MATeacher: Multi-scale Views Learning and Adaptive Unsupervised Weight for Semi-supervised Pest Region Detection

Agricultural production is extremely important, yet it also faces the challenges posed by pests. Object detection techniques based on deep learning have shown promise, but the lack of annotated data has limited their performance. And the general semi-supervised object detection framework cannot achieve good performance in pest detection scenarios due to the small, complex, and dense nature of pests. To solve these problems, MATeacher is put forward in this paper. Based on the characteristics of pest regions, Multi-scale Views Learning (MSVL) is introduced for joint consistency training from features of high and low resolution views. We also consider the information of different datasets and different stages of the model from an overall perspective. Through the Adaptive Unsupervised Weighting (AUW), we generate the unsupervised loss coefficient, avoiding inefficient and cumbersome manual design. The experiment on the corn borer dataset shows that our method can get an mAP of 68.2 using just 20% supervised data, 6.7 higher than the supervised baseline. Compared with all the state-of-theart (SOTA) models in comparison, MATeacher performs better. Also, compared to the fully supervised baseline, our method achieves around 93% of its performance with only 20% supervised data, almost matching full supervision. MATeacher provides an effective solution to the lack of annotated data in the task of pest region detection.