Adaptive Probability Estimator for Entropy Encoding/Decoding

This invention relates to the coding of signals and, more particularly, to a probability estimator for entropy encoding/decoding.

It is known that entropy, e.g., arithmetic, encoding and decoding requires a probability estimate of the symbol to be encoded and subsequently decoded. In arithmetic encoding and decoding, more accurate probability estimates result in greater data compression. To this end, it is desirable that the probability estimates adapt to changing underlying symbol probabilities.

Prior known probability estimator arrangements have included some ability to adapt but have been limited in the adaptation rate used because of the need to estimate symbol probabilities with relatively small values. Indeed, the effective adaptation rate in prior arrangements was constant and was independent of the actual values of the probabilities being estimated. This leads to less accurate probability estimates and, consequently, lower efficiency in the encoding and decoding of the symbols.

The problems and other limitations of prior known probability estimators are overcome, in accordance with an aspect of the invention, by optimizing the rate of adaptation to the estimated probabilities of symbols to be encoded and/or decoded.

More specifically, if the values of the probabilities being estimated are not small a "fast" adaptation rate is realized in generating them and if the values of the probabilities being estimated are small a necessarily slower adaptation rate is realized in generating them.

In a specific embodiment, the adaptation rate is optimized by ideally matching it to the actual probability value being estimated. In particular, the adaptation rate is optimized to be proportional to the inverse of the smallest value probability being estimated. This is achieved, in one example, by first determining whether an at least first characteristic of a set of prescribed parameters meets a prescribed criterion, namely, whether it exceeds an at least first threshold value and if the at least first characteristic exceeds the at least first threshold value, adjusting the set of prescribed parameters in a prescribed manner.

In an exemplary embodiment, the at least first prescribed characteristic is the minimum value of the set of prescribed parameters for a given context and the at least first threshold value is a small value, for example, eight (8). Each element in the prescribed set of parameters is a function of a context sensitive accumulation, i.e., count, of received symbols.

In the Drawing:

FIG. 1 shows details of an arrangement employing an encoder and remote decoder employing aspects of the invention; and

FIG. 2 depicts a flow chart illustrating the operation of elements of the adaptive probability estimator employed in the encoder and decoder shown in FIG. 1.

FIG. 1 shows details of entropy encoder 101 in simplified block diagram form, including aspects of the invention, which receives data symbols s(k), encodes them into a data stream a(i) and interfaces them to a transmission media 102 for transmission to remote entropy decoder 103. Entropy decoder 103, also including aspects of the invention, interfaces to the transmission media to obtain the receiver data stream and decodes it into replicas of the transmitted symbols s(k). Symbols s(k) include elements [0, . . . , S-1], namely, s(k) .epsilon.[0, . . . , S-1]. Thus, the symbols may be multilevel or binary as desired.

Accordingly, encoder 101 includes, in this example, arithmetic encoder unit 104, context extractor 105, adaptive probability estimator 106 and line interface 107. Symbols s(k) and probability estimates p (k) are supplied to arithmetic encoder unit 104 and employed therein in known fashion to generate an encoded data stream a(i). Such arithmetic encoder units are known in the art. See, for example, an article entitled "Compression of Black-White Image with Arithmetic Coding", IEEE Transactions On Communications, VOL. COM.-29, No. 6, June 1981, pages 858-867, and U.S. Pat. No. 4,633,490 issued Dec. 30, 1986 for arithmetic encoders/decoders used to encode and decode symbols having binary elements. Also see an article entitled, "Arithmetic Coding For Data Compression", Communications of the ACM, Volume 30, No. 6, June 1987, pages 520-540, for an arithmetic encoder/decoder used to encode/decode symbols having multilevel elements. Line interface 107 interfaces the encoded data stream a(i) to transmission media 102 which, in turn, supplies the data stream to remote decoder 103. To this end, line interface 107 includes appropriate apparatus for formatting the data stream into the signal format employed in transmission media 102. Some well known examples of possible transmission media 102 are T-carrier trunks, ISDN basic subscriber lines, local area networks and the like. Such line interface apparatus is known in the art.

Context extractor 105 simply obtains the context c(k), where c(k) .epsilon. [0, . . . , C-1], of received symbol s(k). That is to say, context extractor 106 generates a unique context (or state) for symbol s(k) based on prior supplied symbols. By way of example, and not to be construed as limiting the scope of the invention, for an image compression system, symbol s(k) is representative of the color of a current pixel to be encoded and the context c(k) may be determined by the colors of prescribed prior pixels. For example, the color of a pixel (P) adjacent and prior to the current pixel in the same line and the color of a pixel (A) in a prior line directly above the current pixel may advantageously be used to generate a context c(k) for symbol s(k) in a binary application. Thus, c(k) is zero (0) if both pixel P and pixel A are white; c(k) is one (1) if pixel P is white and pixel (A) is black; c(k) is two (2) if pixel P is black and pixel A is white; and c(k) is three (3) if both pixels P and A are black. Also, see the U.S. Pat. No. 4,633,490 for another context extractor (state generator) which may be employed in a binary application. It will be apparent to those skilled in the art how such binary context extractors can be extended to obtain the context for multi-level applications. A representation of the extracted context c(k) is supplied to adaptive probability estimator 106.

Adaptive probability estimator 106 is advantageously employed to generate probability estimates p (k)=(p.sub.0 (k), . . . p.sub.s-1 (k)) for incoming symbol s(k) .epsilon. [0, . . . , S-1] and associated context c(k) .epsilon. [0, . . . , C-1]. To this end, adaptive probability estimator 106 maintains an array {n.sub.s,c } having dimensionality S by C, where each element n.sub.s,c of the array is an accumulation, i.e., a "count", of the occurrences of symbol s in context c, and s and c are dummy indices identifying the location of n.sub.s,c in the array. Adaptive probability estimator 106 can be readily implemented by appropriately programming a computer or digital signal processor. It is envisioned, however, that a superior mode of implementation is in a very large scale integrated (VLSI) circuit configuration on a semiconductor chip.

The flow chart shown in FIG. 2 depicts operation of elements in adaptive probability estimator 106 in generating more accurate probability estimates, in accordance with an aspect of the invention, by optimizing the rate of adaptation to the estimated probabilities of symbols to be encoded. Accordingly, operation of adaptive probability estimator 106 is started via start step 201. Thereafter, operational block 202 initializes k=0 and the counts of n.sub.s,c for all s .epsilon. [0, . . . , S-1] and c .epsilon. [0, . . . , C-1] to be n.sub.s,c =N.sub.s,c, where N.sub.s,c are some predetermined values. Operational block 203 obtains a new context c(k). It is noted that the new context can be the same as a previously obtained context. Then, operational block 204 obtains the sum Z of the counts for the obtained context c(k) for all s .epsilon. [0, . . . , S-1], namely ##EQU1## Operational block 205 causes adaptive probability estimator 106 (FIG. 1) to output the probability estimates which are, in turn, supplied to arithmetic encoder unit 104 (FIG. 1). Since, this is the first run these probability estimates are based only on the initialized conditions and the obtained context c(k). In subsequent runs, the probability estimates are based on the sum of the counts, i.e., accumulations, of the occurrences of symbols s(k) for context c(k). Thus, step 205 causes the probability estimates to be output, namely, ##EQU2## Operational block 206 obtains symbol s(k) to be encoded. Operational block 207 causes the count for the obtained symbol s(k) and context c(k) to be incremented by 1, namely, n.sub.s(k),c(k) is incremented by 1.

Operational block 208 obtains an at least first and an at least second characteristics of a prescibed set of parameters. In this example, each element of the prescibed set of parameters is a function of a context sensitive accumulation, i.e., count, of received symbols to be encoded. That is, the prescribed set of parameters are "accumulated" occurrences of the symbols s(k) for context c(k), namely, n.sub.0,c(k), . . . , n.sub.S-1,c(k). The at least first characteristic, in this example, is the minimum one of the accumulated occurrences for context c(k), namely,

The at least second characteristic, in this example, is the maximum one of the accumulated occurrences for context c(k), namely,

Conditional branch point 209 tests to determine, in accordance with an aspect of the invention, if either the at least first characteristic is equal to or greater than an at least first threshold value, namely,

or the at least second characteristic is equal to or greater than at at least second threshold value, namely,

It is important to note that the use of the at least first characteristic (MIN) allows, in accordance with an aspect of the invention, the optimization of the adaptation rate of adaptive probability estimator 106 (FIG. 1). In prior arrangements, only a maximum threshold value was employed. A significant problem with such a prior arrangement is that it is necessary to use either a large threshold value so that smaller value probabilities can be represented or a small threshold value to obtain fast adaptation. The small threshold value, however, makes it impossible to represent small value probabilities. Additionally, the large value threshold leads to a relatively slow adaptation rate. These problems are resolved by advantageously employing, in accordance with an aspect of the invention, the at least first characteristic which, in this example, is MIN as set forth in equation (4) and a small threshold value T.sub.1, which in this example, is eight (8). Thus, in this example, each of the possible symbol occurrences for context c(k), namely, [0, . . . , S-1], must occur at least eight times before the condition of equation (5) is met. Consequently, the use of the at least first characteristic, i.e., MIN, and the at least first threshold value T.sub.1 =8, yields an adaptation rate that is ideally matched to the actual probability value being estimated. By way of example and not to be construed as limiting the scope of the invention, for a binary application and a probability being estimated of one-half (1/2), the accumulated occurrences are adjusted after seeing the context c(k) approximately 8+8=16 times; for a probability being estimated of one quarter (1/4), the accumulated occurrences are adjusted after seeing the context c(k) approximately 8+24=32 times; and for a probability being estimated of one-eighth (1/8), the accumulated occurrences are adjusted after seeing context c(k) approximately 8+56=64 times. Thus, it is seen that the adaptation rate is faster for the larger (not small) probability values being estimated and is necessarily slower for the smaller probability values being estimated. The adaptation rate adjustment will be apparent from steps 209 and 210.

The at least second characteristic, in this example, MAX in accordance with equation (4), is employed in conjunction with the at least second threshold value T.sub.2 to assure against arithmetic overflow in the accumulation of the occurrences of symbols s(k) in context c(k). Unless one of the probabilities being estimated has an unusually small value, MAX will not be the characteristic that causes the parameter adjustment. In one example, the value of T.sub.2 is 2048. It is noted that other characteristics of the set of parameters may also be employed. For example, the sum Z obtained in step 204 could be used in place of MAX.

Thus, returning to step 209 if the prescribed criterion of either the condition of equation (5) (MIN.gtoreq.T.sub.1) or the condition of equation (6) (MAX.gtoreq.T.sub.2) is met, operational block 210 causes an adjustment in the accumulated symbol elements in context c(k). In this example, the adaptation rate adjustment is realized by step 210 in conjunction with step 209 causing a proportionate adjustment of the accumulated values, i.e., counts a so-called halving of the accumulated occurrences for context c(k) for all s .epsilon. [0, . . . , S-1], namely, setting

Although in this embodiment the counts are proportionately adjusted in the same manner when the condition of either equation (5) or equation (6) is met, it would be advantageous in some applications to adjust the counts differently for each of the above conditions. This adjustment proportionately of the accumulated occurrences makes the probability estimates more dependent on more recent occurrences of the symbols in context c(k). Thus, as implied above, by causing, in accordance with an aspect of the invention, the adjustment of the accumulated occurrences to occur in accordance with equation (5), i.e., MIN.gtoreq.T.sub.1, the adaptation rate is ideally matched to the actual probabilities being estimated. Again, the adjustment of the accumulated occurrences of symbols s(k) in context c(k) which occurs in response to equation (6), i.e., MAX.gtoreq.T.sub.2, is to protect against a possible arithmetic overflow condition in the rare situation when a very small probability value is being estimated.

Thereafter, conditional branch point 211 tests to determine if the symbol s(k) is the last symbol to be encoded/decoded. It is noted that the number of symbols to be encoded is typically known. If not known an indication of the number of symbols would be supplied to adaptive probability estimator 106. If the test result in step 211 is YES, the operation of the elements of adaptive probability estimator 106 is ended via END step 212. If the test result in step 211 is NO, control is returned to step 203 and appropriate ones of steps 203 through 211 are iterated until step 211 yields a YES result.

Returning to step 209, if the test result is NO, control is transferred to step 211 to determine if the symbol s(k) is the last symbol to be encoded (decoded). Again, if the test result in step 211 is YES, the operation of the elements of adaptive probability estimator 106 is ended via END step 212. If the test result in step 211 is NO, increment index k by 1 in step 213, control is returned to step 203 and appropriate ones of steps 203 through 211 are iterated until step 211 yields a YES result.

Decoder 103 includes, in this example, line interface 108, arithmetic decoder unit 109, context extractor 110 and adaptive probability estimator 111. Line interface 108 performs the inverse function of line interface 107 and deformats the incoming signal, in a known manner, to obtain the data stream a(i). Arithmetic decoder unit 109 performs the inverse function of arithmetic encoder unit 104. To this end, the received data stream a(i) and probability estimates p (k) from adaptive probability estimator 110 are supplied to arithmetic decoder unit 109 and used therein in known fashion to obtain the symbols s(k). Again, such arithmetic decoder units are known in the art. See again the article entitled "Compression of Black-White Image with Arithmetic Coding" and U.S. Pat. No. 4,633,490, cited above, regarding binary applications and the article entitled "Arithmetic Coding For Data Compression", also cited above, for multilevel applications. Context extractor 110 is identical to context extractor 105 in structure and operation and is not described again. Similarly, adaptive probability estimator 111 is identical to adaptive probability estimator 106 in structure and operation and is not described again.