Image Classification of Car Paint Defect Detection Based on Convolutional Neural Networks

Abstract In the study of using images to display car paint defects, current need is use deep Convolutional Neural Networks (CNN) identify and classify different types so as give full play application image processing in field automatic defect detection. Using collected images, defects dataset established. The preprocessing process original data three classification models based on CNN are visually displayed. First, 7 body including bubble, dust, fouling, pinhole, sagging, scratch, shrink has been established, with a total 2468 images. model MobileNet-V2, Vgg16, ResNet34 selected for training. As result, after 30 training iterations, MobileNetV2 algorithm achieved 94.3% accuracy, accuracy Vgg16 high 99.9%, maintained at 99.2%. To sum up, detection, learning great potential deserves further development.