Empirical Examination of Color Spaces in Deep Convolution Networks

Abstract

In this paper we present an empirical examination of deep convolution neural network (DCNN) performance in different color spaces for the classical problem of image recognition/classification. Most such deep learning architectures or networks are applied on RGB color space image data set, so our objective is to study DCNNs performance in other color spaces. We describe the design of our novel experiment and present results on whether deep learning networks for image recognition task is invariant to color spaces or not. In this study, we have analyzed the performance of 3 popular DCNNs (VGGNet, ResNet, GoogleNet) by providing input images in 5 different color spaces(RGB, normalized RGB, YCbCr, HSV , CIE-Lab) and compared performance in terms of test accuracy, test loss, and validation loss. All these combination of networks and color spaces are investigated on two datasets- CIFAR 10 and LINNAEUS 5. Our experimental results show that CNNs are variant to color spaces as different color spaces have different performance results for image classification task.