A Dual Approach to Image and Video Inpainting using 3D-Edge-Curvelet Transform and CEGAN Architecture

This research paper presents innovative methods for both video and image inpainting, addressing the removal of moving foreground and background objects in videos as well as the restoration of corrupted or missing regions in images. The video inpainting method involves a sequence of key stages, including Pre-Processing, Motion Inpainting, and Background Inpainting, with a crucial step being Image Transformation during Pre-Processing. The novel '3D-Edge-Curvelet Transform' is introduced for video inpainting, implemented using MATLAB 2013a. Evaluation is conducted on three videos (beach, crossing ladies, and jumping girl) based on six frames selected to showcase substantial motion in both foreground and background. Remarkable results, in terms of CPU Computation Time and Peak Signal to Noise Ratio (PSNR), are observed in frames 37 (beach), frame 59 (crossing ladies), and frame 109 (jumping girl).In addition, a distinct image inpainting method is introduced, which utilizes Context Encoding Generative Adversarial Networks (CEGANs) along with an innovative loss function to guarantee the perceptual and semantic consistency of in painted images. This approach incorporates a multi-scale framework to effectively address images with missing regions of different sizes. Empirical findings demonstrate its exceptional performance compared to other GAN-based image inpainting methods and established transformation techniques in Mathematics and Machine Learning.