Defect detection of photovoltaic glass based on level set map

• Published in: Neural computing & applications 2022-07, Vol.34 (13), p.10691-10705
• Main Authors: Dong, Shuai, Chen, Chen, Liang, Yihui, Zou, Kun, Liu, Guisong
• Format: Article
• Language: English
• Subjects: Artificial Intelligence; Computational Biology/Bioinformatics; Computational Science and Engineering; Computer Science; Contours; Data Mining and Knowledge Discovery; Datasets; Feature extraction; Generative adversarial networks; Image annotation; Image Processing and Computer Vision; Optical properties; Original Article; Performance enhancement; Photovoltaic cells; Probability and Statistics in Computer Science; Semantic segmentation
• ISSN: 0941-0643; 1433-3058
• DOI: 10.1007/s00521-022-07005-x

Defect detection of photovoltaic (PV) glass products is a challenge due to its complex optical properties and lack of defective samples. Aiming at this problem, a multi-task framework, in which an auxiliary semantic segmentation task based on the generative adversarial network assists the main defective classification task, is proposed. The auxiliary task guides the feature extractor to focus on the contour of the glass. Level set map (LSM), a new representation of contour which is the integration of a mask, is introduced into the framework to further improve the performance. Unlike level set loss, which is an indirect contour constraint on the segmentation loss in the form of a regular term, LSM is a direct annotation parallel with mask. Recovering LSM and mask simultaneously can affect each other positively. Two synthetic PV glass datasets, named SynSmall and SynBig, and two real-world PV glass datasets, named Mask3 and Defect3, are established to validate the proposed method. Two groups of experiments on the glass datasets are designed to inspect the feasibility and performance of the proposed framework from different aspects. Furthermore, a group of experiments involving separate segmentation tasks on SynSmall and Mask3, together with two public datasets TOMNet and Human, are executed to illustrate the performance of LSM. Experimental results show that the proposed framework can improve the accuracy of PV glass defective detection task significantly and LSM can improve the segmentation accuracy by filtering isolated wrongly recovered regions.