Light Teacher: Lightweight Semi-supervised Object Detection Algorithm

A two-stage semi-supervised object detection algorithm based on the ResNet-50 backbone, featuring efficient feature extraction capabilities and excellent accuracy performance. However, this network still faces issues in practical applications, such as high computational costs and low-quality pseudo-labels. To address this, a more lightweight yet high-performance semi-supervised object detection algorithm, called Light Teacher, is proposed. Firstly, a lightweight backbone network, MobileECANet, is designed to reduce the model’s computational load and accelerate its running speed while ensuring accuracy. Secondly, a Context Adaptive Feature Pyramid Network (CAF-FPN) is designed as the feature extraction network to further harness rich contextual information and suppress inconsistencies across different scales. Finally, the Confluence method is employed to replace non-maximum suppression, which not only helps reduce false positives and missed detections in pseudo-labels but also simplifies the post-processing workflow. Experimental results show that under conditions using only 10% labeled data from the COCO dataset and on the VOC dataset, the proposed Light Teacher model improves accuracy by 1.35% and 1.53%, respectively, compared to the original Mean Teacher model.