Comparison of Full Training and Transfer Learning in Deep Learning for Image Classification

Abstract

The deep learning algorithms on a small dataset are often not efficient for image classification problems. Make use of the features learned by a model trained on large similar dataset and saved for future reference is a method to solve this problem. In this work, we present a comparison of full training and transfer learning for image classification using Deep Learning. Three different deep learning architectures namely MobileNetV2, InceptionV3 and VGG16 were used for this experiment. Transfer learning showed higher accuracy and less loss than full-training. According to transfer learning results, MobileNetV2 model achieved 98.96%, InceptionV3 model achieved 98.44% and VGG16 model achieved 97.405 as highest test accuracies. The full-trained models did not achieve as much accuracy as that of transfer learning models on the same dataset. The accuracies achieved by full-training for MobileNetV2, InceptionV3 and VGG16 are 79.08%, 73.44% and 75.62% respectively.