Automated Initialization Method in Convolutional Neural Networks using BO Applied to Plant Disease Classification

Abstract

The Convolutional Neural Network (CNN) stands out as the most effective deep learning model for image classification due to its utilization of convolutional kernels for feature extraction. The initialization methods of these kernels significantly impact a CNN performance, with proper initialization leading to faster convergence and improved overall performance, while poor initialization can hinder the learning process.Our paper introduces a novel algorithm called BO-IKM, which leverages Bayesian Optimization to determine the optimal kernel initialization methods for each convolutional layer in a CNN. This systematic approach enhances the model's accuracy and precision by identifying the best initializers. We validated the effectiveness of BO-IKM using the "Plant Pathology 2020" dataset, a challenging image classification problem. The results were compelling, showing significant improvements in accuracy and precision for CNN models optimized with BO-IKM compared to those using standard initialization methods. These findings underscore the potential of BO-IKM to enhance CNN performance across various image classification tasks.