Ziguan Cui

Structural similarity optimal MB layer rate control for H.264

Conventional Rate Control (RC) schemes for video coding mostly take objective metric as distortion measurement, which can not achieve optimal subjective quality. This work applies Structural Similarity (SSIM) based subjective distortion to rate distortion optimization and RC in H.264, and proposes a SSIM optimal macroblock (MB) layer RC scheme. First, a SSIM quadratic distortion model is proposed based on extensive experiments and theoretical analysis. Then an improved quadratic Rate Quantization (R-Q) model is combined to obtain the solution of SSIM optimal MB layer quantization step (Qstep) by Lagrange multiplier method. Experimental results show that the proposed scheme preserves more image structural information and thus acquires better subjective quality compared with objective metric optimal MB layer RC JVT-O016 and classic JVT-G012.

Multi-Channel Mixed-Pattern Based Frame Rate Up-Conversion Using Spatio-Temporal Motion Vector Refinement and Dual-Weighted Overlapped Block Motion Compensation

In this paper, a novel motion compensated frame rate up-conversion (MC-FRUC) algorithm is proposed to enhance the visual quality of video sequences. First of all, the multi-channel mixed pattern (MCMP) is proposed to design a block matching criterion which uses a few computations to reveal the variances of one luminance channel and two chrominance channels in a video frame. Second, in basis of the forward and backward initial motion vector fields (MVFs) estimated by a variant of 3DRS algorithm, the spatio-temporal motion vector refinement (ST-MVR) algorithm is proposed to obtain the more smoother MVF by implicitly adding the spatio-temporal smooth constraint into the process of motion vector refinement, and then the highly fault-tolerant motion vector smoothing (HFT-MVS) algorithm is proposed to prevent the emergence of outliers in MVF. Finally, in order to reduce the edge blurring and occlusions, the proposed dual-weighted overlapped block motion compensation (DW-OBMC) algorithm uses the forward and backward MVFs to jointly produce the interpolated frame. Experimental results show that the proposed algorithm can significantly improve both objective and subjective quality of the interpolated frame with a low computational complexity, and provide the better performance than the existing algorithms.

Vehicle type classification via adaptive feature clustering for traffic surveillance video

Vehicle type classification has become an important part of intelligent traffic. However traditional methods can not deal with the varying situations in the reality. In this paper, a novel method is proposed to handle this task in the real road traffic surveillance video. In order to distinguish different vehicles, we categorize vehicles into three types: compact cars, mid-size cars, and heavy-duty vehicles. For a certain video, our method has four steps. First, a deep convolutional neural network is used to detect vehicles in the candidate region and a data set would be generated. Second, the main features of vehicles can be extracted using a fully-connected network. Also, for the sake of higher accuracy, weak labels given by pre-trained extreme learning machine (ELM) are fused into the final features, adding prior information proportionally. Third, K-means is implemented to learn three vehicle-type cluster centers adaptively. Finally, vehicle type will be recognized according to the closest distance principal. Experimental results show that the recognition rate outperforms other traditional methods, verifying the feasibility and effectiveness of the proposed method.

Active tracking using color silhouettes for indoor surveillance

Pan-Tilt-Zoom (PTZ) cameras play an important role in surveillance systems. In this paper, we propose an active tracker using color silhouettes with a single camera. We firstly apply dilation and erosion operators of morphology to binary difference image to get color silhouettes. We also record the color silhouette of the target which we are interested in. Secondly, we measure the similarity of color silhouette between the observation and the candidates of silhouettes. We exploit the most similar one to update that of the tracked target. Finally, we control the PTZ camera to move according to the location of the tracked target. The experimental results show that the proposed algorithm can effectively track people even though she/he is fully occluded.

Edge Directed Single Image Super Resolution Through the Learning Based Gradient Regression Estimation

Single image super resolution (SR) aims to estimate high resolution (HR) image from the low resolution (LR) one, and estimating accuracy of HR image gradient is very important for edge directed image SR methods. In this paper, we propose a novel edge directed image SR method by learning based gradient estimation. In proposed method, the gradient of HR image is estimated by using the example based ridge regression model. Recognizing that the training samples of the given sub-set for regression should have similar local geometric structure based on clustering, we employ high frequency of LR image patches with removing the mean value to perform such clustering. Moreover, the precomputed projective matrix of the ridge regression can reduce the computational complexity further. Experimental results suggest that the proposed method can achieve better gradient estimation of HR image and competitive SR quality compared with other SR methods.

An improved image dehazing algorithm based on dark channel prior

Image dehazing algorithm based on dark channel prior has been proved to be effective, but it cannot still guarantee accurate transmission. To solve this problem, we firstly propose a more reasonable estimation of atmospheric light, because bias in the atmospheric light estimation will cause an inaccurate transmission. Secondly, we improve the estimation of transmission in bright area so as to alleviate color distortion. After image recovery, we carry out denoising to improve the image quality. Finally, we use a blind image quality assessment method based on property of Human Visual System, and the experimental results show that this improved algorithm is more effective.

A low complexity MB layer rate control algorithm based on motion similarity for H.264

This paper presents a simple but effective macroblock (MB) layer rate control (RC) scheme for H.264/AVC with low complexity. First, to reduce computation cost and inaccuracy of linear mean absolute difference (MAD) prediction at MB layer adopted in JVT-G012, MAD is computed directly according to the difference between current original MB and the reference blocks pointed by estimated MV using intensive motion similarity. Then, MB header bits are predicted based on spatial-temporal correlation because it is not constant due to complicated coding modes and high compression efficiency of H.264. Finally, MB target bit rate is allocated according to its complexity and the parameters of quadratic R-D model are updated using coded MBs with high spatial-temporal correlation and motion similarity not the last coded data points. Simulation results show that the proposed scheme achieves an average PSNR gain of 0.37 dB, meets better with target bit rate, produces more consistent quality for the MBs in a frame and thus improves visual quality compared to classic JVT-H017 RC algorithm, simultaneously has lower computation complexity and suits for real-time application.

Image complexity adaptive intra-frame rate control algorithm for H.264/AVC

Aiming at the bad effect of intra-frame rate control (RC) of H.264, this paper proposed an image complexity adaptive intra-frame RC algorithm. First, in order to acquire intra-frame coding complexity more accurately, our algorithm analyzes intraframe by Sobel operator and establishes edge direction histogram for each 4×4 block, then gets the most probable intra prediction mode and corresponding reconstructed blocks, finally obtains the residual picture which is close to the actual coding residual. The mean absolute value of residual picture was used to represent the intra-frame coding complexity, combined with our empirical rate-quantization (R-Q) model and target bits allocation method from exhaustive experiments, and the intra QP was determined accurately. Extensive experimental results show that our algorithm not only obtains more accurate intra-frame output bitrate, more consistent subjective quality and more steady PSNR fluctuation compared with existing algorithm, but also prevents buffer overflow and frame skip effectively.

A novel saliency detection model based on curvelet transform

Visually salient object or region detection in images is an active research field in recent years. Inspired by that curvelets can provide multi-scale sparse representation of objects with edges and textures, in this paper, we propose a novel saliency detection model based on fast discrete curvelet transform (SDCT) to detect more compact salient objects in an image. First, fast discrete curvelet transform is used to acquire multi-scale representation of feature maps in CIELab color space. Then the feature maps are transformed to feature salient maps based on dissimilarity measure between patches in a global manner. Finally, the complementary feature salient maps at each scale and each color channel are merged linearly to obtain unitary saliency map. Experimental results on MSRA saliency benchmark database show that the proposed SDCT model outperforms the most state-of-the-art saliency detection models in spatial and frequency domain with higher overall performance, especially acquires more compact salient object and suppresses background saliency effectively, which is desirable for many computer vision applications.

Simple and effective image quality assessment based on edge enhanced mean square error

Simple and effective image quality assessment (IQA) method is very desirable in many image and video processing applications, such as coding, transmission, restoration and enhancement. Classic pixel absolute error based objective IQA metrics such as mean square error (MSE) and corresponding peak signal to noise ratio (PSNR) are widely used for various applications due to low computation and clear physical meanings, but have also been criticized for poorly correlated with subjective evaluation. Inspired by that human visual system (HVS) is more sensitive to image local edge distortion than flat or texture areas, in this paper, we propose a novel edge enhanced MSE (EE-MSE) to emphasize edge distortion effects on IQA. Experimental results on LIVE database release 2 show that the proposed EE-MSE IQA metric is competitive with state-of-the-art HVS-based IQA metrics, while has lower computational complexity and is more suitable for optimization task.