LightFusionNet: A Compact and Efficient Tri-Branch Network for Phytopathogen Detection in Horticultural Crops

• Published in: International Conference on Inventive Computation Technologies (Online) pp. 508 - 515
• Main Authors: Kaleeswari, G, Sundarrajan, R
• Format: Conference Proceeding
• Language: English
• Published: IEEE 23.04.2025
• Subjects: Accuracy; Computational efficiency; Computational modeling; Computational overhead; Crops; Diseases; Early Exit; Explainable AI; Feature extraction; Feature Fusion; Horticultural Crops; Neural networks; Sustainable Agriculture; Training; Transformers
• ISSN: 2767-7788
• DOI: 10.1109/ICICT64420.2025.11005006

Cardamom and coffee are valuable horticultural crops of economic importance but are highly susceptible to various diseases that can cause significant damage to yield and quality. Early diagnosis of these diseases by optimized architecture is essential for crop health management and adopting sustainable agriculture. In this work, an optimized deep learning architecture for cardamom and coffee plant pathogen diagnosis is introduced. The proposed methodology is the Tri-Branch Feature Fusion Network (TFFNet), wherein different feature extraction methods are combined using MobileNetV2, InceptionV3, and Vision Transformer (ViT). The proposed TFFNet architecture enhances feature discovery together with selective attention through Swish along with Entmax(1.5) networks to achieve enhanced classification of Healthy and Diseased Cardamom and Coffee crops. A Lion Optimizer is used to optimize the hybrid structure of MobileNetV2, InceptionV3, and ViT by dynamically varying learning rates from gradient history, optimizing performance. Dynamic neural networks with early exit are also utilized to balance accuracy with low computational overhead to improve the efficacy. The model performs with good accuracy with a training accuracy of 98.23%, validation accuracy of 96.14%, F1 score of 0.961, precision of 0.960, recall of 0.963, training loss of 0.009, and validation loss of 0.010. These findings reflect the potential of the model for effective disease diagnosis. Future directions will involve studying the influence of disease on resistance mechanisms, and integrating Explainable AI (XAI) approaches for increased transparency and trust in prediction.