ISBN: 978-981-14-4787-7 DOI: 10.18178/wcse.2020.06.020
An Improved CAM for Weakly Supervised Fabric Defect Detection
Abstract—Deep learning technology has proven applicable in fabric defect detection, but the detection performance depends on the large-scale labelled training sets. Due to the wide variety of fabrics texture and the defects, it is a very arduous task to collect and label a large dataset for each type of fabrics. To alleviate this issue, weakly supervised object detection method was adopted for the fabric defect detection in this paper. We combine the Class Activation Map (CAM) and LZF-Net into the Simplified-LZFNet that achieved good classification accuracy and localization in our fabric datasets. Experimental results demonstrate that the average detection accuracy of the proposed model drops by 2% in the weak supervision setting, but obtain the excellent localization. Thus it improves the practicability of the defect detection method based on deep learning.
Index Terms—Fabric Defect Detection, Weakly Supervised Object Detection, Class Activation Mapping, LZF-Net.
Zhaochen Huo, Zhoufeng Liu, Chunlei Li
Zhongyuan University of Technology, CHINA
Cite: Zhaochen Huo, Zhoufeng Liu, Chunlei Li, " An Improved CAM for Weakly Supervised Fabric Defect Detection " Proceedings of 2020 the 10th International Workshop on Computer Science and Engineering (WCSE 2020), pp. 117-124, Shanghai, China, 19-21 June, 2020.