Reducing Computational Cost in Chili Leaf Disease Classification Using Optimized RegNet

Abstract

Early detection of chili leaf diseases is crucial for ensuring agricultural productivity and preventing economic losses. Traditional methods of disease identification rely on manual inspection, which is time-consuming, labor-intensive, and prone to human error. In response, deep learning techniques, particularly Convolutional Neural Networks (CNNs), have been widely adopted for the automated classification of diseases. However, many CNN models suffer from high computational demands, making them impractical for deployment on resource-constrained devices. This study presents an optimized CNN approach for chili leaf disease classification, which strikes a balance between high accuracy and computational efficiency. To identify the most suitable model for this task, we evaluate multiple existing CNN architectures, including MobileNet, ShuffleNet, ResNet, and VGG16. Our findings indicate that RegNet demonstrates superior classification performance. To further enhance its efficiency, we apply pruning techniques to reduce unnecessary parameters and floating-point operations (FLOPs), followed by fine-tuning using damping. The results show that RegNetX- 148MF, with a damping value of 0.01, achieves an outstanding validation accuracy of 97.65%, outperforming larger models while maintaining a significantly reduced computational load. The novelty of this research lies in optimizing RegNet for high accuracy with reduced computational costs, thereby enabling its deployment in low-resource environments, such as edge computing and IoT-based innovative farming systems. By leveraging damping-based fine-tuning, we improve model stability, reduce gradient oscillations, and achieve faster convergence without increasing model complexity. Our findings highlight the potential of lightweight CNN architectures optimized with damping for real-time agricultural disease detection, offering a practical solution for precision agriculture and sustainable farming practices. This approach enables farmers to efficiently diagnose plant diseases with minimal computational resources, thereby contributing to improved crop health management and enhanced food security.