Tile Defect Detection Based on the Improved S2C-YOLOv5

• Published in: Artificial Intelligence Vol. 13604; pp. 429 - 440
• Main Authors: Song, Limei, Xiao, Jiatong, Wu, Zhichao, Liu, Mengya, Xiang, Zhonghao
• Format: Book Chapter
• Language: English
• Published: Switzerland Springer 2022; Springer Nature Switzerland
• Series: Lecture Notes in Computer Science
• Subjects: Deep learning; Recognition; Surface defect detection; Tile
• ISBN: 9783031204968; 3031204964
• ISSN: 0302-9743; 1611-3349
• DOI: 10.1007/978-3-031-20497-5_35

For the problem of complex tile surface pattern background, different defect features, low efficiency, and high leakage rate of manual and traditional recognition, this paper proposes a deep learning-based tile surface defect detection method. First, the YOLOv5 backbone network is replaced with a lightweight network ShuffleNetV2, and then, the convolutional block attention module (CBAM) is added. Finally, a lightweight tile detection system is constructed by using CrossEntropyLoss instead of the loss function in the original network. The comparison experiments with six networks such as YOLOv4, Faster RCNN, and SSD show that the algorithm detects tile defects with 95.10% Precision and 92.91% mAP, which is 3.02 percentage points better than the defect recognition accuracy of the network before improvement. It solves the industry-wide problem of long-term reliance on manual visual recognition for tile surface defect detection, realizes automatic and high-precision detection of tile defects, and the resulting tile defect detection technology can be put into stable and reliable operation in the ceramic industry.