An Impact of Different Images Color Spaces on the Efficiency of Convolutional Neural Networks

Abstract

A common standard while working with convolutional neural networks is using an RGB color space as an input of the network. All popular benchmarks datasets use this standard with three channels (red, green and blue) and with 8 bits per channel. We can modify images and use lower bits size per channel or use different color spaces standards like HSV, CMYK or YIQ. The main contribution of this paper is testing the influence of an image color space on convolutional neural networks. In our experiments we use DenseNet - a state of the art model and CIFAR 10 - a image benchmark dataset. Our experiments show that choosing proper color space have an impact on the final efficiency of the network and using the most popular RGB color space might not always be the best choice.