Learned Image Compression with Fixed-point Arithmetic
Heming Sun, Lu Yu, Jiro Katto
Abstract
Learned image compression (LIC) has achieved superior coding performance than traditional image compression standards such as HEVC intra in terms of both PSNR and MS-SSIM. However, most LIC frameworks are based on floating-point arithmetic which has two potential problems. First is that using traditional 32-bit floating-point will consume huge memory and computational cost. Second is that the decoding might fail because of the floating-point error coming from different encoding/decoding platforms. To solve the above two problems. 1) We linearly quantize the weight in the main path to 8-bit fixed-point arithmetic, and propose a fine tuning scheme to reduce the coding loss caused by the quantization. Analysis transform and synthesis transform are fine tuned layer by layer. 2) We exploit look-up-table (LUT) for the cumulative distribution function (CDF) to avoid the floating-point error. When the latent node follows non-zero mean Gaussian distribution, to share the CDF LUT for different mean values, we restrict the range of latent node to be within a certain range around mean. As a result, 8-bit weight quantization can achieve negligible coding gain loss compared with 32-bit floating-point anchor. In addition, proposed CDF LUT can ensure the correct coding at various CPU and GPU hardware platforms.