Design and Evaluation of Generative Adversarial Network for Automated Image Colorization System

For the preservation of historical materials and the enhancement of visual representation, image colorization is crucial. Image colorization adds color to black-and-white or grayscale images, revitalizing old photographs and papers. It affords us a more immersive and relatable experience of history. Additionally, colorization imparts realism and liveliness to visual content. It facilitates audience engagement, emotional expression, and the development of visually captivating content that elicits a response and imparts a memory. The process of image colorization utilizing Deep Learning (DL) models is investigated in this study. The pre-processing techniques utilized to ensure data integrity and consistency are sourced from Kaggle. The image colorization challenge is built using DL models, notably Generative Adversarial Networks (GANs) and Convolutional Autoencoders (CAEs). These algorithms learn to colorize grayscale images through the use of training data. Peak Signal-to-Noise Ratio (PSNR), Learned Perceptual Image Patch Similarity (LPIPS), Structural Similarity Index (SSIM), and Fréchet Inception Distance (FID) are some of the metrics utilized to assess the models. The evaluation results indicate that the GAN model exhibits superior performance in image colorization compared to the CAE model. This is supported by higher PSNR and SSIM values, as well as lower values for LPIPS and FID.