DEEP LEARNING BASED MUSHROOM CLASSIFICATION SYSTEM

The classification and processing of shiitake mushrooms often rely on labor-intensive tasks, requiring significant human effort to select high-quality mushrooms over extended periods. In this paper, we propose a high-efficiency channel pruning mechanism to enhance the YOLOX deep learning method, the latest iteration of the YOLO series algorithm, for identifying and grading mushroom quality. Firstly, we construct the YOLOX model through transfer learning after expanding the image dataset. Secondly, we optimize the built model using a channel pruning algorithm. Finally, we further fine-tune the pruned model through knowledge distillation and utilize the image dataset to train the optimized YOLOX network model. Experimental results demonstrate that our improved YOLOX method effectively inspects the surface texture of shiitake mushrooms, achieving mean average precision (mAP) and F1 score (FSP) of 99.96% and 57.3856, respectively, while reducing the model size by more than half.