A Deep Generative Convolutional Autoencoder Image Colourizer

Abstract

Image colourization converts greyscale images into colour images. It is used in fields like history to enhance visual realism. Manual and automated methods have been used for image colourization. Automated methods have included Convolutional Neural Networks (CNNs), Generative Adversarial Networks (GANs) and Autoencoders (AE). This paper presents an image colourizer model based on a Convolutional Autoencoder (CAE) that generates high-resolution RGB images from greyscale images without user intervention. The model brightens and sharpens the output of the CAE by applying a Laplacian filter. The CAE was built from encoder and decoder layers, each incorporating four convolutional layers, and either a Tanh or a Sigmoid activation function as the output layer, and either an Adam or an RMSprop optimizer, defining four CAE variants (1 to 4). The CAE was trained on Kaggle’s Greyscale Landscape dataset over 45 epochs, with best performance from variant 4, which incorporated the RMSprop optimizer and the Tanh activation function. Variant 4 was at its best at 25 epochs, achieving an FID score of 51.7. The output of the CAE is then input to a 3×3 Laplacian filter. The Laplacian was found to give the most appealing results with a central value of 5.3.