An adaptive output-feature knowledge distillation of YOLOv5 for identifying magnetic tiles

Identifying magnetic tiles is vital in the sorting and transfer process of magnetic tile manufacturing. Target detection algorithms empowered by deep learning facilitate accurate and fast identification for magnetic tiles. However, many existing deep learning algorithms are limited by the computing power of edge devices to be deployed easily in real production. In this paper, we propose a lightweight target detection method based on the output-feature knowledge distillation of the compressed YOLOv5 network to identify magnetic tiles. Adaptive weighting is used for the output feature distillation to transfer the helpful feature information from the teacher network to the student network in improving the efficiency of the model. To verify the effectiveness of the proposed method, the accuracy, speed, and size of the compressed model in identifying magnetic tiles are tested and evaluated, respectively. The experimental results show that the lightweight model obtained by our method reaches 98.1% in mAP, which is 1.1% higher than the original model. The average inference speed of a single image is 11.1ms faster than that of the teacher network, but only 16.2% of the teacher's model in terms of the number of model parameters.