Semantic Segmentation of Tile Surface Defects Based on CCH-DeepLabV3+

Magnet tile surface defect images have the problems of complex texture features, defects are highly similar to the background, and tiny defects are often obscured by the complex background, which leads to the poor effect of existing magnet tile defect segmentation algorithms in dealing with defect boundary details. To address the above problems, this paper proposes a semantic segmentation algorithm CCH-DeepLabV3+ based on DeepLabV3+ model for surface defects of magnetic tiles. In order to reduce the information loss caused by downsampling, this paper introduces the Haar wavelet downsampling method, which retains more spatial detail information; the model is trained using the enhanced contrast learning based supervision strategy, which improves the segmentation ability of the model; in the model prediction stage, the CA coordinate attention mechanism is introduced to enhance the model’s ability to recognize and process the key features of defects. In this paper, a series of experiments are conducted based on the Magnetic-Tile-Defect dataset, and the experimental results show that compared with the original DeepLabV3+ model, the improved model improves the mIoU by $2.74 \%$ and the mPA by $3.87 \%$ without basically improving the number of parameters, which also has certain advantages compared with the other deep-learning-based semantic segmentation algorithms.