Optimizing Transfer Learning and Fine-Tuning Hyperparameters in Image Classification Problems with Firefly Algorithm
Vinicius Diego Mendes Silva, Gustavo F. V. de Oliveira, Fabrício A. Silva, Marcus Henrique Soares Mendes · 2024
Image classification is crucial in computer vision, mainly with Convolutional Neural Networks (CNNs). This paper optimizes transfer learning and fine-tuning hyperparameters of CNNs pre-trained on ImageNet for still image classification. Hyperparameter tuning is a complex task that impacts the classification results. The Firefly Algorithm (FA) was used to optimize these hyperparameters across four datasets with Xception and ResNet-152 architectures. Experiments show that FA enhances model performance, achieving state-of-the-art accuracy on three datasets: Multi-Class Weather (99.11%), Pistachio (100%), and D0 (99.89%). Despite being time-consuming, this approach offers a viable method for improving image classification, mainly with smaller datasets.