FG-LCNet: A two-stage foreground-guided network for whole-tree litchi counting

Accurate litchi counting from whole-tree images is essential for yield estimation, orchard management, and plant phenotyping, but remains challenging in real orchards because fruits occur in dense, heavily occluded clusters and vary markedly in scale, illumination, and appearance across ripening stages, particularly when green fruits resemble surrounding foliage. Existing methods have shown promise, but their robustness in complex orchard environments remains limited. To address these challenges, we propose FG-LCNet, a two-stage foreground-guided litchi counting framework. In the first stage, an enhanced fruit-cluster detector improves the localization of small and ambiguous clusters under complex canopy backgrounds. In the second stage, the detected foreground regions are fed into a density-regression network with hybrid attention, while a consistency-based training strategy is introduced to improve robustness to appearance and illumination variations. To support this study, a large-scale litchi counting dataset was established, consisting of 1126 whole-tree images collected from five orchards and spanning three ripening stages, with approximately 120,000 fruit-level dot annotations and more than 20,000 cluster-level bounding boxes. FG-LCNet achieved the best overall counting performance, with an MAE of 7.44 and an RMSE of 11.01. It showed clear advantages in high-density fruit-cluster scenarios and cross-orchard validation, while maintaining competitive results across orchard-region and maturity-stage subsets. The framework further retained inference efficiency suitable for practical deployment. These results indicate that FG-LCNet provides an effective solution for robust litchi counting and offers potential for other clustered fruit-counting tasks.