Automatic zipper tape defect detection using two-stage multi-scale convolutional networks

Abstract

Defects inevitably occur during the manufacturing process of the zipper, significantly affecting its value. Zipper inspection is of significant importance in ensuring the quality of the zipper products. Traditional zipper inspection requires skilled inspectors and is labor-intensive, inefficient, and inaccurate. Currently, automated zipper defects inspection with high precision and high efficiency is still very challenging. In this paper, we propose a novel zipper tape defect detection framework based on fully convolutional networks in a two-stage coarse-to-fine cascade manner. For our special application, the zipper tape defects have multi-scale characteristics. Most of the existing deep learning methods have great advantages in detecting the large-scale defects with prominent features, but are prone to fail in detecting the small-scale ones due to their less remarkable features as well as their general location in a large background area. Thus, we propose to detect first the large local context regions containing the small-scale defects using a multi-scale detection architecture with high efficiency, which integrates a new detection branch by fusing the features in the shallow layer into the high-level layer to boost the detection performance of the context regions. Then we finely detect the small-scale defects from the local context regions detected in the first stage, which can be regarded as large-scale objects that are more easily detected. Extensive comparative experiments demonstrate that the proposed method offers a high detection accuracy while still having high detection efficiency compared with the state-of-the-art methods, coupled with good robustness in some complex cases.