Data Augmentation Technique in Neural Network Training

Abstract

The data augmentation technique is used to increase the number of images in an image bank for training a neural network. The technique generates new images from an original image, using elementary operations such as rotation, shift, zoom, noise, contrast enlargement and translation. The new images created are different from the original image, even having the same image as the source, the operations used make the image different when compared point by point with the original image, which provides the neural network with a greater number of possibilities for your training. The data augmentation technique is widely used in cases in which the training set is very small and is not sufficient for the neural network to extract the characteristics of a given class. The technique was used to enlarge an image bank of orange photos that will be classified by a Hopfield network with respect to quality and size criteria. Due to the scarcity of images of bad oranges, in order to balance the image bank so that the analysis of results is coherent, the technique was applied in a set with 59 images of oranges, being 50 good and 9 bad. The image bank was expanded to 100 oranges, 50 good and 50 bad. The results obtained were satisfactory and consistent with a high accuracy classifier.