AUTHOR=Cheng Chenming , Lei Jin , Zhu Zicui , Lu Lijian , Wang Zhi , Tao Jiali , Qin Xinyan 

TITLE=A novel multiscale feature enhancement network using learnable density map for red clustered pepper yield estimation

JOURNAL=Frontiers in Plant Science

VOLUME=Volume 16 - 2025

YEAR=2025

URL=https://www.frontiersin.org/journals/plant-science/articles/10.3389/fpls.2025.1548035

DOI=10.3389/fpls.2025.1548035

ISSN=1664-462X

ABSTRACT=IntroductionAccurate and automated yield estimation for red cluster pepper (RCP) is essential to optimise field management and resource allocation. Traditional object detection-based methods for yield estimation often suffer from time-consuming and labour-intensive annotation processes, as well as suboptimal accuracy in dense environments. To address these challenges, this paper proposes a novel multiscale feature enhancement network (MFEN) that integrates a learnable density map (LDM) for accurate RCP yield estimation.MethodsThe proposed method mainly involves three key steps. First, the kernel-based density map (KDM) method was improved by integrating the Swin Transformer (ST), resulting in LDM method, which produces higher quality density maps. Then, a novel MFEN was developed to improve feature extraction from these density maps. This network combines dilation convolution, residual structures, and an attention mechanism to effectively extract features. Finally, the LDM and the MFEN were jointly trained to estimate both yield and density maps for RCP.Results and discussionThe model achieved superior accuracy in RCP yield estimation by using LDM in conjunction with MFEN for joint training. Firstly, the integration of LDM significantly improved the accuracy of the model, with a 0.98% improvement over the previous iteration. Compared to other feature extraction networks, MFEN had the lowest mean absolute error (MAE) of 5.42, root mean square error (RMSE) of 10.37 and symmetric mean absolute percentage error (SMAPE) of 11.64%. It also achieved the highest R-squared (R²) value of 0.9802 on the test dataset, beating the best performing DSNet by 0.98%. Notably, despite its multi-column structure, the model has a significant advantage in terms of parameters, with only 13.08M parameters (a reduction of 3.18M compared to the classic single-column network CSRNet). This highlights the model’s ability to achieve the highest accuracy while maintaining efficient deployment capabilities. The proposed method provides an robust algorithmic support for efficient and intelligent yield estimation in RCP.