Lightweight frequency-aware hybrid vision transformer for fine-grained plant disease recognition

Plant diseases can substantially reduce crop yield and quality, while accurate disease recognition on edge devices remains challenging because images often contain subtle lesion textures and complex backgrounds. This study proposes a lightweight frequency-aware Dual-Embedding Hybrid Vision Transformer, termed DE-HVT, for fine-grained plant disease recognition. Here, "frequency-aware" does not denote an explicit or orthogonal separation into independent frequency bands. Instead, it refers to complementary frequency-biased processing by two spatially aligned embedding branches: OBPE preserves slowly varying structural cues, whereas the CDC-based TexPE branch has a zero response at zero spatial frequency under the linearised $\theta =1$ analysis and therefore exhibits difference-sensitive, high-pass-like responses to lesion boundaries and textures. Their outputs are fused by element-wise addition before MHSA, while ConvFFN introduces local spatial filtering after global token interaction. Experiments were conducted on the Plant Pathology and PlantVillage datasets using the same training and evaluation protocols for all compared models. Across three independent runs, DE-HVT achieved a mean Top-1 accuracy of $93.51\pm 0.44\%$ and a mean macro-averaged F1 score of $93.64\pm 0.40\%$ on the more challenging Plant Pathology dataset. On the controlled-background PlantVillage dataset, all evaluated models achieved near-ceiling recognition performance. The model contains 4.10 million parameters and requires 0.85 GFLOPs. On a Snapdragon 865-based mobile platform, it used 46.86 MB of memory and recorded a maximum observed inference latency of 111.96 ms per image. These results indicate that DE-HVT provides a favourable balance between recognition performance and computational efficiency within the evaluated datasets and under the tested mobile-device conditions.