Research And Application Of Vehicle Scratch Damage Determination Method Based On Faster RCNN

In the field of vehicle insurance claims,the traditional loss assessment method has some disadvantages such as complicated loss assessment process and subjective factors.Along with the development of the target detection technology,in the small claims of vehicle scratches scenarios,using the target detection technology greatly simplify the procedure of fee,also has improved detection accuracy,but for scratch detection model is bigger,deployed in poor real-time performance of the model test on a mobile device,Therefore,on the premise of losing as little detection accuracy as possible and aiming at improving the realtime performance of model detection,this paper proposes a vehicle scratch detection method based on Faster RCNN.The main work is as follows:The feature extraction network in the Faster RCNN algorithm was improved,and the convolutional layer structure of the feature extraction network was optimized based on the convolutional layer structure design idea of Mobile Net V3 network,which reduced the calculation of model parameters and increased the real-time performance of the detection model.Aiming at the problem of loss of detection accuracy caused by shrinking model size,m-FPN feature extraction network is proposed based on feature pyramid,which improves sensitivity to image details and makes up for the loss of detection accuracy of model to a certain extent.In this paper,a mapping method of target candidate frame on feature graph is proposed,which maps target candidate frame generated by regional suggestion network to corresponding feature graph,effectively reducing the loss of feature information in target candidate frame.The bilinear interpolation method is used in the pooling layer of the region of interest to pool the target candidate box,which overcomes the limitation of the pooling method in the algorithm and improves the accuracy of the detection model for detecting target location.Two indexes,value and model size,were used to compare the model,and the results showed that the model size was reduced to one third of the original under the condition of less precision loss.Based on the above improvement and innovation work,through the demand analysis of the vehicle scratch recognition system,the design and implementation of the vehicle scratch detection model as the core of the vehicle scratch recognition system,and has shown good detection performance on mobile devices represented by smart phones.