CEG-YOLO: A lightweight edge-optimized framework for in-field rice panicle counting

Abstract

Rice panicle number per unit area is a key determinant of yield, but manual counting remains time-consuming and labor-intensive. This study proposes CEG-YOLO, a lightweight deep learning model for in-field rice panicle detection and counting using consumer-grade RGB imaging devices. The model introduces three improvements to YOLOv11s to address specific challenges in field scenarios: C2f-Fast replaces standard convolutions with depthwise convolutions to reduce computational cost for edge deployment; SPPF-ECA integrates an attention mechanism to suppress complex background interference; and GhostConv reduces feature redundancy to improve detection of dense panicles. A dataset of 5,175 images was collected from four rice cultivars planted at three densities using an iPhone 12. The proposed model achieved 93.9% average precision (AP) on the test set, outperforming YOLOv11s which achieved 89.1%, while reducing parameters to 7.8 million and floating-point operations (FLOPs) to 16.5 billion. Robustness evaluation yielded coefficients of determination (R²) values of 0.91 to 0.94 for lighting, 0.89 to 0.94 for planting density, and 0.90 to 0.94 for cultivar. A cross-year field test using an NVIDIA Jetson Orin NX edge device on 120 quadrats in 2025 achieved R² of 0.91, root mean square error (RMSE) of 4.0, and mean absolute error (MAE) of 3.3 at 20 frames per second, confirming practical deployability. This study demonstrates that smartphone-based proximal phenotyping with an optimized deep learning model can provide accurate, low-cost rice panicle counting for breeding and production applications.

Plant phenotyping relevance

イネ穂数という植物形質をRGB画像から推定する深層学習モデルを開発し、精度・頑健性・実地展開性能を検証しており、フェノタイピング手法が研究の中心である。

Code and data availability

公開論文であることは確認できましたが、現在の公式API・許可済み取得経路では本文を自動取得できませんでした。

No evidence-backed public reproduction asset is currently recorded.