Answers are grounded in this paper and cite their sources.
Maize tassel counting is critical for breeding and yield evaluation, yet faces occlusion and background interference. To reduce reliance on costly manual annotations, semi-supervised learning offers a solution but suffers from confirmation bias due to pseudo-label noise. To solve the above problems, this paper presents a dual uncertainty-driven semi-supervised counting network (DUET-Net). First, we introduce the parallel perception attention (PPA) module to enhance feature extraction capabilities specifically for small targets. Second, a multi-task network including density regression, classification and uncertainty auxiliary branches is proposed. Designed density level constraints and uncertainty ranking loss serve as multi-task regularization terms to improve prediction accuracy in the main branch of density regression. Furthermore, this study innovatively proposes a dynamic uncertainty masking training strategy. In the supervised phase, the uncertainty-hybrid hard mining (UHHM) strategy employs density map prediction errors and regression uncertainty. In the unsupervised phase, the dynamic reliability filtering (DRF) strategy involves a filtering mechanism based on classification information entropy and regression uncertainty. This method integrates progressive threshold adjustment to facilitate gradual pseudo-label learning, effectively mitigating noise accumulation during semi-supervised training. Finally, the experimental results on a self-constructed dataset of maize tassels demonstrate that DUET-Net outperforms existing semi-supervised methods in the range of 5% to 50%. Particularly at an annotation ratio of 50%, the mean absolute error and root mean square error are 8.49 and 10.95. Moreover, experimental results on public datasets such as wheat and apples have substantiated the effectiveness of the model presented in this paper.
No takes yet. Share an insight, caveat, or question.
Wang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: