Benchmarking Optimizers in Transfer Learning for Automated Weed Image Recognition

Abstract

This study investigates the performance of different optimization algorithms within Transfer Learning for weed image analysis. Utilizing pre-trained Convolutional Neural Networks (CNNs), we compare Adam, SGD, and RMSprop optimizers for fine-tuning, aiming to enhance weed classification accuracy with limited data. The research evaluates each optimizer's impact on model convergence, accuracy, and robustness across diverse datasets. Experiments, conducted using MATLAB R2020a, employ the AlexNet architecture and a dataset of farming images from the Vidarbha region, Maharashtra, India. Results highlight significant variations in performance based on optimizer selection, demonstrating the critical role of optimization in achieving efficient and effective weed image analysis. This comparative analysis provides valuable insights for researchers and practitioners seeking optimal optimizer choices in Transfer Learning applications for agricultural image processing.