Random Pixel Selection through Image Cropping for Data Augmentation and Classification

Abstract

Data augmentation is one of the strategies addressing the challenge of lack of data, which constitutes an obstacle to getting high accuracy during training Deep Neural Networks. Through this paper, we propose a new method based on the random selection of pixels by cropping some rows and columns from a given image to create new images with a reduced number of pixels. It allows the creation of images with new rows and columns which differ from the original image. This assures enriching and diversifying the original dataset while trying to capture the regions of interest and dismiss the other regions, which leads to better generalization of the training model, The accuracy of the proposed method on Deep Convolution Neural Network applied to the Kaggle Cat VS Dog dataset achieves 77.05% with 40000 samples, whereas the accuracy when applied the original dataset (with 10000 samples) was 73.65%.