ViT-Based Automatic Grayscale Image Colorization with a Hybrid Loss Function

Yet, automatic colorization of black and white images is still a challenging task in computer vision, given the effective feature extraction and perceptual consistency which this task requires.A new advancement in deep learning is introduced in this study, one that will increase accuracy and speed in colorization.It replaces the conventional ResNet blocks with Inception-based building blocks to capture the multi-scale features and integrates the pretrained vision transformer (ViT) into modeling distance-dependent long-range spatial information.A hybrid loss function is introduced, which combines pixel-wise precision (MSE) with perceptual similarity and adversarial feedback to ensure realism and fidelity in the generated images.On all the evaluation metrics, including obtaining a better MSE and FID score and increasing PSNR and SSIM value, the proposed method outperformed the state-of-the-art methods, all measured in spite of other benefits such as extra rendering of natural textures, realistic skin colors, and consistent heavenly colors even in the complex portions of the image.Besides, it shows reduced inference time and rapid convergence time during training, which proves that the model can be used in real-time applications while having constraints regarding resources.In a nutshell, the proposed method presents a realistic yet high-quality solution to automatic colorization in images, and it has excellent prospects for real-life situations in historical photo restorations and generating digital content.