Research on Deep Semi-Supervised Object Detection Based on Active Learning

Abstract

As a key paradigm connecting supervised learning and unsupervised learning, deep semi-supervised object detection technology aims to enhance the generalization ability and training efficiency of models in real-world scenarios where labeled data is scarce. This method effectively alleviates the dependency of traditional fully supervised models on intensive manual annotation by synergistically utilizing a small number of high- quality labeled samples and large-scale unlabeled data. It particularly demonstrates significant advantages in high-cost annotation fields such as remote sensing image analysis, medical image diagnosis, and autonomous driving perception. Experimental verification shows that the collaborative architecture integrating active sampling and semi-supervised learning achieves performance breakthroughs on multiple datasets, including Pascal VOC, MS-COCO, as well as specialized datasets such as remote sensing and underwater sonar. It not only significantly reduces annotation costs but also exhibits good adaptability in challenging scenarios such as domain shift, class imbalance, and small object detection.