Abstract:

Implicit Neural Representation (INR) is an emerging technology for representing multimedia signals, such as RGB images, with small data size. A key issue in INR is the accuracy of high-frequency details in RGB images under the limited data size. Many studies in machine learning have discussed activation functions and coding to improve the accuracy. This paper aims at the same goal and proposes a novel communication-oriented solution for INR. To represent high-frequency details to the user, the proposed scheme exploits analog transmission for the residual signals between the original image and the decoded image derived from the trained INR. The integration of the INR and analog transmission provides high-frequency details to users with low traffic, and further improves the image quality as a function of the wireless channel quality between the transmitter and each user. Evaluations using an RGB image dataset show that the proposed scheme achieves better image quality than the existing INR-based image compression and standard image codecs under the same amount of traffic.