DefinePK

Abstract

In recent years, video files have significantly increased in size, posing challenges for storage and transmission. One effective solution has been to reduce the length of videos. In this study, we introduce Deep Bi VC, a dual-branch framework designed to improve video compression. The model was developed to address these challenges through two distinct compression strategies. Initially, video sequences were pre-processed by segmenting them into groups of five consecutive frames to enable efficient processing. For the first stage, we implemented an Invertible Neural Network (INN) to develop an image compression module. This component focuses on compressing the first and last frames of each group. Subsequently, a video compression module was developed, utilizing motion prediction techniques to interpolate the intermediate frames between key frames. Experimental evaluations using PSNR and MS-SSIM metrics demonstrated that Deep Bi VC outperformed several state-of-the-art methods. On the VUG dataset, the model achieved a PSNR of 37.16 and an MS-SSIM of 0.98 at 3.2 bits per pixel, indicating superior compression performance.