Machine Learning Enabled Hybrid Image Compression With Implicit Neural Representation For Wireless Semantic Communications

Semantic communication is considered a fundamental breakthrough technology that leverages machine learning intelligence in future 6G wireless networks. A fundamental challenge in semantic communication is to represent the multimedia content’s meaningfulness with a limited number of symbols or bits. In this paper, we propose a hybrid approach to prepare and compress multimedia data traffic for semantic wireless communications. In the proposed approach, a hybrid image compression architecture is developed that combines Implicit Neural Representations (INR) intelligence for important content with JPEG 2000 encoding for less critical data. Specifically, the luminance (Y) channel of the YUV color space is extracted and encoded by overfitting the magnitude values of each pixel to a Multi-Layer Perceptron (MLP), with the machine learning model weights and biases stored as the semantics representation before wireless transmission. Meanwhile, the chrominance (U and V) channels are aggressively compressed using the traditional JPEG 2000 codec in binary bit format. At the decoding stage, the Y channel is reconstructed from the MLP via INR machine learning intelligence and combined with the U and V channels decoded from the traditional JPEG 2000 codec. The experimental results demonstrate that this hybrid approach, applied to the prioritized YUV format, outperforms traditional INR compression in the RGB format with approximate 1 to 2 dB gain at similar bit rates.