ICAG-Net: An Interactive CNN–Transformer Architecture With Attention-Guided Gated Fusion for Crop Disease Detection

Abstract

Smart agriculture based on the use of Artificial Intelligence for crop disease detection to ensure food security. Fungal disease is one of the major causes that affects the quality of the vegetables. Convolutional neural networks (CNNs) and vision transformers (ViTs) enable the detection of crop diseases at an early stage, allowing farmers to take preventive measures and minimize further losses. The proposed method introduces an ensemble of Custom CNN to understand multilevel local features and Pretrained ViT to capture global dependencies as well as contextual information from the dataset. An interactive cross attention module (ICAM) facilitates bidirectional information exchange between CNN and transformer token representations, while an attention guided gated fusion (AGGF) mechanism adaptively combines complementary features. The Indian Crop Visual Disease Dataset (ICVDD-5) has been developed in a real field for the proposed work with the help of domain experts. The dataset contains 880 diverse images depicting both healthy and diseased specimens of five vegetable crops. The crops selected for this research initiative include Brinjal, Cabbage, Chili, Okra, and Tomato. These five crops are examined for about 21 distinct disease classes. Comprehensive ablation studies are conducted to prove the contributions of each architectural component, including CNN-only, ICAM-disabled, and AGGF-disabled configurations. Experimental results demonstrate that the proposed ICAG-Net achieves a test accuracy of approximately 70%–73% with improved macro-F1 score compared to baseline CNN models under identical training settings. The novelty of this work lies in an extensible solution for real world crop disease diagnosis systems and offers insights into hybrid CNN–transformer architectures for small scale agricultural datasets.