Carpet Defect Detection by Transfer Learning Combing Classification and Semantic Segmentation

Abstract

Nowadays, with the development of industrial production technology, defect detection has become an indispensable part of industrial production. However, due to various types of products and defects, it can be extremely difficult to identify and locate those defects precisely and accurately. The current major trend in defect detection is using convolutional neural networks and semantic segmentation techniques to better minimize the error rate of human eye recognition and highly improve efficiency. Our work is based on semantic segmentation method and combines it with transfer learning technique enabling our model to train on a relatively small dataset without compromising the performance, and use CNN to firstly classify input images in order to further reduce the number of images to improve computational efficiency and accuracy. Then through incorporating state-of-the-art semantic segmentation model U-Net++, our model achieves the best performance compared to UNet under transfer learning scenario. We compare our model with the state-of-the-art U-Net. Then we use mIOU and pixel accuracy to measure the models’ performance under two scenarios. Results illustrated that through transfer learning scenario, our model achieves the highest scores over other methods.