Research And Implementation Of Vehicle Part Recognition Based On Deep Learning Algorithm

With the breakthrough of computer vision processing technology,target detection technology is widely used in all walks of life.Vehicle part recognition utilizes the target detection technology,which can realize the detection of vehicle targets in the image,and at the same time identify the vehicle parts in the Fig.according to the part segmentation standard of the vehicle.In the auto insurance industry,the combination of vehicle part recognition and damage recognition technology can identify and detect each part of the vehicle and the damage to that part,and quickly give repair solutions and repair costs,replacing manual measurement and calculation,saving labor and time costs.In the car-sharing industry,the evaluation of the situation of each part of the car after the user has used it can also be done with the help of vehicle part recognition technology,which identifies and detects the parts and conditions of each vehicle in the pictures taken by the user,making the process that originally needed to be done manually greatly compressed.For some practitioners or researchers related to vehicle computer vision and target detection,vehicle part recognition technology is a very efficient tool that can be used to automatically label data sets,saving the time and money costs associated with manual labeling.In this paper,for the existing deep learning target detection technology,according to the analysis of application scenario requirements,the target detection algorithm Faster RCNN is optimized in terms of both accuracy and scalability,and the structure of the network is optimized by using k-means clustering algorithm to obtain the optimal anchor frame according to the dataset annotation information,replacing the feature extraction network VGG16 to solve the deep network gradient dispersion problem.The multi-scale features are extracted and fused,combining specific features and high semantic features to improve the detection accuracy.A soft non-maximum suppression algorithm is used to filter low confidence prediction frames while preventing objects with too much overlap from being missed,ultimately improving detection accuracy.Meanwhile,in order to train the landed model,thousands of clear vehicle part datasets are obtained by means of crawlers and human photography,and other means,and processing such as screening and labeling of the datasets is performed.At the same time,according to the characteristics of each part of the vehicle,different methods are adopted to improve the accuracy of local detection,using binarized open-operator filtering to process and segment the rear lights,and introducing Mirror-NET to detect windows and windshields that have reflective physical properties,etc.Experiments show that the training model achieves an average accuracy of 86.3% for 24 vehicle parts on the dataset,which is 8.1% better than the original algorithm.Finally,the software system was built using the algorithm,combined with Ali cloud server,We Chat applet and Web.