Improving Unsupervised Deep Learning Methods for Detection and Colorization of Grayscale Images Based on Contextual Content

This research presents a context-aware, unsupervised deep learning framework for automatic image colorization, addressing challenges like color bleeding and context inconsistencies.The framework integrates a Contextual Attention Mechanism (CAM) into a Generative Adversarial Network (GAN) architecture, which includes three key modules: feature extraction, CAM for refining spatial and semantic relationships, and an adversarial generator for high-quality colorization.Unlike previous methods that require paired grayscale-color datasets, this approach learns directly from unlabeled grayscale images, improving scalability.Extensive experiments on ImageNet and BSDS500 demonstrate superior performance with PSNR, SSIM, and FID values of 32.8dB, 0.96, and 11.2, respectively.The method shows faster training and inference compared to prior GAN-based approaches.Future work includes interactive colorization, real-time edge device deployment, and applications in medical imaging and satellite image enhancement.