Tire Defect Depth Learning Detection Technology Under The Condition Of Small Sample Set

As one of the most important parts of motor vehicles,the quality of tires directly affects the driving safety of motor vehicles.Tire defect detection is an important part of tire production,and it is a subject of great practical value to realize intelligent defect detection of tire X-ray image.However,in the actual industrial production environment,products with defects only account for a small part,and it is difficult to provide sufficient defect samples for target detection model training.In view of the current situation and shortcomings of quality inspection in tire production,this paper studies the tire X-ray flaw detection technology under the condition of small sample set,combined with the mainstream target detection model,to provide a feasible intelligent detection scheme for tire defects in small sample set.First,explore appropriate data enhancement methods to effectively expand the disease flaw data of a small sample set.Then YOLO v5s object detection model is applied to tire defect detection,which verifies the feasibility of intelligent tire defect detection technology.At the same time,in view of the characteristics of tire defects and the shortcomings of YOLO v5s model structure,an improved method is proposed in the feature extraction and feature fusion stage of the model,which improves the performance of the defect detection model.The specific research contents are as follows:(1)Aiming at the problem of insufficient training samples,explore various data enhancement methods for effective data amplification.Firstly,several main algorithms for data enhancement are analyzed,and the adversary generation network and Mixup algorithm are selected for data amplification.After comparative analysis,appropriate methods are selected for effective data amplification.(2)In view of the problem that the added feature channel contains a large amount of invalid channel information during the sampling process of YOLO v5s backbone network,which will lead to serious loss of tire defect features,and considering the characteristics of the neck network that needs to integrate the characteristics of the specific layer of the backbone network,this paper proposes to integrate CA attention mechanism at the fourth and sixth layers of the backbone network,and use the characteristics of CA attention mechanism,It makes the network feature focus on the effective flaw feature information in the extraction phase,improves the feature extraction ability of the model backbone network,and ensures that the shallow features of the neck network fusion contain significantly less invalid channel information,and the fused features are better.The experimental results show that the mAP of YOLO v5s CA can reach 87.6%,which is 4.5 percentage points higher than YOLO v5s model.(3)In the feature fusion stage of the model,considering different deep and shallow features,their weights should not be exactly the same.The weighted feature fusion mechanism in the weighted bidirectional feature pyramid is used to improve the feature fusion structure of the model.Through network learning,the corresponding fusion weights are assigned to deep and shallow features,effectively improving the ability of model feature fusion.The method is applied to the improved YOLO v5s model in(2),and finally the mAP of the model is further improved to 87.9%.