CNN classification with blending mode data augmentation

Abstract

Convolutional neural networks (CNNs) have revolutionized the task of image classification, a frequently occurring computer vision problem. Progress in network architecture has been the leading factor in this advancement, with the evolution of deeper networks with fewer parameters. Alongside CNN network advancements, dataset augmentations have been implemented to expand the learnable data the network can model from. Limited dataset size is a major issue in building neural network models, particularly due to the problem of overfitting, which arises from models fitting training data too well, at the expense of capturing general trends in data, leading to large test errors. In this thesis, we augment available image data by implementing blending modes to expose the full tonal range contained in each training image. Additionally, we implement a horizontal flip transformation to create mirror versions of training images. These data augmentations are shown to reduce overfitting in CNN models. We explore different combinations of blending mode layers to maximize the validation accuracy of the network model.