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