|M.Sc Student||Radovsky Olga|
|Subject||Deblocking of Block-Transform Compressed Images Using|
Blending Functions Interpolation
|Department||Department of Electrical Engineering||Supervisor||Mr. Moshe Israeli (Deceased)|
Block-transform codecs (BTC) are among the most common compression tools available for still images and video sequences. These codecs divide the image into disjoint blocks and apply a transform on each individual block as part of the encoding process. At low bitrates the images exhibit a visually annoying phenomenon, known as the blocking effect. Various techniques have been published that reduce the blocking effect, but they often introduce excessive blurring, ringing and in many cases they produce poor deblocking results at certain areas of the image. This thesis presents a novel postprocessing technique for reducing the blocking effect in images and video sequences, which can be applied to conventional BTC, without introducing additional information. A basic deblocking methodology is described, which is aimed at very low bitrate images, in particular images where all the blocks are almost homogeneous. An important aspect of our deblocking algorithm is the novel approach for blocking effect definition. In this thesis the blocking effect is classified into two types of artifacts: boundary artificial discontinuities and corner artificial discontinuities. The image is deblocked by subtracting from every block an interpolatory surface striving to cancel the above discontinuities. This surface is obtained by using geometric techniques of bivariate blending function interpolation. The basis functions, which are used by the blending interpolation are of predefined form for every block and obey certain constraints. Application of the basic algorithm to images compressed at higher bitrates produces further damaging details inside the blocks and smoothes large areas of the image. In order to prevent this behaviour, the basis functions are modified for every block by a predefined factor called a grade. Adaptive grades are based on two points of view: fidelity to the true statistical nature of real images and human perceptual characteristics. This scheme is referred to as adaptation by distortion grading (DGA). Finally, we suggest another deblocking algorithm which utilizes the basic approach. The deblocking consists of two stages: background luminance averaging of the blocky image and then application of the basic approach, with basis function adaptation, to the image which is the difference between the averaged image and the blocky image. We refer to this scheme as deblocking by background luminance averaging (BLA). In order to achieve better deblocking by keeping the deblocked image as close to the initial image as possible, a new family of adaptive quantization constraint set (AQCS) is applied.
Experimental results show that the proposed class of algorithms provides better performance as compared to several of the best known methods in both objective and subjective image quality and enable real-time applications due to its non-iterative property.