Chaoke Pei

Fast mode selection scheme for H.264/AVC inter prediction based on statistical learning method

H.264 adopts variable block size motion estimation and Rate-Distortion-Optimization based mode decision to improve video quality and compression ratio. These techniques have made H.264 better than other existing video coding standards. However, they are computationally intensive and time-consuming. In this paper, a fast mode selection scheme is proposed for H.264 inter prediction. Firstly, the first few frames are encoded and thresholds are acquired through a statistical learning process. Then, for the rest of frames, motion estimation and mode decision are only performed for the candidate modes which are selected with the proposed fast mode selection scheme. The proposed approach is applicable to all existing motion search algorithms. Besides, thresholds are on-line computed separately for each sequence. Results show that the total encoding time is saved by 57.2% on average with negligible video quality degradation.

A Real Time Ball Detection Framework For Soccer Video

Despite a lot of research efforts, it is still challenging to detect the ball in soccer video, especially in long shot frames. In this paper, we propose a novel real time ball detection framework for broadcast soccer video in long shot frames. Firstly each frame of the video is preprocessed to remove the regions outside ground field and build a saliency binary map. Then, the ball candidates are detected and examined by a rule-based algorithm proposed in this paper. At last, making use of Kalman filter as an effective supplement, the ball location is decided. Experimental results demonstrate that the proposed framework is quite effective to recognize the ball.

A model of visual attention detection based on phase spectrum

Visual attention detecting has drawn great research interest due to its useful application in the area of multimedia. Recently, a fast spectral residual approach in frequency domain was proposed to find the Region of Interest in an image. It was proved to be able to give better performance than the classic approaches. In this paper, we propose a new model for Region of Interest detection based on the phase spectrum. First, we analyze the reason that the salient region can create “pop-out” in the feature maps of the phase spectrum in spatial domain. Second, in order to obtain better feature maps, we propose to blur the image and then get the feature maps based on the phase spectrum of the blurred image. Last, we present a model to generate Region of Interest accordingly. Experimental results show that our approach achieves better and more robust results than that from the state-of-the-art.

An early mode decision algorithm for H.264 optimization

The H.264 standard provides significantly high compression efficiency with multiple block size Motion Estimation adopted. However, the encoding time of Motion Estimation and Discrete Cosine Transform is dramatically increased as a result. An early mode decision algorithm is proposed in this paper to control the computation complexity. It derives a threshold to make mode decisions in advance. Experimental results show that 75.11% encoding time and 84.25% Motion estimation time can be saved on average with little PSNR dropped.

A ROI Detection Model for Soccer Video on Small Display

With the proliferation of small-display mobile devices, the video sequences which are captured for normal size screen, may give small-display viewers uncomfortable experience, especially for some kinds of sports videos. We focus on soccer video, and propose to establish a saliency binary map and several character rules to detect the ball and decide the region-of-interest accordingly. Experimental results demonstrate that both the ball detection and the region-of-interest determination are effective.

A fuzzy logic based fast inter mode selection algorithm in H.264/AVC

H.264 adopts variable block size motion estimation and rate-distortion optimization based mode decision. They are two of the techniques that H.264 exploits to improve visual quality and achieve higher compression ratio. However, the computational complexity and encoding time are enormously increased. In this paper, a fuzzy logic based fast inter mode selection algorithm is proposed for H.264 inter prediction. With the proposed algorithm, candidate encoding modes with maximum probabilities are selected for motion estimation and unnecessary modes are cast off. Results show that about 54% on average of the total encoding time is saved with negligible quality degradation.

Fast algorithm for multi-reference and variable block size motion estimation in H.264/AVC

H.264 adopts Variable Block Size Motion Estimation and multiple reference frames to improve compression rate and video quality. However, these techniques increase the computation complexity and the encoding time enormously. In this paper, a fast scheme is proposed for H.264 inter prediction. The scheme includes two parts. One is the frame selection algorithm which enables the most likely reference frames and discards the rest. The other is the mode selection algorithm that selects modes with more possibility of being the best encoding mode. Results show that 52.91% of the encoding time and 68.37% of the ME-Time can be saved on average with negligible quality degradation.

A PFT Visual Attention Detection Model Using Bayesian Framework

Visual attention refers to the perceptual quality that makes an object or a region pop out relative to its neighbors and seize humans visual attention. Recently, a new fast approach based on phase spectrum of Fourier Transform (PFT) was proved to be effective and also parameter-free. In this paper, we present a novel improved saliency detection model using PFT as well as the Bayesian framework. The bottom-up saliency is gathered based on PFT in several color channels and the Bayesian framework is used to incorporate top-down information with this bottom-up saliency. Experiments show that our fast PFT-based Bayesian model achieves better and more robust results than that from the state-of-the-art.