Enhanced Mobile GC Vit Architecture for Efficient Image Classification with Application to Plant Disease Detection

Efficient and accurate automated diagnosis of plant diseases remains a challenge for deployment on resource-constrained edge devices. While hybrid vision transformers like GCViT balance accuracy and efficiency, they often lose critical high-frequency details such as fine lesion textures and leaf margins that are essential for fine-grained disease classification. To address this gap, we propose the Enhanced High-Frequencies Global Context Visual Transformer (EHF-GCViT), a novel hybrid architecture designed to explicitly enhance high-frequency feature retention within a lightweight framework. The core innovations of EHF-GCViT include: first, a customized, lightweight convolutional refinement block based on depthwise separable operations that acts as a learnable pre-processor to preserve discriminative spatial details before tokenization; second, a gated convolutional block that replaces the final transformer stage, reducing the model memory footprint from 46.36 MB to 34.48 MB; and third, an adaptive normalization strategy to stabilize the training of the integrated heterogeneous layers. Extensive experiments on the PlantVillage tomato disease dataset demonstrate that EHF-GCViT achieves superior performance, surpassing the baseline GCViT, standard Vision Transformers (ViT), and CNN benchmarks (e.g., ResNet) in accuracy, precision, recall, and F1-score. These results validate that explicitly modeling high-frequency features within a hybrid transformer design provides a more memory-efficient and accurate backbone for practical plant disease detection systems targeting edge deployment.

@article{Bahrouni2025,
  title     = {Enhanced Mobile GC Vit Architecture for Efficient Image Classification with Application to Plant Disease Detection},
  journal   = {International Journal of Advanced Computer Science and Applications},
  volume    = {16},
  number    = {12},
  year      = {2025},
  publisher = {The Science and Information Organization},
  author    = {Mohamed Jawher Bahrouni and Faouzi Benzarti and Mohamed Touati and Sadok Ben Yahia},
  doi       = {10.14569/IJACSA.2025.01612108},
  url       = {https://doi.org/10.14569/IJACSA.2025.01612108}
}