Image recoloring using generative adversarial neural networks

Abstract

This paper presents an in-depth exploration of a Neural Network designed to recolor grayscale images with minimal input requirements. The paper delves into the intricate process of training the network, which involves carefully selecting a fitness function and creating an effective adversarial network. Throughout the paper, various alternatives are considered and evaluated until a suitable approach is identified for further training. Notably, the implementation adopts a random batch sampling approach to gather images in each batch selection, allowing for diverse and comprehensive training. Moreover, several techniques, including Batch Normalization, Leaky ReLU, and Label Smoothing, are strategically employed to tackle challenges related to generalization and achieve a balanced interplay between the generator and discriminator. The experimental results are thoroughly discussed, showcasing the substantial progress achieved in addressing the problem at hand. Remarkably, the Neural Network attains a Structural Similarity Index (SSIM) of -0.5944 on the test set and -0.5922 on the training set, signifying its proficiency in accurately recoloring grayscale images. This paper contributes valuable insights into the realm of image recoloring using neural networks and demonstrates the effectiveness of the proposed methodology in achieving good results.