Transmitting Television as Clusters of Frame-to-Frame Differences

Abstract

A number of redundancy reduction techniques are used in a coder that is about eight times more efficient than simple PCM. The coder is capable of transmitting Picturephone® signals at an average rate of one bit per picture‐element (2 Megabits per second). When there is movement in the scene, most transmission time is devoted to the parts of the picture that change significantly. The data are generated irregularly but the data flow is smoothed prior to transmission in a buffer that holds about one frame of data. The redundancy reduction techniques used and the behavior of the coder are discussed both from an intuitive and from a statistical viewpoint. The positions of elements that change are signaled by addressing the first element of a run of changes and marking the end of the run with a special code word. The changes of luminance are transmitted as frame‐to‐frame differences using variable‐length code words. When rapid motion makes the buffer more than a quarter full, only differences for every second element are transmitted, the values of the intervening changed elements being set equal to the average of their neighbors. If the buffer continues to fill, the threshold that determines which changes are significant is raised from 4/256 to 7/256 of the maximum signal value. When violent motion causes the buffer to fill completely, replenishment is stopped for about one frame while the buffer empties. Subsampling and raising the threshold are not objectionable because viewers rarely detect the small impairments introduced in moving images. Observers are critical, however, of small impairments in stationary scenes. Thus, to maintain high quality in stationary areas, the entire picture is forcibly updated every three seconds by transmitting 8‐bit luminance values for three lines of every frame. A record of the coder's behavior is available as a 16‐millimeter movie film.

Interframe Adaptive Data Compression Techniques for Images.

“…An inherent assumption here is that the illumination from one frame to the next remains unchanged. Generally, the moving pixels occur in clusters [13], and constitute the so called moving areas. In [52], prediction is based on the previous pixel of the same scan line and the technique is essentially an intraframe one dimensional DPCM.…”

“…In [13], the conditional replenishment method of [48] has been improved, and some of the techniques of [441,together with some other adaptations, have been used to result in a more efficient coder wlt increased complexity. We have simulated this technique for comparison purposes and a brief description is given in chapter III.…”

“…Most of the coders described in [13,24,42,44] are designed for a data rate of 0.25 -1.0 bits/pixel and for a signal with 1 MHz bandwidth, the data rate is .5 -2 Mbits/sec. In [261 a very low bit rate coder, .1 bit/pixel,or 0.2 Mbits/sec., has been described which reproduces the stationary areas quite well, but scenes containing moderate and large motions are visibly smeared and blurred.…”

“…In [261 a very low bit rate coder, .1 bit/pixel,or 0.2 Mbits/sec., has been described which reproduces the stationary areas quite well, but scenes containing moderate and large motions are visibly smeared and blurred. It combines cluster coding of [13], a higher order prediction given by the line-to-line difference of the frame difference signal, subsampling in spatial and temporal directions as needed, temporal filtering, etc., to achieve this low data rate. Low pass temporal filtering of the signal is done to reduce the entropy of the prediction error.…”

“…Here a pixel by pixel translational motion is recursively estimated and the interframe prediction is based on the estimated motion-displaced-location in the previous field. The prediction error is cluster coded similar to [13] The fact that the multiframe data cannot be satisfactorily modeled by separable statistics (covariance), a model such as described in [67] could yield poor coding performance. But we do need the transform domain variances to design a coder without resorting to trial and error.…”