Improved Video Compression Using Variable Emission Step ConvGRU Based Architecture

Video content over the Internet is growing rapidly. There arises a need for more powerful and proficient video compression techniques to handle beating stress of voluminous video over the limited bandwidth. Traditional video compression mechanisms are hand-designed, and their architecture is an amalgamation of different modules designed in such a way that different modules are optimized individually instead of achieving end-to-end optimization of the whole network. The positive upshots of deep learning in image compression emerged a breakthrough for video compression as well. ConvGRU, a convolutional recurrent neural network, comprises productive edges of both RNN and CNN. The proposed architecture consists of ConvGRU as basic building blocks implemented in both fixed and variable bit rate models. The experimental results demonstrated that randomized emission step ConvGRU-based architecture gives better performance and provides a base framework for further optimization enhancements.