Implicit Neural Representation for Video Coding Through Progressive Feature Extraction

Jihoo Lee, Je-Won Kang

Abstract: Implicit neural representation for video (INRV) utilizes a neural network to represent a video without relying on an explicit pixel-wise representation of RGB values, and, therefore an INRV-based video coding method aims to compress a neural network and transmit its parameters through a bit-stream. In this paper, we propose a novel rate-distortion (R-D) optimized INRV-based video coding method through a progressive feature extraction module (PFEM). The PFEM consists of a series of residual blocks (RBs) that can progressively control the quality of the reconstruction, by searching the optimal network architecture. By increasing the number of RBs, the quality of a reconstruction frame is improved. However, the quality comes at the expense of overheads in transmission. In the proposed method, the number of the RBs are decided to provide the best trade-off between the frame quality and bits. Our decision model is based upon the R-D theory. Experimental results demonstrated that the proposed method provides a superior coding gain to the conventional video coding standards.