Dumper carriage weld defect detection method and device and storage medium

Abstract

The invention relates to the field of defect detection, and discloses a dumper carriage weld defect detection method and device and a storage medium, and the method comprises the steps: collecting dumper carriage weld defect images to construct an original data set; marking the original data set, and dividing the original data set according to a preset proportion; constructing an improved compartment welding seam detection model based on a YOLOv8n network structure; inputting the training set into the improved model and the original YOLOv8n model for training; and different degrees of salt and pepper noise are added in the original data set, and the detection performance of the original model and the improved model under the extreme noise condition is tested. According to the method, the problems of high missing detection rate and false detection rate of weld defect detection in the prior art are solved, so that the weld defect detection precision is improved, the calculation complexity of the YOLOv8n model is reduced, and the YOLOv8n model can be better deployed on industrial edge equipment.

Description

Technical Field

[0001] The present invention relates to the field of defect detection, and in particular to a method, equipment and storage medium for detecting defects in a dump truck compartment weld. Background Art

[0002] With the rapid development of industrial automation and intelligent manufacturing, the reliability and safety requirements of mechanical structures are becoming increasingly higher. Among many mechanical structures, dump truck compartments have attracted much attention due to their wide application in industries such as construction, mining, and logistics. The quality of the welds in dump truck compartments is directly related to their load-bearing capacity and service life. Therefore, the detection of weld defects has become a key link in ensuring the safety of the compartments.

[0003] Welds are an important way to connect metal parts, and their quality directly determines the stability and safety of the structure. However, the quality of welds will be affected by the workers' skills, welding methods, welding environment, etc., resulting in a variety of weld defects, such as holes, biased welding, spattering, and undercuts. These weld defects will reduce the structural strength of the carriage and may even cause safety accidents. Therefore, weld defects must be detected to ensure that the product quality of the carriage meets the standards.

[0004] Traditional weld defect detection methods mainly rely on manual visual inspection, which is inefficient, inaccurate, and greatly affected by the subjective factors of the inspectors. With the development of deep learning technology, more and more researchers have begun to explore the use of deep learning models to automatically detect weld defects. Deep learning models, especially convolutional neural networks, have shown excellent performance in image recognition and classification, providing a new perspective for the automatic detection of weld defects. The YOLO (You Only Look Once) series of algorithms have been widely used in the field of industrial inspection for their fast and accurate target detection capabilities, but in actual applications, it is found that the model network structure is relatively complex, and when the characteristics of the defects are small, there are problems of missed detection and false detection.

[0005] In the existing Chinese patent with publication number CN117470849A, a weld defect detection method based on an improved YOLO algorithm is disclosed. This method adds an attention mechanism to the C3 module of the backbone network to enhance the backbone network's ability to extract weld features. At the same time, the ordinary convolution in the C3 module of the backbone network and the neck network is replaced with Ghostconv to improve the detection efficiency of the model. However, this method only provides feasibility verification of the weld defect detection function, and has no requirements for the detection accuracy and detection speed of the improved model, which cannot meet the requirements of high precision and high efficiency in industrial environments.

[0006] In the existing Chinese patent with publication number CN116468684A, a pipeline weld defect detection method based on an improved YOLO network is disclosed. This method uses two improved residual network structures to extract features in the upsampling feature extraction module, achieving a double increase in network model detection accuracy and detection speed, and designs a regularized cross entropy loss function to improve the sample imbalance problem. However, it is not possible to verify whether the proposed improved method has good generalization ability through other data sets.

[0007] The existing Chinese patent with the announcement number CN113034478B discloses a method and system for identifying and locating weld defects based on a deep learning network. This method improves the backbone network and detection head of the YOLOv4 model to enhance the improved model's ability to locate tiny weld features. However, due to the complexity of the detection environment, the confidence score of the identified tiny weld defects is low and does not have sufficient credibility.

[0008] A lot of noise is generated during the welding process. The presence of this interference information may lead to missed detection and false detection during the welding defect detection process. At the same time, in order to better deploy on industrial edge devices, the computational complexity of the algorithm model must be reduced as much as possible. Therefore, it is necessary to develop a weld defect detection method that takes into account both high precision and lightweight, and this method also needs to have good detection performance in extreme noise environments. Summary of the invention

[0009] The purpose of the present invention is to propose a method, device and storage medium for detecting defects in the weld seam of a dump truck compartment, so as to solve the technical problems that the current weld seam detection in the compartment has low precision, complex calculation and is easily affected by the noise environment, resulting in high missed detection rate and false detection rate of the detection results and long time consumption.

[0010] Specifically, the present invention provides a method for detecting weld defects in a dump truck compartment, comprising the following steps:

[0011] S1, collect the weld defect images of dump truck compartments to build the original data set;

[0012] S2. After annotating the original data set using annotation tools, divide it into a training set, a validation set, and a test set according to a preset ratio;

[0013] S3. Construct an improved car weld detection model based on YOLOv8n network structure;

[0014] S4, input the training set into the improved car weld detection model based on the YOLOv8n network structure and the original YOLOv8n model for training, so that the two models can learn the mapping relationship between features and targets, and then test the detection accuracy of the two models on the test set;

[0015] S5. Add different degrees of salt and pepper noise to the original data set to test the detection performance of the original YOLOv8n model and the improved car weld detection model based on the YOLOv8n network structure under extreme noise conditions to verify the robustness and generalization ability of the improved algorithm.

[0016] A storage medium stores instructions and data for implementing a method for detecting weld defects in a dump truck compartment.

[0017] A dump truck compartment weld defect detection device comprises: a processor and a storage medium; the processor loads and executes instructions and data in the storage medium to implement a dump truck compartment weld defect detection method.

[0018] The beneficial effects provided by the present invention are:

[0019] Compared with the prior art, the present invention replaces the detection head with a lightweight detection head DESLDH with detail enhancement, enhances the improved model's positioning capability for weld defect features, and reduces the computational complexity of the improved model; redesigns the neck network of the YOLOv8n model, reduces the loss of weld defect feature information during the feature fusion process of the improved model, improves the detection accuracy of the improved model, and thus reduces the missed detection rate and false detection rate of the improved model; introduces ADown downsampling to replace part of the conv downsampling in the backbone network and feature extraction network of the YOLOv8n model, enhances the improved model's positioning capability for tiny weld features such as holes and sputtering, and improves the improved model's detection accuracy for tiny features; redesigns part of the C2f module of the YOLOv8n model using multi-scale high-efficiency convolution MSHEConv, and improves the computational efficiency of the improved model through the C2f-MSHEC module; finally, through the channel pruning method, removes the redundant structure in the improved model network, and further reduces the computational complexity of the model. In addition, the present invention tests the improved model by adding noise to the weld defect data set, and verifies that the improved model has good robustness and generalization ability. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 It is a schematic flow chart of the method of the present invention;

[0021] Figure 2 A schematic diagram of defect types of a weld defect dataset provided by an embodiment of the present invention;

[0022] Figure 3 A network structure diagram of the improved YOLOv8n model provided in an embodiment of the present invention;

[0023] Figure 4A module schematic diagram of a lightweight detection head DESLDH with detail enhancement designed by the present invention provided in an embodiment of the present invention;

[0024] Figure 5 A schematic diagram of a module of a feature focusing diffusion network FFDN designed by the present invention provided in an embodiment of the present invention;

[0025] Figure 6 A schematic diagram of the FocusFeature module in the FFDN network designed by the present invention provided in an embodiment of the present invention;

[0026] Figure 7 A schematic diagram of a module for ADown downsampling in YOLOv9 provided in an embodiment of the present invention;

[0027] Figure 8 A structural diagram of the multi-scale efficient convolution MSHEConv designed by the present invention provided in an embodiment of the present invention;

[0028] Fig. 9 A schematic diagram of a module of a C2f-MSHEC designed by the present invention provided in an embodiment of the present invention;

[0029] Fig.10 A comparison diagram of the F1 curves of the improved model of the present invention and the original model provided in the embodiment of the present invention.

[0030] Fig.11 A visual comparison diagram of the improved model of the present invention and the original model on the test set provided by the embodiment of the present invention;

[0031] Fig.12 It is a working schematic diagram of the hardware device of the present invention. DETAILED DESCRIPTION

[0032] To make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0033] Before formally explaining the present invention, relevant technical terms in the present invention are explained. In addition, the scheme of the present invention is also generally explained first for easy understanding.

[0034] Please refer to Figure 1 , Figure 1 It is a schematic diagram of the process of the present invention.

[0035] The present invention provides a method for detecting weld defects in a dump truck compartment, comprising the following steps:

[0036] S1, collect the weld defect images of dump truck compartments to build the original data set;

[0037] As an example, the data set used for the experiment of the present invention was collected in the welding workshop of a certain automobile company. The device used for data collection was a mobile phone. The object of collection was the weld data of the bottom plate of the dump truck in the welding workshop. The pixels of the collected original image were 640 pixels * 640 pixels. A total of 596 original images were collected to construct the original data set. The data set has four defect categories: undercut, partial weld, porosity, and spatter. The four defect categories of the weld data set are shown in the attached figure. Figure 2 shown.

[0038] S2. After annotating the original data set using annotation tools, divide it into a training set, a validation set, and a test set according to a preset ratio;

[0039] As an embodiment, step S2 is specifically as follows:

[0040] First, create a folder named hanfeng, and create two new folders, images and xml, under the folder to store the 596 images of the original dataset and their corresponding label files; then, add the class.txt document to the xml file to mark the four defect categories of the dataset; in addition, use the labelimg annotation tool to annotate the bounding boxes of the 596 dump truck weld defect images to obtain the original dataset with label files; finally, convert the voc format label file to YOLO format through Python code.

[0041] The total number of images in the data set is only 596. A small amount of data sets can easily affect the training accuracy of the model and may also cause overfitting. To avoid this, the present invention uses the following 8 image enhancement methods for the original data set: 1) horizontal flip enhancement; 2) vertical flip enhancement; 3) mirror symmetry enhancement; 4) affine transformation; 5) random rotation; 6) adding Gaussian noise; 7) contrast change; 8) scale transformation. After image enhancement, the number of images in the data set is 5364. Afterwards, the present invention randomly divides the data set according to the ratio of 7:2:1, and obtains a training set of 3755 images, a verification set of 1073 images, and a test set of 536 images.

[0042] S3. Construct an improved car weld detection model based on YOLOv8n network structure;

[0043] It should be noted that the improved carriage weld detection model based on the YOLOv8n network structure described in step S3 is specifically:

[0044] First, the detection head of the original YOLOv8n network structure is replaced with a lightweight detection head DESLDH with detail enhancement;

[0045] Secondly, the neck network of the original YOLOv8n is replaced with a feature focused diffusion network FFDN; at the same time, some Conv downsampling in the backbone network and neck network of the original YOLOv8n network structure is replaced with Adown downsampling;

[0046] Finally, some C2f modules of the original YOLOv8n model are replaced with C2f-MSHEC modules, and redundant structures are removed by channel pruning.

[0047] As an example, please refer to Figure 3 , Figure 3 It is an overall schematic diagram of the improved carriage weld detection model based on the YOLOv8n network structure.

[0048] from Figure 3 It can be seen that the improved model consists of three parts: backbone network, neck network and detection head.

[0049] The backbone network includes an Adown downsampling module and a C2f_MSHEC module.

[0050] In the neck network, feature information of different scales is captured through a set of parallel deep convolutions, and the captured rich multi-scale weld feature information is diffused to each detection scale through convolution and upsampling operations. At the same time, a partial Adown downsampling module is also introduced.

[0051] For the detection head, a lightweight detection head DESLDH with detail enhancement is adopted.

[0052] The following application will focus on the details of the improved model.

[0053] First, the present invention designs a lightweight detection head DESLDH with detail enhancement to replace the detection head of the original model. Figure 4 shown.

[0054] First, after DESLDH receives the feature map input by the neck network, each branch unifies the number of channels of the feature map output by the neck network through a 1x1 convolution with group normalization. Secondly, the feature information of three different scales is fused through shared convolution, while reducing the number of parameters and calculations of the model. Finally, the 3x3 ordinary convolution in the shared convolution is replaced with detail enhancement convolution DEConv, which enhances the model's ability to represent feature detail information and reduces the missed detection rate and false detection rate of the improved model.

[0055] Secondly, the present invention designs a feature focused diffusion network FFDN to replace the neck network of the original model, and its network structure is shown in the attached figure. Figure 5 and Figure 6 As shown, first, the present invention designs a feature focusing module, which utilizes a set of parallel deep convolutions to capture rich feature information of different scales, thereby enhancing the improved model's ability to locate multi-scale weld defect feature information; secondly, this study designs a unique diffusion mechanism, which diffuses the captured rich multi-scale weld feature information to each detection scale through convolution and upsampling operations, thereby reducing the feature information loss in the feature fusion process, thereby improving the detection accuracy of the improved model.

[0056] Furthermore, the present invention replaces part of the Conv downsampling with ADown downsampling, and its module schematic diagram is shown in the attached figure. Figure 7 As shown in the figure, this module realizes feature extraction of weld defect images of different scales through average pooling and maximum pooling operations, and combines spatial features of different scales through splicing operations to retain the main feature information of weld defects, thereby enhancing the improved model's perception and positioning capabilities for some tiny weld defects such as porosity and spatter, and improving the model's detection performance for tiny weld defects.

[0057] The calculation formula for average pooling is:

[0058]

[0059] The calculation formula for the maximum pooling is:

[0060]

[0061] Where: k h ,k w Represents the height and width of the pooling kernel, x i,j Represents the pixel value of the input feature map (i, j).

[0062] Finally, the present invention designs a multi-scale high-efficiency convolution MSHEConv, whose structure diagram is shown in the attached figure. Figure 8 As shown, the first branch of the module runs in a non-working state, which reduces the amount of calculation and parameters of the module and improves the detection efficiency of the model for weld images.

[0063] After that, replace the bottleneck Conv in the C2f module with the multi-scale high-efficiency convolution MSHEConv, construct Block_MSHEC, and replace the original bottleneck in the C2f module, and finally construct a lightweight C2f-MSHEC module, as shown in the attached diagram. Fig. 9 As shown in the figure, this module reduces the computational complexity of the model training process and improves the detection efficiency of the model.

[0064] As an embodiment, the above four improved modules are improvements to the YOLOv8n network structure, and the overall structure diagram of the final improved model is shown in the attached figure. Figure 3 As shown. The present invention improves the YOLOv8n model through a series of lightweight improvement measures. The final improved model maintains high detection performance under the premise of greatly reducing the number of parameters and model volume, but the reduction in its computational complexity is not large, indicating that the network structure of the improved model still has redundancy. To this end, the present invention adopts the channel pruning method in structured pruning to remove the filters and their connected feature maps that have little impact on the output accuracy in the model network structure, thereby compressing the model network structure and improving the computational efficiency of the improved model.

[0065] S4, input the training set into the improved car weld detection model based on the YOLOv8n network structure and the original YOLOv8n model for training, so that the two models can learn the mapping relationship between features and targets, and then test the detection accuracy of the two models on the test set;

[0066] It should be noted that step S4 in the present invention is specifically as follows:

[0067] Feature extraction: The weld defect images in the training set are input into the backbone network of the YOLOv8n model and the improved model. The backbone network will perform feature extraction operations on the features of the input images.

[0068] Feature fusion, the weld defect images of the training set will be input into the feature fusion network of the neck after the feature extraction operation of the backbone network. In this process, the feature fusion network of the neck will fuse the weld defect features of different scales and output the learned features to the detection head.

[0069] Bounding box prediction and non-maximum suppression. The detection head maps the learned feature information to the prediction task and predicts the bounding box of the weld defect based on the extracted features. Since the network model may predict multiple overlapping bounding boxes, non-maximum suppression operations are also required to retain only the bounding box with the highest score.

[0070] Classification and confidence scoring,For each predicted bounding box, the network model classifies the weld defect and gives a,corresponding confidence score.

[0071] Output results: After being processed by the above network model, the final detection results will be output;

[0072] The trained YOLOv8n model and the improved model are tested on the test set to evaluate the detection performance of the model.

[0073] In order to verify the effect of the present invention, the present invention carried out module ablation experiments respectively, and the data set used in the experiment was a self-made weld defect data set.

[0074] Experimental example:

[0075] The experimental environment and training parameters of the present invention are shown in Table 1:

[0076] Table 1 Experimental environment and training parameters

[0077]

[0078]

[0079] In the embodiment of the present invention, all experimental results are mainly measured by three evaluation indicators: precision P (Precision), recall R (Recall), average precision AP (Average Precision) and average precision mAP (Mean Average Precision) of all categories to measure the detection accuracy of the model. The calculation method of each evaluation indicator is as follows:

[0080]

[0081] Where: TP - the number of times the positive samples are correctly identified as positive samples; FP - the number of times the negative samples are incorrectly identified as positive samples; FN - the number of times the positive samples are incorrectly identified as negative samples;

[0082] The present invention also introduces two evaluation indicators, namely, the number of floating-point operations (GFLOPs) and the number of parameters (Params) during model reasoning, to measure the computational complexity of different models.

[0083] The ablation experiment of the present invention is shown in Table 2:

[0084] Table 2 Ablation experiment results

[0085]

[0086]

[0087] Note: √ stands for adding module; D stands for DESLDH module; F stands for FFDN module; A stands for ADown module; M stands for C2f-MSHEC module; P stands for channel pruning; bold fonts indicate optimal values.

[0088] It can be seen from the ablation experiment that after the five improvements of the YOLOv8n model in this paper, the performance indicators of the improved model have reached the optimal values ​​of the six groups of ablation experiments. Compared with the YOLOv8n model, the precision P and recall R of the improved model are increased by 4.4% and 7.8% respectively, and the average precision mAP of the two models are increased by 2.7% and 15.1% respectively. The number of parameters and the amount of calculation of the model are reduced by 46.7% and 28.4% respectively, achieving an effective balance between high precision and lightweight of the model.

[0089] Afterwards, the present invention compares the F1 curves of the YOLOv8n model and the improved model, as shown in the attached figure. Fig.10 As shown in the figure, the F1 value of the improved model is 6% higher than that of the YOLOv8n model, which proves that the comprehensive performance of the improved model proposed in the present invention is better than that of the YOLOv8n model. In addition, the present invention also intuitively displays the detection effects of the YOLOv8n model and the improved model through a visual comparison chart, as shown in the attached figure. Fig.11 As shown in (1) and (3) in Figure 2, the YOLOv8n model misidentifies non-defective areas and misses sputtering defects, while the improved model correctly identifies all defects in the image. Fig.11 From (2) and (4), we can see that both models correctly identified all weld defects, but the confidence score of the improved model is better, which proves that the improved model has better detection performance than the YOLOv8n model, fully verifying the effectiveness of the improvement of the present invention.

[0090] S5. Add different degrees of salt and pepper noise to the original data set to test the detection performance of the original YOLOv8n model and the improved car weld detection model based on the YOLOv8n network structure under extreme noise conditions to verify the robustness and generalization ability of the improved algorithm.

[0091] As an example, different degrees of salt and pepper noise are added to the self-made weld defect dataset images to simulate the noise such as smoke and metal splash in the actual welding process, test the detection performance of the YOLOv8n model and the improved model under extreme noise conditions, and verify the robustness and generalization ability of the improved algorithm. The experimental dataset is a self-made weld defect dataset with salt and pepper noise added.

[0092] Experimental example:

[0093] Robustness Verification Experiment of the Present Invention

[0094] Table 3 Robustness verification on the weld defect dataset with added noise

[0095]

[0096] Note: Bold fonts indicate optimal values

[0097] From the experimental results of the robustness verification experiment, it can be seen that with the lower computational complexity, the model's two average precisions mAP are improved by 5% and 20.8% respectively, proving that the improved model also has good robustness in extreme noise environments.

[0098] Finally, the present invention also conducted a comparative experiment with the mainstream algorithm, as shown in Table 4.

[0099] In order to compare the accuracy of different algorithm models in weld defect detection, the present invention conducts a performance comparison experiment between the improved model and the three mainstream algorithms currently available based on a self-made weld defect dataset.

[0100] As can be seen from Table 4, the improved model proposed in the present invention has achieved the optimal values ​​of 13 mainstream algorithms in terms of precision P, recall rate R, average precision mAP and parameter quantity. Although the computational complexity of the improved model is 1.7G higher than that of the YOLOv5n model, its precision P and recall rate R are respectively improved by 3.7% and 6.2% compared with the YOLOv5n model, and the average precision of the two items is respectively improved by 2% and 23.5%, while its parameter quantity is reduced by 9%. In summary, the improved model proposed in the present invention has better detection performance than the YOLOv8n model in the process of weld defect detection of dump truck models, and its computational complexity is also lower than that of the YOLOv8n model, and has good industrial application potential.

[0101] Table 4 Comparative experiments of mainstream algorithms

[0102]

[0103]

[0104] Note: Bold fonts indicate optimal values

[0105] See also Fig.12 , Fig.12 4 is a schematic diagram of the working of the hardware device of an embodiment of the present invention, wherein the hardware device specifically comprises: a dump truck compartment weld defect detection device 401, a processor 402 and a storage medium 403.

[0106] A dump truck compartment weld defect detection device 401: The dump truck compartment weld defect detection device 401 implements the dump truck compartment weld defect detection method.

[0107] Processor 402: The processor 402 loads and executes the instructions and data in the storage medium 403 to implement the method for detecting weld defects in a dump truck compartment.

[0108] Storage medium 403: The storage medium 403 stores instructions and data; the storage medium 403 is used to implement the method for detecting weld defects in a dump truck compartment.

[0109] The beneficial effects of the present invention are:

[0110] Compared with the prior art, the present invention replaces the detection head with a lightweight detection head DESLDH with detail enhancement, enhances the improved model's positioning capability for weld defect features, and reduces the computational complexity of the improved model; redesigns the neck network of the YOLOv8n model, reduces the loss of weld defect feature information during the feature fusion process of the improved model, improves the detection accuracy of the improved model, and thus reduces the missed detection rate and false detection rate of the improved model; introduces ADown downsampling to replace part of the conv downsampling in the backbone network and feature extraction network of the YOLOv8n model, enhances the improved model's positioning capability for tiny weld features such as holes and sputtering, and improves the improved model's detection accuracy for tiny features; redesigns part of the C2f module of the YOLOv8n model using multi-scale high-efficiency convolution MSHEConv, and improves the computational efficiency of the improved model through the C2f-MSHEC module; finally, through the channel pruning method, removes the redundant structure in the improved model network, and further reduces the computational complexity of the model. In addition, the present invention tests the improved model by adding noise to the weld defect data set, and verifies that the improved model has good robustness and generalization ability.

[0111] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for detecting weld defects in a dump truck compartment, characterized in that: The method comprises the following steps: S1, collect the weld defect images of dump truck compartments to build the original data set; S2. After annotating the original data set using annotation tools, divide it into a training set, a validation set, and a test set according to a preset ratio; S3. Construct an improved car weld detection model based on YOLOv8n network structure; S4, input the training set into the improved car weld detection model based on the YOLOv8n network structure and the original YOLOv8n model for training, so that the two models can learn the mapping relationship between features and targets, and then test the detection accuracy of the two models on the test set; S5. Add different degrees of salt and pepper noise to the original data set to test the detection performance of the original YOLOv8n model and the improved car weld detection model based on the YOLOv8n network structure under extreme noise conditions to verify the robustness and generalization ability of the improved algorithm.

2. A method for detecting weld defects in a dump truck compartment as claimed in claim 1, characterized in that: The original data set in step S1 includes four defect categories: undercut, offset welding, pores and sputtering.

3. A method for detecting weld defects in a dump truck compartment as claimed in claim 1, characterized in that: The improved carriage weld detection model based on the YOLOv8n network structure described in step S3 is specifically: First, the detection head of the original YOLOv8n network structure is replaced with a lightweight detection head DESLDH with detail enhancement; Secondly, the neck network of the original YOLOv8n is replaced with a feature focused diffusion network FFDN; at the same time, some Conv downsampling in the backbone network and neck network of the original YOLOv8n network structure is replaced with Adown downsampling; Finally, some C2f modules of the original YOLOv8n model are replaced with C2f-MSHEC modules, and redundant structures are removed by channel pruning.

4. A method for detecting weld defects in a dump truck compartment as claimed in claim 3, characterized in that: The data processing process of the lightweight detection head DESLDH with detail enhancement is as follows: DESLDH accepts the feature map from the neck network input, and each branch unifies the number of channels of the feature map output by the neck network through a 1x1 convolution with group normalization; Feature fusion of feature information of three different scales is performed through shared convolution; The 3x3 ordinary convolution in the shared convolution is replaced by detail enhancement convolution DEConv to enhance the model's ability to represent feature detail information.

5. A method for detecting weld defects in a dump truck compartment as claimed in claim 3, characterized in that: The feature-focused diffusion network (FFDN) adopts a set of parallel deep convolutions to capture feature information of different scales, and diffuses the captured rich multi-scale weld feature information to each detection scale through convolution and upsampling operations.

6. A method for detecting weld defects in a dump truck compartment as claimed in claim 3, characterized in that: Replace some C2f modules of the original YOLOv8n model with C2f-MSHEC modules, specifically: Construct the MSHEConv module and replace the bootleneck Conv in the C2f module with the MSHEConv module to obtain the C2f-MSHEC module.

7. A storage medium, characterized in that: The storage medium stores instructions and data for implementing a method for detecting weld defects in a dump truck compartment as described in any one of claims 1 to 6.

8. A dump truck compartment weld defect detection device, characterized in that: include: A processor and a storage medium; the processor loads and executes instructions and data in the storage medium to implement a method for detecting weld defects in a dump truck compartment as described in any one of claims 1 to 6.

Images