Data Augmentation On The Convolutional Neural Network For Image Classification

Résumé: In the field of computer vision, image classification is considered as one of the most dominated research domains. Machine Learning, specifically, deep learning becomes the most useful tool that handles image classification tasks. The Deep Learning model needs to be trained on a huge number of samples, which leads us to one of the most popular issues facing this field "The lack of data". Data Augmentation increase the number of training images. In this thesis, we clarify the impact of the amount of training data and the effect of data augmentation on the performance of CNN models in image classification. We evaluate this on Kaggle dataset by manipulating it in a different manner we created three more datasets from the original one. the first one contains 8% of the original dataset, the second one is generated by applying seven image manipulation technique(rotation, shifting, horizontal flipping) on the second dataset with random parameters, the last one was created using supervised data augmentation with specific parameters instead of the random one, then we train each one of the 4 datasets on two different deep learning architecture ResNet and AlexNet. The results obtained show that the more data we feed to the model the better performance we get, using data augmentation increase the level of accuracy besides of using supervised data augmentation show a little better performance than the random augmentation but even a small percentage may make the difference

Publié dans la revue: