A method of choosing a pre-trained convolutional neural network for transfer learning in image classification problems

Abstract

A method of choosing a pre-trained convolutional neural network (CNN) for transfer learning on the new image classification problem is proposed. The method can be used for quick estimation of which of the CNNs trained on the ImageNet dataset images (AlexNet, VGG16, VGG19, GoogLeNet, etc.) will be the most accurate after its fine tuning on the new sample of images. It is shown that there is high correlation (ρ ≈ 0.74, p < 0.01) between the characteristics of the features obtained at the output of the pre-trained CNN’s convolutional part and its accuracy on the test sample after fine tuning. The proposed method can be used to make recommendations for researchers who want to apply the pre-trained CNN and transfer learning approach to solve their own classification problems and don’t have sufficient computational resources and time for multiple fine tunings of available free CNNs with consequent choosing the best one.