Abstract Image colorization transforms grayscale images into realistic color representations. It is a challenging area of research in computer vision due to uncertainty in mapping intensity values to chromatic information. Earlier approaches, often based on optimization or reference-guided strategies, depend extensively on manual input or exemplary images. In contrast, recent advancements in deep learning methods enable more autonomous and reliable colorization. This article offers a deep transfer learning framework to accomplish high quality and efficient colorization of any gray-scale image. A pre-trained Visual Geometry Group (VGG19) network is developed as the backbone to extract both fine-grained textures and high-level semantic features. This network is designed with 16 convolutional & 3 fully connected layers. To get more prominent contrast and color in output image, the Contrast Limited Adaptive Histogram Equalization (CLAHE) is applied as a post-colorization step. CLAHE improves contrast and overall color quality of the output image. Experiments ware conducted on multiple datasets like ImageNet, COCO-Stuff and Places365. The experimental result proves that the proposed approach achieves outstanding performance in terms of PSNR and SSIM compared with prevailing techniques. Assessments of visual eminence also show improved color vibrancy and perfect contrast adjustment. The approach can be effectively used for restoring old photos, improving black and white movies, and enhancing black scanned medical images.
Ghosh et al. (Wed,) studied this question.