Auto-Image Colorization Of Grayscale Images Using Machine Learning Techniques

There are lots of old pictures in black-and-white and grayscale. When a grayscale image is painted, it can show the clarity of the past image. Many methods were proposed to perform a color transformation on greyscale images. Some include using manual methods and chemicals to more modern ones using graphic design software. However, all of these methods require special skills and a long time. We need a system that can automatically perform color transformations on gray images. The system developed in this study uses the Self-attention and Spectral Normalization GAN (SSN-GAN) algorithm, which automatically transforms grayscale images. This method was chosen because it can solve various image-to-image translation problems. This method produces a better initial score and a more stable training process. Self-attention is used to increase the initial score of the image generated by the GAN. The addition of spectral normalization makes the training process on the GAN run more stable. At the same time, by increasing the learning rate on the generator and discriminator, the model can produce a better image. In this study, researchers used two loss functions, namely, using SSIM Loss and without SSIM Loss. The highest accuracy value in the test with an epoch value of 50 and using SSIM loss, the accuracy value obtained is 59.9