Research And Implementation Of Vehicle Scratch Recognition System

With the dramatic increase in car ownership,traffic problems occur frequently,and the resulting scratch damage is the most common.Currently,the scratch images in car insurance policies need to be checked by manual inspection to verify the type and damage degree,which has low efficiency,high leakage rate and high work intensity.This paper designs a car scratch recognition system based on YOLOv5 to deal the problems of low accuracy,slow speed and high labor cost of scratch verification in car insurance policies.Firstly,a dataset about car scratches is established.The web images are crawled by a crawler program,and the images from insurance companies are merged and filtered to build a dataset of 800 vehicle scratches.Secondly,an Open CV-based algorithm is designed to achieve darkening and blurring of highlight areas.The YOLOv5 algorithm is improved,the attention mechanism is added to the algorithm,and the original non-maximum suppression algorithm is replaced by weighted boxes fusion.Based on this improved algorithm,a model applicable to car scratch recognition is constructed and deployed in the system,and the software on the PC realizes picture recognition,record display and image display function,and provides picture recognition interface and record function on the web side.Finally,validation and tests are performed on the dataset.The test results show that the average accuracy of the algorithm reaches 92.6%,which is 4.4% higher than the original,and the recognition frame rate reaches 37 FPS.Compared with the manual recognition,the algorithm ensures its recognition accuracy and the recognition speed is significantly improved.The algorithm can effectively improve the recognition accuracy and speed to meet the practical requirements of the system in realistic conditions.