Perceptual Image Compression And Processing Based On Edge Modeling

The prevalence of mobile devices, such as smart phones, PDA, etc, changes our dailylives. People now can capturing，viewing and sharing photos via diverse ways. In spite ofthe conveniences that mobile devices give to us, they also bring new challenges to thecurrent wireless communication systems.Firstly, the wireless communication is limited by bandwidth, especially when multipleusers accessing the servers simultaneously. Each user can only get a narrow piece ofbandwidth and the images have to be compressed as much as possible. Thus there is anurgent need for efficient low bit rate coding techniques. Secondly, the ubiquity ofinteractive applications, such as electronic map and online shopping，puts more pressureto the current low bit rate codecs. In these scenarios, good user experience may costadditional bit budget thus the bit cost tradeoff should also be considered in the design oflow bit rate codec. Thirdly, the mobile devices are also strongly restricted by thethroughput, power, etc. The encoder complexity, instead of compression performance,should have the first priority in these scenarios. Lastly, the purpose for wireless imagecommunication is to serve humans. It is the human visual system which finally judges thequality of images. As been approved by a mount of perceptual experiments, theL2-normdistortion metric cannot reflect on the perceptual quality of images. Therefore, at low bitrate condition, the design of image compression algorithms should be in pursuit of higherperceptual quality instead of minimizing some-norm distortion metric.This paper is motivated by the above problems. Considering the importance of edges tothe human visual systems, we investigate the efficient modeling of edges and theedge-guided coding techniques. We also extend the edge models to solve the restorationof degraded images. We investigate three efficient edge models for different applications：1) piecewise auto regressive (PAR) model for image restoration,2) graph model forregion of interest (ROI) detection, and3) edge tree model for the adaptive scan andcoding of edge pixels. To be specific, our works consists of the following four parts.We propose an edge-based region-of-interest (ROI) image coding technique forlow bit-rate visual communication. An image is compactly coded into twosemantic levels: background sketch and object textures. The background sketch isa down-sampled version of the input image. The decoder reconstructs thebackground first by adaptive interpolation and lets users identify the ROI, andthen requests the ROI textures to be transmitted. A distinct advantage of theproposed ROI technique is a compact edge-based descriptor of natural object boundaries. The expensive ROI geometry computations are carried out at theencoder only on demand, keeping decoder complexity low to benefit wirelessdevices. The new system outperforms the dynamic ROI coding of JPEG2000inboth visual quality and Peak Signal-to-Noise Ratio. Furthermore, bothbackground and texture coding can be made compliant with any existingcompression standardWe develop a novel, psychovisually motivated, edge-based low bit-rate imagecodec. It offers a compact description of scale invariant second order statistics ofnatural images, the preservation of which is crucial to perceptual quality of codedimages. Although being edge based, the codec does not explicitly code the edgegeometry. To save bits on edge descriptions, a background layer of the image isfirst coded and transmitted, from which the decoder estimates the trajectories ofsignificant edges. The edge regions are then refined by a residual codingtechnique based on edge dilation and sequential scanning in the edge direction.Experimental results show that the new image coding technique outperformsexisting ones in both objective and perceptual quality, particularly at low bit rates.We propose a new model-based soft decoding technique to restore the widely usedJPEG streams. The image is modeled as a2D piecewise stationary autoregressiveprocess, and the decoding task is formulated as an optimization problem with theconstraint given by the quantization intervals which are freely available at thedecoder. The autoregressive model serves as an important regularization term ofthe objective function of the optimization. And the autoregressive modelparameters are solved on the decoded image locally using a weighted total leastsquare method, where a novel bilateral dualside weighting scheme is proposed tominimize the influence of the blocking artifact on the final estimation. Extensiveexperimental results suggest that the proposed algorithm systematically improvesthe quality of JPEG images and also outperforms existing JPEG postprocessingalgorithms in a wide bit-rate range both in terms of PSNR and subjective quality.We propose a new restoration technique to interpolate degraded images.Resolution upconversion of a degraded image is an ill-posed inverse problem thatis even harder than video superresolution due to the lack of redundantobservations from reference frames. To overcome this difficulty an adaptive2DPAR model is used to strengthen the constraints on the solution of the inverseproblem. The PAR model can be fit to local image waveforms by adjusting itsparameters. But estimating the model parameters needs the knowledge of the very original high-resolution pixels to be estimated by the model. We resolve thischicken-and-egg dilemma by adaptive nonlinear least-squares joint estimation ofboth model parameters and original pixels. This non-linear estimation problem issolved by the structured total least-square method, constrained by the degradationfunction (e.g., the point spread function of a camera plus noises) that forms theobserved low-resolution image. As such, this work offers a unified generalframework for joint upsampling, deconvolution and denoising. Moreover, theupsampling can be carried out at an arbitrary scale rather than power of two.Experiments show that the proposed technique outperforms current methods inboth PSNR and subjective visual quality, and its advantage becomes greater forlarger scaling factors.