Transfer Learning with a Fine-Tuned CNN Model for Classifying Augmented Natural Images

Abstract

Convolutional neural network has proven to be highly efficient in image classification, but this approach has some disadvantages. Notably, a large number of images are required for training, and great time is need for training to achieve high degree of accuracy in the classification. Transfer learning with a pre-trained model can be used to overcome these problems. There are two approaches to transfer learning, feature extraction approach and fine-tuning approach. Based on the used dataset, both approaches were evaluated, and the two approaches yielded high accuracy with a little better result for fine-tuning approach. To overcome overfitting issues that may occur during transfer learning, data augmentation and dropout techniques were applied. The dataset studied contained 6899 images with 8 distinct unbalanced classes, but only 5600 images with balanced classes were analyzed. The classes include airplane, car, cat, dog, flower, fruit, motorbike and person. We used Google’s Inception-v3 model that is trained on ImageNet database and can classify images into 1000 object categories like pencil, Zebra and many animals. We retrained the Inception model to classify the used natural images using the Keras Library with Tensorflow as backend and achieved an overall accuracy of 99.7% for the test data.