Improved Lightweight GFocal Method for Tile Detection

In the field of industrial production, tile defect detection is one of the key aspects of quality control. Aiming at the problems of high computational complexity and poor detection ability of small targets in industrial tile defect detection, a lightweight tile defect detection method based on GFocal is proposed. The method uses FasterNet as the backbone, which improves the coarse feature extraction ability of the model and increases the detection speed. Secondly, DFC-Involution is introduced in the multi-feature fusion part, which reduces the model parameters and improves the feature fusion ability of the model. Quality Focal Loss and Distribution Focal Loss are utilized to solve the inconsistency and inflexibility of classification and quality prediction. Finally, the BCKD module of knowledge distillation is utilized to replace the backbone with a smaller mobilenetv2, which significantly reduces the model parameters. Experimental results show that the model achieves an average accuracy (mAP) of 77.51 in tile defect detection and the parameters are reduced by 73% compared with GFocal. Compared with the industrial mainstream algorithms, the improved algorithm realizes a balance between parameters and accuracy.