Cut-Stitch: A Simple and Effective Data Augmentation Method for Industrial Inspection

Haigen Hu, Jingshan Hong, Kangkang Song, Jinwei Zhu, Weilun Ren

Abstract: Deep convolutional neural networks have important application value in industrial defect detection. However, due to the low defect output rate, there are limited available training data, making it very difficult for the model to achieve good generalization performance. Data augmentation is an effective way to solve data deficiency. However, most existing methods usually aimed at general scene objects rather than industrial defects, and the augmented data deviates significantly from the original data distribution, thereby leading to the performance degradation of the modes. To address this issue, we propose a universal data augmentation scheme, called Cut-Stitch, for industrial defect detection. Specifically, some pixel blocks are firstly cut from an original image with a certain pixel width along the height width and width direction. Then the obtained blocks are stitched to generated new images by skip (i.e., every other block) concatenation (block splicing) according to the original order. The images obtained by the proposed Cut-Stitch can maintain a consistent position distribution with the original image, and can improve data diversity. Finally, extensive experiments are conducted to verify the effectiveness of the proposed Cut-Stitch, and the results show that the proposed method can effectively improve the performance of the baseline model and enhance the generalization ability of the model whether in detection or classification tasks. The proposed Cut-Stitch is very promising in the area of industrial defect recognition. The code is publicly available at: https://github.com/HuHaigen/Cut-Stitch.