Image and video compression using deep network

A large fraction of internet traffic revolves around the image and video transfers. All the moments, memories, and experiences that we share online are heavily dependent on strong image and video compression. Strong compression techniques significantly reduce the cost of transmission and storage. Traditional image and video compression techniques are laboriously hand-designed and hand-optimized, and thus become less efficient for current needs. In this work, we explore a series of image and video compression architectures to improve the performance of compression. On the image compression side, we explore the model that integrates auto-encoder and GANs. The results show that WGAN performs better among all the models we tried and is worth exploring in the deep video codec. On the video compression side, our model is based on Wu et al. [43] and [19]. We formulate the video compression problem as a joint rate-distortion optimization problem. This helps us to efficiently throw out a lot of information from the bottleneck layer and achieve good performance with lower bit-rates. Our deep video codec outperforms todays prevailing distance models by Wu et al. on Kinetics dataset in terms of PSNR. From the results we have on PSNR metrics, we believe that we can achieve a significantly better performance in video compression