Improving the Performance of an Image Classification with Convolutional Neural Network Model by Using Image Augmentations Technique

Abstract

        Image classification is one of the challenges for both humans and computers because it is a critical task. Researchers try to solve this task in many ways. In neural networks, Convolutional Neural Network (CNN) has been widely used for images recognition, images classifications, and objects detections. The process of CNN model includes training and testing. This paper aims to improve the performance of an image classification with Convolutional Neural Network model by using image augmentation technique and compare the accuracy of each technique. The dataset that we used in this paper was collected from Canadian Institute for Advanced Research (CIFAR-10) dataset which consists of a total of 60,000 images. The color images are of size 32x32 in 10 different classes. In the first stage, the 10- fold cross validation method is used to divide 50,000 images for training 300 rounds and 10,000 images for testing. In the second stage, in order to create artificial images, 10,000 images were randomly-selected from all categories and applied the image augmentation technique such as color adjustment, rotation, zoom, or flip image to the image. This image augmentations technique were randomly selected only one technique and applied to one image. The training images were 60,000 images in total; 50,000 images from the first stage plus 10,000 new images. Then the computer trained the images for 300 rounds. The evaluation shows that after using image augmentations technique by creating artificial images, the accuracy of Convolutional Neural Network model performs better from 84.79% up to 87.57%