A Method for Segmentation of Surface Defects in Non-flat Area Based on Deep Learning

Abstract

Surface defect detection based on computer vision has become a valuable and promising research field, which has a very high direct impact on the application field of visual inspection, especially industrial production. With the development of computer vision technology, deep-learning has become the most suitable method to solve the problem. By using the sample image as a reference, deep learning enables the inspection system to learn to detect surface defects. However, in the actual industrial environment, fewer defective samples cause difficulties in data collection. This paper proposes a segmentation method, which is specially designed for the segmentation of surface defects in non-flat areas, and demonstrates the detection effect of this method. The method only uses a small number of defective samples for training, which is a very significant advantage for industrial applications. The paper compares the proposed method with related deep-learning detection methods, and the result shows that this method is superior to other methods in detecting surface defects in specific areas on the surface of non-flat areas. We created a new dataset for experimentation based on real cases. The results of experiments show off that the proposed approach apply small number of defected surfaces to fit network parameters, using only approximately 40-50 training samples for each type of defects, and the detection effect is not inferior to the related methods. Because of this few-shot characteristics, the proposed approach has high industrial application value.