A wavelet-based frequency-domain approach for accurate multi-crop disease detection.

Abstract

Timely localization and diagnosis of crop lesions are critical for disease control and reducing pesticide use. However, in-field lesions often resemble leaf textures, vary widely in scale, and suffer from lighting and shadow interference-making simultaneous high accuracy and lightweight inference challenging. We propose WGA-YOLO, a lightweight YOLO variant for crop disease recognition. Central to our design is Wavelet Channel Recalibration (WCR), a DWT-based downsampling module: discrete wavelet transform naturally provides multi-resolution, time-frequency localized representations that explicitly separate low-frequency approximations from high-frequency edge/texture details. WCR fuses high- and low-frequency components and enhances feature representation through their frequency-domain complementarity, thereby preserving semantic and fine texture information during resolution reduction with negligible extra cost. We also introduce PS-C2f, which integrates Pinwheel-shaped convolutions into C2f to better capture tiny lesion details via multi-directional, irregular kernels, and replace SPPF with Dynamic Group Attention Pooling (DGAP) for efficient multi-scale context aggregation. On our PlantDoc_boost dataset, WGA-YOLO improves over YOLOv8n by 3.02 and 2.85% points, while reducing parameters and FLOPs by ~ 0.18 M and ~ 0.3G, demonstrating improved inference efficiency and deployment friendliness while maintaining strong detection performance in field scenarios.

Code and data availability

The paper's PlantDoc_boost dataset (the annotated crop-disease image dataset constructed and analyzed in this study) is explicitly stated to be publicly released on the authors' GitHub repository. No author analysis code or trained model checkpoints are stated as available. The Roboflow corn and tomato datasets are pre