DDU-BiATeacher: A semi-supervised pest detection framework combining active learning

Agricultural pests pose a serious threat to global food security, making timely and accurate detection crucial for effective pest management. Traditional detection algorithms rely heavily on labeled data, which is time-consuming and labor-intensive. To address this, we propose DDU-BiATeacher, an active semi-supervised detection framework that reduces the need for manual annotation while maintaining high detection performance. The DDU(Distance,Divergence, Uncertainty) strategy is employed to iteratively filter unlabeled data, prioritizing the annotation of high-value samples. Bi-Resolution Cooperation (BiResCo) is designed to tackle challenges in learning small object features, while Dynamic Unsupervised Weight (DynUW) and Adaptive Pseudo-Label Filter (APLF) improve the robustness and utility of pseudo labels in semi-supervised processes. Experiments on the maize borer, citrus pests dataset, and Pest24 dataset show that DDU-BiATeacher outperforms existing methods. Over 92% of fully-supervised performance is achieved by the DDU-BiATeacher with only 20% annotation cost. These results highlight its potential as a scalable and efficient tool for pest control in precision agriculture. • The DDU-BiATeacher framework combines active learning and semi-supervised object detection to reduce annotation costs and improve pest detection performance. • A novel active learning strategy integrates priori information and model divergences to select the most informative and the most representative samples. • The Bi-Resolution Cooperation architecture is introduced to enhance feature learning for small pests. • An adaptive pseudo-label filter and dynamic unsupervised weight improve pseudo-labeling and loss weighting, boosting stability and convergence. DDU-BiATeacher achieves SOTA on self-built and public datasets, excelling in hard-pest detection.