Dual Fusion Net: A Transformer-Based Hybrid Dual model Architecture for Highly Accurate Chili Leaf Disease Classification

Abstract

Chili leaf diseases significantly impact agricultural productivity, demanding advanced and reliable AI-based detection systems for timely intervention. This research introduces Dual Fusion Net, a Transformer-enhanced hybrid dual-model architecture that integrates InceptionV3 and DenseNet121 to achieve highly accurate chili leaf disease classification. The parallel CNN backbones extract multi-scale and densely connected deep features, while a Transformer-based fusion module learns global contextual relationships across disease patterns. Experimental results demonstrate that Dual Fusion Net outperforms single-model baselines and recent state-of-the-art chili disease classification frameworks. The proposed method achieved an overall accuracy of 98.36%, surpassing standalone InceptionV3 (94.8%) and DenseNet121 (95.6%) as well as recent published models based on EfficientNet, MobileNet, and CNN–Transformer hybrids. Visualization using Grad-CAM and attention maps validates the enhanced interpretability enabled by the Transformer fusion mechanism. The contributions of this experiment is to development of a novel dual-backbone CNN–Transformer fusion architecture, a significant improvement in classification accuracy over cutting-edge baseline models, and real-time inference capability suitable for smart agriculture systems. In future the model is lightweight Edge-AI optimization, transformer efficiency enhancement, and federated learning integration for scalable and privacy-preserving agricultural disease monitoring.