MemoryTeacher: Memory Fusion and CARF for Semi-supervised Pest Region Detection

With the rapid progress of object detection, intelligent and effective crop pest region detection becomes possible. However, supervised learning-based object detection relies significantly on a large amount of labeled data, which presents challenges in real-world applications. Meanwhile, using vision technology to accurately detect insect infested regions is challenging due to their dense distribution and small target. To overcome these obstacles, in this paper, we put forward an innovative semi-supervised pest region detection framework, MemoryTeacher. The framework that successfully reduces the problems of instability in detection results over multiple iterations. Furthermore, for insect infested regions, missed detection is caused by small size, low resolution, and limited information, we provided a Lightweight feature upsampling operation Content-Aware Reassembly of Features(CARF) module to gather more useful information throughout a vast receptive field. The outcomes of the experiments show that MemoryTeacher performs well on various SSOD evaluation index. our semi-supervised method with ResNet-50 backbone achieves 58.0 mAP given only 20% of annotated corn-borer data, which surpasses previous baselines based on pseudo label method by about 3.2 mAP. Additionally, compared with supervised object detection methods, the enhanced model achieves about 75% effectiveness with only 20% of the training set. This approach provides technological references and theoretical studies for semi-surprised pest region detection.