Efficient Deep Learning Based Video Compression

Overview

This paper proposes a deep learning-based video compression algorithm that improves the efficiency of video transmission and storage. Every day, 1.5 billion hours of videos are watched across YouTube, Netflix, and Facebook, and 23 million new cameras are added into circulation; additionally, around 85% of all internet traffic is in the form of video data. This has resulted in an unprecedented explosion in the sheer volume of video data, occurring in the last 5 years alone. Despite this, today’s methods of compressing and distributing video data are still the same as 20 years ago. This research provides a solution to the explosion in video data by documenting a novel deep learning algorithm called COMPACT that utilizes a Scale Space Auto-Encoder architecture to compress videos. This technology outperforms traditional video compression algorithms in two distinct metrics: size and quality. It produces compressed videos that are 40% smaller than its nearest competitor, H265, and still has higher reconstruction quality. Moreover, the reconstructed video quality using COMPACT is 5 DB higher in Peak-Signal-T o-Noise-Ratio and 0.15 better in Multi-Scale Structural Similarity Index Measure than H265. Implementing COMPACT in the real world could significantly enhance the fields of online conferencing, video streaming, the medical industry, and video surveillance. Hence, proving the efficacy of deep learning-based solutions in the realm of video compression.

Competition history

• JSHS 2023 Category not listed

Resources

• National JSHS abstract book ↗

Related projects

Closest projects by meaning, across every fair and year in the corpus.