to claim 1, characterized in that: The DSAF module of the backbone feature extraction network first uses 1×1 standard convolution for preliminary feature extraction, then uses dynamic snake convolution DSConv, and then uses residual branches to perform adaptive average pooling on the feature map in the H and W dimensions, compressing the H×W×C input feature map into a 1×1×C feature map, and processing it through activation operations and fully connected layers. Then, the feature map size is adjusted to the same size as the main branch output feature map through upsampling operations, and finally the features of different levels are fused through the Concat operation, and then the detail features are further extracted through standard convolution.

3. A lightweight detection model for tank surface defects according to claim 1, characterized in that: The DELA-C2f module of the neck feature fusion network improves the Bottleneck module by fusing 3×3 standard convolution and improved efficient local attention (ELA) to obtain the DE-Block module; the DE-Block module first uses 1×1 standard convolution, then undergoes 3×3 standard convolution and D-ELA modules in parallel, and then performs a Concat fusion operation, and finally obtains the output by adding the input feature map through 1×1 standard convolution.

4. A lightweight detection model for tank surface defects according to claim 1, characterized in that: The prediction output network and the GH-Detect detection module are composed of two cascaded GH-Block modules and two parallel Conv2d modules. The GH-Block module structure introduces GhostConv to lightweight the detection head and then improves the detection performance through two layers of convolution and a skip connection.

5. A method for detecting a lightweight detection model for tank surface defects, characterized in that: The tank channel defect data set is randomly divided into a training set and a validation set in a ratio of 8:

2. The tank channel defect detection model as claimed in claims 1 to 4 is trained with the training set, and the trained model is used to implement tank channel detection on the images to be detected.

6. A method for detecting a lightweight detection model for tank surface defects, characterized in that: Follow these steps: S1. Preprocess the tank channel surface defect image to convert the image size to 640×640, send the processed tank channel image to the backbone network to complete the surface feature extraction of the tank channel surface defect, and send the shallow and deep features to the neck feature fusion network in three ways; S2, the core module of the neck feature fusion network DELA-C2f, consists of the D-ELA module. Its purpose is to improve the ability to focus on defect information during the feature extraction process, effectively reduce the influence of irrelevant information, and generate multi-scale feature maps that are sent to the prediction output network; S3. The prediction output network completes the tank surface defect detection and outputs the detection results of the tank surface defects.

7. According to S2 in claim 6, it is characterized in that The working process of the D-ELA module is as follows: an intermediate branch is introduced on the basis of ELA, which includes point-by-point convolution, normalization and activation function. The module first performs average pooling on the height and width dimensions of the input feature map X, then processes it using Conv1d, and then refines the feature map through group normalization and nonlinear activation function, generating the final output feature maps Y from two directions. h and Y w The middle branch uses the point-by-point convolution operation of PWConv1 to reduce the number of channels of the input feature map X to 1 / 4 of the original number, and then uses PWConv2 to keep the number of channels consistent with the original input feature map. The output calculation formula is Y, and X' is the final output feature map; The feature extraction process of the D-ELA module is as follows: Y h ==σ s (G n (F h (z h ))) (1) Y w ==σ s (G n (F w (z w ))) (2) Where: σ s Represents the Sigmoid activation function, G h represents group normalization, F h Represents the feature map after the H dimension Conv1d operation, z h Represents the feature map after H-dimensional pooling operation, F w Represents the feature map after the W dimension Conv1d operation, z w Represents the feature map after the W-dimensional pooling operation; Y=σ s {B{PWConv2{s r [B(PWConv1(X))]}}} (3) Where: σ s represents the Sigmoid activation function, PWConv1 and PWConv2 represent point-by-point convolution operations, B represents batch normalization, σ r Represents the ReLU activation function, X represents the input feature map; X'=X×Y h ×Y w ×Y (4) Where: Y h Represents the H-dimensional output feature map, Y w represents the W-dimensional output feature map, and Y represents the output feature map of the intermediate branch.

Images