Transformer-based Encoder-Decoder Model for Surface Defect Detection

Recently, deep learning approaches have been gaining popularity in industrial quality control (e.g. surface defect detection), due to their ability for automatically extracting more representative features. In this paper, we propose a two-stage end-to-end approach through a Transformer-based encoder-decoder for surface defect detection. First, we develop a surface defect detection model to train the slicing of input raw images with the same final resolution of the input images and the output images, which better expands the perceptual field. After that, a 1x1 convolution layer is applied to its final layer, thus reducing the number of channels to obtain a single-channel output mask. Then, we combine this single-channel output mask with the output obtained from the last layer of the first stage as the input of the second stage decision layer. Considering different types of sample data, we design two different decision network strategies, namely: plain-up sampling and dynamic-up sampling. Our experimental studies on several publicly available datasets show that the proposed approach is general and effective in detecting defects, and we only need a relatively small number of samples to train the model, which has a good applicability in industrial practice where the sample size is normally limited.