MPN: Multi-scale Progressive Restoration Network for Unsupervised Defect Detection

Abstract

Defect detection is one of the most challenging tasks in the industry, as defects (e.g., flaw or crack) in objects usually own arbitrary shapes and different sizes. Especially in practical applications, defect detection usually is an unsupervised issue, since it is difficult to collect enough labeled defect samples in the industrial scenario. Although a lot of works have achieved remarkable progress with the help of labeled data, how to effectively detect defects with only positive samples (i.e., clean image without any defects) for training is still a troublesome problem. Therefore, in our paper, we adopt the image restoration strategy to address the unsupervised defect detection task. More specifically, we enable model to restore the original image from the defect samples first and then calculate the differences between the restored image and defect image as the final results. To deal with the unsupervised scenario, we first introduce a self-synthesis component to generate pseudo defects for training by Poisson editing, and the generated pseudo defects will be used for training. Targeting at the unsupervised defect detection, we introduce a novel Multi-scale Progressive restoration Network (MPN), which utilizes multi-scale information to detect defects progressively. More specifically, we choose an invariant scale convolution network as the restoration network for image reconstruction. However, it is difficult to detect defects with once restoration. Thus, we adopt the iterative restorations and our model is conditioned on the previous results for progressive defect detection. Our progressive detection is maintained by using a recurrent neural network to memory previous states. Considering that the defects can be arbitrary size, we incorporate the top-down and bottom-up structure into our model to extract multi-scale semantics better. Experimental results on multiple datasets demonstrate that our model achieves better performance than previous methods.