Light-FCCNet: A Compact Multi-Scale Framework for Accurate Field Crop Counting

Abstract

Accurate field crop counting supports phenotyping, crop monitoring, and yield-related analysis, but real field images often contain scale variation, target overlap, cluttered vegetation, shadows, and visually similar backgrounds. These factors make density-map regression difficult because weak target responses and accumulated false responses in non-target regions can both bias the final count. This study proposes Light-FCCNet, a compact multi-scale framework for accurate field crop counting under a limited parameter budget. The framework integrates three coordinated components: a lightweight feature pyramid aggregation module for compact scale-aware representation, a multi-attention fusion module for refining fused features and suppressing background-induced responses, and an FCC loss that combines pixel-level density regression, count consistency, and structural similarity. Light-FCCNet is evaluated on three public field crop counting datasets, namely GWHD, MTC, and URC. In the canonical ablation trajectory consisting of baseline, baseline_p1, baseline_p1_p2, and full configurations, the full model achieves the lowest errors, with MAE values of 13.28 on GWHD, 18.10 on MTC, and 61.23 on URC, corresponding to reductions of 18.0%, 25.3%, and 35.2% relative to the baseline. Compared with representative general counting baselines and our implementation of TasselNetV2++ under the same experimental protocol, Light-FCCNet obtains the lowest MAE on all three datasets while using only 0.92 M parameters. These results indicate that its advantage comes from the coordinated design of compact pyramid aggregation, attention-guided feature refinement, and counting-oriented supervision, rather than from any isolated module alone.

Code and data availability

The supplied blocks describe evaluation on three public datasets (GWHD, MTC, URC) and a PyTorch pipeline, but contain no authors' code deposit, availability statement, or public URL for the Light-FCCNet implementation, trained checkpoints, or split files. The datasets themselves are cited prior public benchmarks, not a

No evidence-backed public reproduction asset is currently recorded.