Answers are grounded in this paper and cite their sources.
With the rapid development of intelligent agriculture, accurate and efficient yield estimation has become critical for the management of high-density crops. However, in scenarios such as the maturity stage of Brassica Chinensis, the extreme object density and occlusion pose significant challenges to traditional counting methods. In this work, we propose AgriDenseNet, a novel density-based counting framework designed to address the complex characteristics of densely overlapped Brassica Chinensis. Built upon the classical CSRNet, our method incorporates three key improvements. First, an ASPP-based Perception Module is inserted between the frontend and backend to enhance multi-scale feature extraction. Second, a hybrid loss function, combining gradient-constrained loss and attention-guided weighting, directs the model's focus toward crucial regions, improving local detail accuracy. Third, a local context enhancement module is proposed to further boost the salience of densely packed targets through additional feature injection. Experiments on our Brassica Chinensis dataset demonstrate the superiority of our model, achieving an MAE of 1.936, surpassing existing state-of-the-art counting models. Furthermore, we validate its generalization ability on the Global Wheat Head Detection (GWHD) dataset(David et al., 2020), where our approach maintains a strong lead. To reduce the reliance on labeled data, we extend our model into a semi-supervised framework. Using only 10% labeled data, the model achieves an MAE of 3.096, significantly improving over the labeled-only baseline. These results highlight the model's effectiveness, generalizability, and practicality for real-world dense crop counting tasks.
No takes yet. Share an insight, caveat, or question.
ZHOU et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: