Comprehensive Analysis of Flow Incorporated Neural Network based Lightweight Video Compression Architecture

The increasing video content over the internet motivated the exploration of novel approaches in the video compression domain. Though neural network based architectures have already emerge as de-facto in the field of image compression and analytics, their application in video compression also result in promising outputs. Adaptive and efficient compression techniques are required for video transmission over varying bandwidth. Several deep learning based techniques and enhancements were proposed and experimented but they didn’t exhibit full optimal behavior and are not end to end trained and optimized. In the zest of a pure and end to end trainable compression technique, a deep learning based video compression architecture has been proposed comprises of frame autoencoder, flow autoencoder and motion extension network for the reconstruction of predicted frames. The video compression network has been designed incrementally and trained with random emission steps strategy. The proposed work results in significant improvement in visual perception quality measured in SSIM and PSNR when compared to some state-of-art techniques but in trade-off with frame reconstruction time sheet.