VimicroABCnet: An Image Coder Combining A Better Color Space Conversion Algorithm and A Post Enhancing Network

Ming Li; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2019, pp. 0-0

The framework of combining a better color space conversion (ABC) algorithm,and a post enhancing network for image coding, called VimicroABCnet[??] , is described in this paper. The ABC algorithm employs the principle component analysis[??] method, to find a new primary base axis offering the highest variance for each individual image. The RGB values of each pixel are pre-processed by a 64x64 template filtering. The pixels are then converted by the proposed ABC algorithm, before being encoded by an open source coder[??]. During decoding, the least square method (LSM) has been introduced to estimate the optimal inverse conversion, instead of using a matrix inversion directly. Another feature of the VimicroABCnet is the enhancing network, which adopts the architecture of a classic ResNet[??], and post-processes the decoded RGB image after ABC. Experiments on the CLIC2019 valid dataset have shown significant RGB-PSNR boost of 0.26db or 7.4% bits save@0.145bpp, and 1.2db/22.5%@1.0bpp, making use of the ABC algorithm; and a RGB-PSNR boost of 0.30db@0.15bpp, making use of the enhancing network, respectively. Combining both techniques, an improvement of 0.56db or 12% bits save@0.15bpp; and a decrease in the compressed file size of about 17.8% are achieved in the transparent track. It is noted that each of the two techniques contributes equally. Methods to speed up the decoder model are also discussed.