Traffic Reduction in Video Call and Chat using DNN-Based Image Reconstruction
Shota Watanabe, Takuya Fujihashi, Shunsuke Saruwatari, Takashi Watanabe · 2019
In this paper, a traffic reduction scheme for videobased call and chat applications that uses deep neural network (DNN) based super resolution is proposed. Specifically, a sender transmits low-quality, low-resolution video frames containing face information in order to reduce the amount of video traffic. The receiver uses DNN-based super resolution to reconstruct highquality, high-resolution video frames from the low-quality video frames. The proposed scheme makes two contributions. First, face features are adopted for parameter optimization of DNNbased super resolution for high-quality image reconstruction, and second, the scheme includes a newly designed loss function that considers face features that allow high-quality face-containing video frames to be reconstructed at the receiver. According to our evaluation results using real video frames of video calls, the proposed scheme reduces the amount of video traffic by more than 90% as compared with conventional schemes that implement the standard video encoder. In this case, the proposed scheme achieves a reconstructed image quality up to 0.85 in terms of structural similarity (SSIM).