Dense detection algorithm for ceramic tile defects based on improved YOLOv8

As a common building decoration material, ceramic tiles have been widely used in modern society, and deep learning inspection methods are increasingly employed for tile quality inspection. However, current methods face issues such as slow detection velocity and diminished precision in ceramic tiles detection. To resolve these issues, this study presents a dense detection algorithm for ceramic tile defects with an improved YOLOv8. The model redesigns the CSPLayer (Cross Stage Partial Layer) structure by incorporating the BiFormer architecture, and the SCConv (Spatial and Channel Reconstruction Convolution) is employed to replace the ordinary convolution in the Neck and Head. Furthermore, the MPDIoU + DFL (Distribution Focal Loss) is adopted as the bounding box regression loss function, and the EMA (Efficient Multi-Scale Attention mechanism) attention module is introduced to improve the significance and precision of the defective feature information detection. Experimental results indicate that the final improved model has a size of 58.6 MB, the mAP@0.5 reaches 95.62%, and the FPS is 145.4.