Using Deep Attention Networks to Extract Defects in Crisscross Background

Nowadays, automatic optical inspection (AOI) has been widely used in advanced manufactory. In AOI area, crisscross background may influence extraction of defect features. A package, semi-finished product of textile industry, usually has cricross background. This study aims to classify four types of package defects, which are wound-in waste, spillover, cobwebs, and dirt. We use a well-known supervised attention-neural-network architecture to classify the four types of package defects effectively. In this study, we use three steps to decide the best strategies. First, we find the best location of channel attention blocks for the deep attention network. After that, we compare two image preprocessing methods to enhance the features of defect. To understand if regularize the background trend will improve the performance or not, we create two kinds of dataset, rotated and non-rotated. Our study improves traditional AOI methods. The experimental results show that the proposed procedures can extract the package defects with interlacing background efficiently.