Transfer Learning by Fine-Tuning Pre-trained Convolutional Neural Network Architectures for Image Recognition

Convolutional neural network (CNN) is an important machine learning model that can be applied to the image recognition problem based on extracting the best features. But creating a new efficient CNN model requires considerable training time, it takes more computational resources and a large dataset for high accuracy. There are many pre-trained CNN models that are trained on millions of images. This manuscript proposes the performance of popular pre-trained convolutional neural network architectures on image recognition. The performances of different types of CNN’s are compared on a single dataset. The CNNs are already pre-trained on ImageNet data, so their performances will give a measure of which type of CNN to use for quick fine-tuning with greater accuracy. This study focuses on the concept of transfer learning and analyzing the performance of four popular networks: Alex Net, VGGNet, Inception and ResNet50. Four different CNN’S are taken and fine-tuned them on furniture images. The analysis shows that ResNet50 and Inception can recognize objects with much better accuracy than AlexNet and VGG16 with a smaller number of epochs.