Colorizing Images with Split-Brain Autoencoders and Convolutional Neural Networks

This research presents a novel methodology for image colorization using the architecture of Split-Brain Autoencoders (SBAE) with Convolutional Neural Networks (CNNs) in the Caffe framework for predicting missing colors in grayscale images. Image Colorization has been an ongoing research topic in computer vision for many years as this is a strenuous task due to the high dimensionality of the color space, the variability of color perception, and the ambiguity of the grayscale images. Our model processes only the ‘A’, ‘B’ color channels from the LAB color-space and combining the channels at the end of the model to generate the output color image reducing the computational overhead and improving efficiency. The proposed deep learning model is trained on ImageNet dataset and evaluated on standard benchmark datasets. Our results reveal that the suggested method has significant improvements in accuracy and visual quality, achieving higher PSNR and SSIM values over prevailing methods.