Automatic colorization of grayscale images using deep learning fusion technique and accuracy analysis

Abstract

Recent advances in technology have opened up new ways for the automatic colorization of grayscale images using different deep-learning techniques. Automatic colorization plays a crucial role in various applications such as coloring low-light or night vision images, historical photo restoration, lights-out factories, medical imaging, and many more. Through an extensive review of the literature, colorization techniques were studied. Initially, different colorization methods and error calculations of colorized images were compared and analyzed. In recent years, Convolutional Autoencoder Neural Networks have outperformed in many fields of image generation, denoising, and style conversion which fully proves the potential of the use of Autoencoders in grayscale image colorization. Progressively, different parameters which affect the colorization were changed and compared. Building upon this knowledge, an improved algorithm using a fusion-based approach was developed. The algorithm concatenates local and global features of the image for plausible colorization results. The model was validated by demonstrating extensively using different types of images and is quantitively as well as qualitatively compared with several existing models. Seven types of error calculation methods are used in this thesis for comparison of the original and the predicted images. Furthermore, due to the lack of quantitative visual representation for the accuracy analysis of colorized images, a new concept of error analysis using heat maps is developed