Rice panicle detection from unmanned aerial vehicle (UAV) imagery via image restoration and pseudo-depth prior

Accurate rice panicle detection using Unmanned Aerial Vehicle (UAV) imagery is a key prerequisite for automated counting and yield-related phenotyping. However, existing detection methods generally assume clear input imagery and rely primarily on two-dimensional RGB appearance, making them vulnerable to motion-induced texture degradation and the strong visual similarity between rice panicles and background foliage in dense canopies. To address these limitations, we propose a novel framework for rice panicle detection via image restoration and pseudo-depth prior fusion. First, we introduce the Spectral-Geometric State Space Network (SG-SSN) to explicitly recover task-relevant high-frequency textures from the temporally accumulated blur commonly generated during UAV cruising. Second, we utilize the Depth Anything V2 vision foundation model to infer pixel-aligned pseudo-depth maps from the deblurred images, providing complementary relative geometric cues. Finally, these dual representations are processed by our proposed Rice Panicle Detection with Pseudo-Depth (RPDet-PD) model, which leverages hierarchical pseudo-depth fusion and frequency-spatial structure preservation to achieve precise panicle localization. Evaluations on a large-scale, multi-site, and multi-variety UAV field dataset show that SG-SSN achieves a PSNR of 32.19 dB on a synthetic temporal-accumulation restoration benchmark. Furthermore, the complete cascaded pipeline utilizing RPDet-PD achieves a mean average precision (mAP@0.5) of 91.6 % under the evaluated field conditions, outperforming the evaluated baseline methods. The experimental results indicate that the proposed cascaded framework provides a feasible approach for rice panicle detection under complex field conditions.