陈志波  (教授)

博士生导师

电子邮箱:

学位:博士

毕业院校:Tsinghua University

   

Last-Bit CNN Classified Fast Mode Decision for Screen Content Coding

点击次数:

Abstract

The coding complexity introduced by the new screen content coding tools has posed a great challenge for practical applications of Screen Content Coding (SCC) standard. In our work, we propose a Last-Bit Convolutional neural network classified Fast Mode Decision (LBC-FMD) algorithm to reduce the coding complexity of SCC encoder. The framework of LBC-FMD comprises two stages. In the first stage, a novel Last-bit Convolutional Neural Network (LBCNN) classified structure is proposed to predict the content types of CUs. By extracting the last-bit of each pixel in gray scale image as the input instead of the gray scale image itself, the classification accuracy is significantly improved. In the second stage, a Fast Mode Decision (FMD) scheme is proposed to reduce the coding complexity as well as maintaining the same coding performance, which leverages the content types predicted by the proposed LBCNN classifier. The FMD scheme includes three modules. The first is content-based candidate modes selection module, which is designed to remove unnecessary modes. The second is candidate modes supplement module, which re-activates the previously disabled modes, and the third is content-aware early termination module to avoid examining unnecessary Intra sub-modes and CU sizes. Compared with SCC reference software, our LBC-FMD algorithm can remarkably reduce the encoding time by 40.01% in average while preserving the compression performance efficiently with only 1.30% Bjontegaard Delta-rate (BD-rate) increase with All-Intra configuration under the standard common testing conditions.

Paper: to be published.

Get Source Code

We implement LBCNN module in Caffe and implement our LBC-FMD algorithm on the standard software HM-16.7+SCM-6.0.

You can download the publicly available release of the source code by clicking THIS link. Please fill THIS FORM and the password will be sent to you.


上一条: Multi-view Vehicle Type Recognition with Feedback-enhancement Multi-branch CNNs

下一条: Blind Stereoscopic Video Quality Evaluator