Enhanced Visual Clarity in Noise-Reduced Image Environment
Nalagatla Mohaneesh Reddy, Thellapally Jagadishwar Reddy, M Udaya Kiran, Veldi Vivek Adharsh, B Natarajan, R. Bhuvaneswari · 2023
This research introduces an advanced deep learning approach aimed at restoring highly degraded aging photographs. While conventional restoration tasks in supervised learning are effective, real-world image degradation is intricate, making it challenging for the network to generalize due to the disparity between artificial images and genuine aged photographs. The proposed research pioneers a distinctive triplet domain translation network, leveraging a comprehensive dataset containing both artificial image pairs and genuine snapshots to address these complexities. The approach involves training two variational autoencoders (VAEs) to encode clean images and aged photos into distinct latent spaces, facilitating translation between these latent representations. The existence of a discernible gap in the latent space enables the proposed translation strategy to effectively generalize to real photos, overcoming real-world degradation challenges. Moreover, this research incorporates a global branch integrated with a partial nonlocal block to tackle structural defects and multiple mixed deteriorations within a single aged image. Additionally, an extra branch is introduced to handle unstructured defects such as noise, blurriness, scratches, and dust spots. The fusion of these branches in the latent space translation significantly enhances the capability to address a wide array of defects commonly found in antique photos. Furthermore, a specialized face refinement network is employed to extract intricate facial details from the old photos, resulting in enhanced perceptual quality of the restored images. The effectiveness of the proposed method is demonstrated through extensive evaluations, showcasing its superior performance in terms of visual quality for restoring ancient images compared to both current state-of-the-art techniques and existing commercial technologies.