Abstract

Accurate, high-throughput quantification of rice panicles is important for yield estimation and breeding-oriented rice phenotyping. However, unmanned aerial vehicle (UAV)-based panicle detection remains challenging because flight-altitude variation produces large target-scale changes. Flooded paddy backgrounds, leaf occlusion, and illumination fluctuations further obscure small panicle targets. To address these challenges, we constructed a composite multi-altitude dataset covering UAV imagery acquired from 3 to 20 m under varying field conditions. We then propose Panicle-DETR, a lightweight detection and counting framework based on a frequency-aware Cross Stage Partial (CSP) backbone. Rather than treating the Fast Fourier Transform (FFT) as a filter by itself, the proposed FasterFD module uses frequency-domain representations with learnable frequency-response reweighting to enhance panicle-related texture cues and reduce redundant background responses. A Lossless Feature Encoder is designed to preserve fine spatial information for small targets across altitude-induced scale changes, while a composite metric loss based on Normalized Gaussian Wasserstein Distance (NWD) and Inner-IoU improves localization for adherent and overlapping panicle clusters. On the composite dataset, Panicle-DETR achieved a Precision of 90.97%, a Mean Absolute Error of 4.28, and an R2 of 0.957 for single-frame panicle counting. With 13.78 M parameters and 53.0 GFLOPs, the framework achieved 16.9 FPS with 1.96 GB peak GPU memory in a battery-powered notebook benchmark, supporting its potential for resource-constrained field-side UAV image analysis.

IPC Classification

Keywords