Visual Saliency Estimation Via HEVC Bitstream Analysis

Abstract Since Information Technology developed dramatically from the last century 50's, digital images and video are ubiquitous. In the last decade, image and video processing have become more and more popular in biomedical, industrial, art and other fields. People made progress in the visual information such as images or video display, storage and transmission. The attendant problem is that video processing tasks in time domain become particularly arduous. Based on the study of the existing compressed domain video saliency detection model, a new saliency estimation model for video based on High Efficiency Video Coding (HEVC) is presented. First, the relative features are extracted from HEVC encoded bitstream. The naive Bayesian model is used to train and test features based on original YUV videos and ground truth. The intra frame saliency map can be achieved after training and testing intra features. And inter frame saliency can be achieved by intra saliency with moving motion vectors. The ROC of our proposed intra mode is 0.9561. Other classification methods such as support vector machine (SVM), k nearest neighbors (KNN) and the decision tree are presented to compare the experimental outcomes. The variety of compression ratio has been analysis to affect the saliency.