A general motion estimation algorithm using Fourier descriptors is developed, and results for a test sequence showing that the estimated 2-D motion parameters are very close to their values are given. A new image sequence coding scheme with combined motion compensation in which the above algorithm is implemented together with block matching is also proposed. A post-estimation technique is added to improve motion compensation efficiency. Mean/residue VQ is used further to encode the motion compensated error signals. A bit rate of 0.1 b/p is achieved for an 8 bit 256*256 moving sequence with good picture quality.>

We have modified the interpolation and quad-tree based still picture Recursive Binary Nesting algorithm to code moving sequences. Central to the efficiency of this is the mechanism by which the extent of the quad-tree decomposition is controlled. Termed block testing, this forms the subject of this paper.

The authors identify the main mechanism through which bits are consumed in the RBN (recursive binary nesting) coding scheme and determine ways in which this can be reduced. To this end, various forms of local motion-compensation are implemented and described, but the results are found to be disappointing. Different quantizers are then considered, but again the improvements were minimal. A scaled quantizer is introduced and shown to provide satisfactory performance. It is concluded that with the consequent increase in transmission efficiency for small block sample points, the use of RBN for quality coding of moving sequences at low channel rates is practicable.>

The authors describe an approach to the coding of moving sequences that combines image segmentation with adaptive thresholding based on a priori knowledge about the scene and motion-adaptive frame interpolation techniques. In videoconference pictures the face is an area of high movement, and it is also the most noticeable part of the picture. Using such knowledge about the scene, the location of the face in alternate frames is determined, and a set of thresholds is chosen that allows segmentation of this part of the image more accurately than of the background. To achieve the high compression ratios required for transmission of videoconference and videophone pictures over the existing networks, only alternate frames are coded and transmitted. At the receiver a skipped frame k is produced by motion-adaptive interpolation of corresponding blocks of frames k-1 and k+1.>

Methods of employing a combination of two recently reported simple transforms (High Correlation Transform and Low Correlation Transform) using block adaptive transform coding with spatial domain activity classification are examined. It is found that for small block size, (viz. 8×8), increased coding efficiency can be had by employing a combination of the two new transforms, particularly for high activity images. The performance of such systems for small block size is found to be close, and sometimes superior, to that of a more conventional block classification adaptive transform coder employing the more complex real number cosine transform.

In this paper, an information lossy preprocessing technique using subsampling is described. Notches and pinholes in the original image are first removed using a set of masks, after which single element runs are doubled to preserve connectivity. The image is then subsampled by taking alternate picture elements (pels) horizontally and vertically. Enlargement of the subsampled image is carried out at the receiver where replication, bilinear, and cubic B-spline interpolation techniques are investigated. The edges of the interpolated images are "smoothed out" by a set of restoration masks, producing a visually acceptable result. A formal subjective experiment is also conducted to determine the order of subjective preference of the resulting restored images. Finally, adaptive block/ location coding (ABLC), with respect to a numerical measure of complexity, is used to code the subsampled image. Compression ratios achieved for the CCITT documents range from 13.31:1 to 63.77:1, a reduction, on average, by about 40 percent when one-dimensional run-length coding using the modified Huffman code (MHC) is applied to the subsampled image and by about 73 percent when MHC is applied to the original data.