Original title: FG-LCNet: A two-stage foreground-guided network for whole-tree litchi counting
Yitao Zhuang, Mingchao Yang, Zhikeng Ke, Riyao Chen +3 · September 2026
Counting litchi fruits on whole trees is tough. The fruits hang in crowded, overlapping bunches, lighting changes constantly, and green fruits blend right into leaves. Researchers created a two-step system called FG-LCNet to solve this. First, a fruit-cluster detector scans the tree canopy to find small, hard-to-see bunches of fruit. Second, the system sends those fruit areas into a density-regression networkA type of neural network that estimates how many objects are in an image by predicting a density map., which estimates fruit counts by examining crowded spots. This part uses hybrid attentionA mechanism that helps the network focus on important parts of the image. to focus on key visual details and trains with a consistency-based strategy to handle lighting and color shifts. The researchers tested the system on a newly built dataset of whole-tree images. FG-LCNet achieved the top overall counting performance, reaching an MAEMean absolute error, the average difference between predicted and actual counts. of 7.44 and an RMSE of 11.01. It performed especially well on dense clusters and across different orchards, while running fast enough for practical use.
Imagine trying to count all the apples in a large tree from a photo. The apples are bunched together, some are hidden behind leaves, and green apples look just like the leaves. It's hard to tell where one apple ends and another begins. In this paper, the system first finds the bunches of apples, then counts the apples in each bunch separately, which makes the counting much more accurate.
Accurate litchi counting helps farmers estimate yields and manage orchards more efficiently. This work makes it possible even in challenging conditions where fruits are dense, occluded, or green like leaves. That means more reliable harvest predictions and better resource allocation.
The abstract doesn't explicitly state limitations, but the method is evaluated only on litchi counting; its performance on other fruits is not demonstrated. The dataset, though large, comes from five orchards and may not cover all possible conditions. The two-stage framework adds complexity compared to single-stage methods.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
FG-LCNet: A two-stage foreground-guided network for whole-tree litchi counting — Yitao Zhuang, Mingchao Yang, Zhikeng Ke et al. — Plant phenomics (Washington, D.C.) — doi:10.1016/j.plaphe.2026.100268
Abstract licence: CC BY
Explanation generated by Synthesis from the abstract — it is not the original paper.