| Solid wood panels made from raw wood are a strong,durable,naturally textured,beautiful,comfortable,healthy,and environmentally friendly material that exudes a natural wood aroma.They are a high-quality material for decorating houses,making high-grade furniture and crafts,and are widely used in construction and furniture decoration.However,due to the presence of defects such as live knots,dead knots,and cracks in the boards,their appearance and physical properties are severely affected.Wood processing companies need to remove surface defects from the boards.Therefore,combining deep learning with board defect detection to analyze and identify different types of board defects is essential to improve the yield of wood and ensure the production quality of wood.Online data augmentation and offline data generation are adopted to extend the number of samples and increase the generalization ability and robustness of the model for the problem of insufficient samples of wood defects and poor network robustness.Mosaic data augmentation and Mix Up data augmentation,DCGAN networks,are used to generate new images to solve the problem of unbalanced wood defect categories.In view of the problems of low accuracy,slow detection rate and large model parameters,the feature fusion module of YOLOX target detection algorithm is improved and ECA is added ECA(Efficient Channel Attention)mechanism,ASSF(Adaptively Spatial Feature Fusion)mechanism,improved confidence loss function is Focal Loss.The positioning loss function is EIoU,which improves the feature extraction ability of the algorithm and the accuracy of detection.The experimental results show that the improved model detects the four types of oak defects with great improvement.Improved EAE-YOLOX,96.68% mAP,46.7fps.There are significant advantages over other classical target detection algorithms.Considering the depth and width of the model,a depth-separable convolution and an optional multi-version algorithm were used to reduce the model parameters and computational effort to find the optimal model,EAE-YOLOX-tiny,with a mAP of 94.92% and a speed of 52.09 fps.The effect of different input size models on the algorithm model was tested to cross-validate the effect of both oak and pine datasets on the accuracy of the algorithm.The software interface of the panel defect detection system was designed.The requirements were first analyzed,the general framework of the software and the detection process were designed,then the software interface was written using Qt,and finally the operation process of the real panel defect detection software was demonstrated to show satisfactory results. |
