An Incremental Surface Defect Detection Method by Fused Unsupervised and Supervised Methods

Abstract

Surface defect detection is an essential procedure during industrial production. It is a challenge to establish an effective model for the surface defects inspection of products. Because defect samples are few and varied. Current supervised learning methods for object detection require large amounts of defect data, which is difficult to collect in the industrial scene. The unsupervised method based on image reconstruction often reconstructs defects. In this paper, we propose a novel surface defect detection method by fused supervised and unsupervised approaches to accurately inspect various surface defects. For unsupervised module, it employs a convolutional autoencoder (CAE) to reconstruct the defect-free image. For the supervised module, use CAE to inspect the defective area for the defective images. A novel loss function is proposed to detect defects by making the residual image between the output image of CAE and the artificial defect im to close to the defect label image. So, by adding a semantic label with all zero values to the defect-free image, the residual image of different tasks is jointly close to their respective semantic labels. Therefore, a unified loss function is used to unify the unsupervised and supervised methods. The experimental results show that the proposed method achieves better inspection accuracy.