Plug-and-play high-frequency feature enhancement for plant image super-resolution.

Abstract

Introduction High-resolution plant imagery is vital for phenotyping, disease monitoring, and precision agriculture. However, image acquisition in real-world conditions is frequently limited by sensor resolution, cost, and environmental noise, resulting in low-quality images. While deep learning-based super-resolution (SR) approaches show promise, they often fail to recover fine structural details that are essential for plant science applications. Methods We propose a plug-and-play high-frequency feature enhancement (HF-FE) module that can be seamlessly integrated into existing SR architectures. The module selectively amplifies high-frequency information, thereby improving the reconstruction of subtle details such as leaf venation, lesion boundaries, and texture patterns, while maintaining computational efficiency. Performance was evaluated on three diverse plant datasets: an oil palm dataset for large-scale plantation imagery, the UAV-based AqUAVPlant dataset for aquatic plants, and the Plant Pathology 2020 dataset for crop disease imagery. Results Across all datasets, models incorporating the HF-FE module achieved consistent improvements over state-of-the-art (SOTA) baselines, with notable gains in PSNR and SSIM. Visual assessments further confirmed enhanced clarity of fine structural features, particularly in challenging plant imaging scenarios. Discussion The proposed HF-FE module provides a flexible and effective enhancement strategy for plant image SR. By improving the fidelity of reconstructed plant imagery, it supports more accurate visualization and analysis, offering a methodological advancement that contributes to intelligent plant sensing, supports digital agriculture, and facilitates sustainable crop management.

Plant phenotyping relevance

植物画像の超解像度復元手法を開発し、複数の植物画像データセットで性能評価しているため、表現型取得を支える計算手法が中心です。

Code and data availability

The paper evaluates an HF-FE super-resolution module on three public plant image datasets (Oil Palm, AqUAVPlant, Plant Pathology 2020), but these are cited prior-work benchmarks, not paper-specific assets. The supplied blocks contain no data availability statement, no author code/model release, and no repository or URL

No evidence-backed public reproduction asset is currently recorded.