BRIC: Bottom-Up Residual Vector Quantization for Learned Image Compression

This paper presents Bottom-up Residual vector quantization for learned Image Compression (BRIC). This novel deep learning-based image compression method quantizes latent representations beginning with the lowest resolution and proceeding with residual quantization to the highest resolution. BRIC outperforms the top-down approach from prior work, reducing average bit rates for a given image quality on three vehicular datasets. BRIC is designed with a shallower architecture than prior work, which enables faster encoding/decoding speeds and lower peak memory usage, both critical for practical application. The importance of BRIC is further demonstrated in vehicular networks, where reduced bit rates can double frame delivery rates under challenging channel conditions in a C-V2X network and significantly increase the probability of achieving target frame rates in 5G networks. This bottom-up quantization approach inherently captures a notion of feature importance, making it well-suited for future incorporation into semantic communication systems that utilize unequal error protection during transmission. Overall, BRIC represents an important step towards enabling efficient, reliable, and scalable visual data sharing on the roadway, significantly enhancing the performance of vehicular networks.