Images generally have features of local smoothness and global similarity. Their redundancies can be eliminated or reduced by segmenting an image into variable size blocks and by coding them depending on pixel features in each block. This report considers on the lossless image compression method using variable block size segmentation recently proposed by N.Ranganathan, et al., and proposes an improved method of it. Our idea focuses on a reference pixel for a block being coded to get a shorter codeword and a coding method of a 2×2 pixel block which occurs frequently to get better compression efficiency.

View full abstract

Continuous-tone images are generally composed of areas where pixel values change smoothly and areas, such as edges, where pixel values change dynamically. Until now, the prediction coding methods have used one method or an algorithm to predict the pixel value to be coded, even though the features of images change locally. To make the prediction difference to be smaller, this report proposes a new approach to select a prediction method from two methods adaptively. The features of this approach are : selection of a prediction method using a 2-dimensional pattern which shows which method caused the minimum difference at the neighboring pixels, and the parallel calculations of prediction differences at the pixel to be coded.

View full abstract

We propose a video archiver using frame sorting method as a coding method for archive motion pictures and investigate a basic concept by computer simulation. Frame sorting method is a way to sort similar frames by estimation function like frame correlation coefficient for changing the order of frames temporally. By coding images sorted by this method, it is possible to archive many sources of images more effectively because of the reduction of redundancy than before. As the results of computer simulations, effective archiving is realized compared with a conventional method.

View full abstract

Although MC (motion compensation) methods based on arbitrarily-shaped regions can reduce interference such as blocking-effects, these methods generally require a large number of bits to encode region-shapes. To cope with this problem, we propose a new MC method using Voronoi diagrams and apply it to video coding. Since a Voronoi diagram is reconstructed by only positions of points called generators, region-shapes can be encoded at a low coding rate. First, the Voronoi-based MC is carried out, then the conventional square-block-based DCT coding of MC error signals is conducted. Simulation results indicate that this scheme can produce images of higher quality than the conventional scheme, which uses square-block-based MC and DCT, from the viewpoints of SNR as well as subjective evaluation.

View full abstract

This paper proposes a new motion compensation method. First, the current frame is partitioned into varibable-size blocks in the three steps according to the local energy of warping prediction errors. In order to overcome unnatural distortion due to warping prediction, block matching is then selectively applied to the small blocks where discontinuous motion is detected. Finally the method transfigures these blocks into variable-shape blocks so that their boundaries can run parallel to those of moving objects. Simulation results indicate that the method provides more than 2 dB additional improvement in SNR of predicted images over the conventional block matching and produces natural images with less interference.

View full abstract

In this summer, SIGGRAPH'96,which is the festival of Computer Graphics, digital image processing and interactive technology, was held in New Orleans, Louisiana USA. In this report, I show you the outline of this meeting. SIGGRAPH'96 has more kinds of event than former SIGGRAPH. They are electronic theater, paper session, panel session, art/designer sketches, exhibition, animator sketches, digital bayou, behind the scenes, applications and etc.

View full abstract

In the case of pattern recognition using neural networks, it is very difficult for researchers or users to design them. In this paper, a method of learning and designing feedforward neural networks is discussed. In the proposed method, a neural network is regarded as one individual and neural networks whose structures are same as one species. These networks are evaluated by grade of training and they evolve according to an evolution rule proposed in this paper. The designing and training neural networks which performs Handwritten KATAKANA recognitions are described and the efficiency of proposed method is discussed.

View full abstract

Image morphing is an active fleld of research and recent efforts aim at improving both warping functions and user interface. To specify feature points in two images, the user interface takes up much time and it allows the warping speciflcation by the user represents very tedious work. Therefore, we propose a new semi-automatic algorithm based on snake model to specify the features in an image. It allows a user to extract a contour that defines features such as lips, mouth, profile, etc., by only specifying endpoints of the contour. The algorithm uses these two points as anchor points for the images, and then generates feature points of the image using these anchor points. The proposeal algorithm helps the user to easily define the exact position of a feature. It may also reduce the time in establishing the feature correspondence between two images.

View full abstract

The rendering time of special effects is very important problem for non-linear editing systems that edit the digitized image in the harddisk. This paper report a result of experiment of antialiasing of superimposed character. We picked up two cases through a study of several combinations of methods. Then, we did a series of experiments for these two cases. We tested rendering time, composite image quality and usage conditions. We also report what we decided be the best method in the development of our video editing system.

View full abstract