Network adaptation for color image semantic segmentation

Image analysis using deep learning has made significant progress in the last few decades, and importance of pre-processing input images become evident. However, adapting a network structure suitable for not been considered. In this study, simple adaptation method color image is described. The illustrated semantic segmentation, which mainly takes as input. inspired by chrominance subsampling, practical video analysis. human visual system less sensitive to differences than it brightness, based on phenomenon, possible improve existing networks providing resolution chroma information luminance encoder design applying group convolution concept. proposed helps achieve improved results without changing complexity baseline model, especially helpful applications with limited resources, such autonomous driving, augmented reality. Experiments were performed combination datasets (i.e. CamVid, Cityscapes KITTI-360) ENet, ERFNet, Deeplabv3plus mobilenetv2). show that improves performance structures increasing number parameters.