Deep neural network compression for semantic video communications

Deep neural network-based video coding systems are emerging to become mainstream video coding standards. However, a key challenge is the efficient transmission of neural network parameters between transmitters and receivers. We introduce a novel neural network transmission framework that integrates Versatile Video Coding (VVC) and semantic communication principles to transfer the parameters of the Deep Neural Network (DNN) between the transmitter and the receiver by exploiting the spatial correlation between the weights and biases to optimize them for each video scene. The effectiveness of the proposed network compression model was evaluated by comparing the bitrate required to transmit ten different videos with that of the state-of-the-art NNCodec. Experimental results demonstrate that the proposed approach consistently achieves lower bit rates while maintaining comparable video quality, outperforming NNCodec, and is a potential solution to reduce bandwidth requirements for transmission of neural network parameters in video transmission, as well as in more generalized applications.