Detecting textile micro-defects: A novel and efficient method based on visual gain mechanism

In modern textile industrial processes, fast and efficient detection of textile defects plays a crucial role in textile quality control. Recently, as a critical machine-learning method, faster region-based convolutional neural network (Faster RCNN) have arisen as a promising framework, providing competitive performance for object detection. However, detecting small-scale objects, such as micro-defects on textile, is still a challenging task for Faster RCNN. To address the challenge, this paper aims to develop a new detection model to improve the ability of detecting small-scale objects. First, by analyzing the relationship between the attention mechanism and the visual gain mechanism, we find that the attention-related visual gain mechanism can modify response amplitude without changing selectivity and improve the acuity of visual perception. Then, the relevant mechanisms are further incorporated into the Faster RCNN model to build a new model called Faster VG-RCNN. To evaluate the proposed detection model, a unique textile micro-defect database is built as the benchmark for micro-defect detection. Furthermore, we conduct extensive experimental validations for various design choices. The experimental results show that the proposed Faster VG-RCNN outperforms the existing detection methods. In particular, compared to Faster RCNN, Faster VG-RCNN improves the detection precision from 90.1% to 94.3%. (C) 2020 Elsevier Inc. All rights reserved.