A new training strategy: Coordinating distillation techniques for training lightweight weed detection model

Abstract

The excessive use of herbicides and the continuous growth of weeds pose a severe threat to agricultural production while also presenting potential environmental and public health risks. Site-Specific Weed Management effectively addresses this issue but relies heavily on real-time and accurate weed detection algorithms. This study innovatively proposed a new training strategy (TASA) to develop a lightweight weed detection algorithm based on YOLOv5. The heart of TASA's innovation was to alleviate information conflicts among multiple distillation techniques and between distillation techniques and student models and to stop distillation at the appropriate time. We used channel pruning technology to compress the model volume and introduced Knowledge Distillation (KD) during fine-tuning to recover the model performance maximally. Concurrently, TASA was used to coordinate multiple distillation techniques to assist in recovering the model training. The experimental results indicated that the volume of the Optimized YOLOv5s was reduced by 79.2% compared to the YOLOv5s, while the mean Average Precision (mAP) and F1-score (F1) reached 97.4% and 95.1%, respectively, with only decreased by 1.2% and 1.5%. Additionally, the detection speed on the CPU increased by 86.64%, reaching 38.423 frames per second (FPS). Meanwhile, we had also developed an online detection system based on PyQt5 and deployed it on Raspberry Pi. The system detected weeds in real-time, achieving a mAP of 96.3% and an FPS of 25.521 when the input image resolution was 416 × 416.