Defect detection for melamine-impregnated paper decorative particleboard surface based on deep learning

In the context of advancing automation in panel furniture manufacturing, the need for stringent quality control of panel parts has grown. This study delved into utilizing deep learning for detecting surface defects in melamine-impregnated paper decorative particleboard, focusing on edge breakage, scratch, and surface damage. Data augmentation and equalization techniques were devised to enhance defect identification accuracy. Strategies to improve the YOLOv8 algorithm were developed based on various improvements such as DCN and DSC convolutional structures, EMA attention mechanism, BiFPN feature fusion, and Loss function improvement. Then, the optimal effect model was found through ablation experiments and comparative analysis. The precision, recall, and mAP@50 for the identified defects improved from 77.4%, 69.6%, 71.8% to 83.6%, 77.7%, 78.3%. Compared to Faster R-CNN and CB-Net, which perform well in other comparative models, our model's precision is 9.4% and 5.1% higher respectively. This research effectively addressed the limitations of traditional algorithms, such as limited data and challenges in detecting minor target defects, offering valuable insights for advancing automated visual inspection systems in panel furniture production.