Electric Vehicle Scratch Recognition And Detection Based On Deep Learning

The new energy vehicle represented by electric vehicles is the future of the automobile industry.The way of time-sharing is the best choice for electric vehicle promotion.The timely detection of scratches on electric vehicles can effectively check out accidents or faults in the user's driving behavior,and can avoid some users from evading responsibility and concealing accidents after an accident.At the same time,it can help to realize automatic diagnosis of vehicle health,and ensure the safety of users and company operations.This topic studied the automatic identification and detection of electric vehicle scratches based on deep learning.The main research contents are as follows:(1)The car scratch recognition based on convolutional neural network was studied to automatically determine whether there are scratches on the surface of the vehicle.A deep learning model for identifying scratches on electric vehicles is designed.We used the one-dimensional convolutional layer instead of the fully connected layer,which can greatly reduce the number of parameters without reducing accuracy.Dropout was introduced to suppress overfitting.The subject collected data through the vehicle surround system and expanded image data.Compared with other models,our model had high recognition accuracy and occupied less space.(2)After analyzing the traditional object detection method and comparing several methods based on deep learning,the high-precision end-to-end model SSD was selected as the basic architecture of object detection.A network model suitable for vehicle scratch detection was designed.The feature extraction layer is streamlined lightweight mobile network MobileNetV2.Location and classification layers were cropped.And the proportion of default boxes was changed to suit the characteristics of the scratch.The space occupation of the model is greatly reduced without sacrificing accuracy.(3)A cascaded model based on detection and classification was proposed.The object detection network was used as a preliminary screening method for the region,and it could obtain a series of suspected regions.Then the classification network was used for more accurate judgment.The spatial pyramid pooling was introduced to solve the problem of different regional scales.Experiments were carried out to verify validity of the model.Experiments showed that the cascade model can effectively improve the accuracy of scratch recognition.