Arithmetic Coding Context Model That Accelerates Adaptation for Small Amounts of Data

The present invention relates to a method and device for arithmetically encoding and, more particularly, to a context model that accelerates adaptation for small amount of data while still maintaining high compression ratios for larger amounts of data.

Data compression systems are known in the prior art that encode a stream of digital data signals into compressed digital code signals and decode the compressed digital code signals back into the original data. Data compression refers to any process that attempts to convert data in a given format into an alternative format requiring less space than the original. The objective of data compression systems is to effect a savings in the amount of storage required to hold or the amount of time required to transmit a given body of digital information.

To be of practical utility, a general purpose digital data compression system should satisfy certain criteria. The system should have reciprocity. In order for a data compression system to possess the property of reciprocity it must be possible to re-expand or decode the compressed data back into its original form without any alteration or loss of information. The decoded and original data must be identical and indistinguishable with respect to each other. The property of reciprocity is synonymous to that of strict noiselessness used in information theory. Some applications do not require strict adherence to the property of reciprocity. One such application in particular is when dealing with graphical data. Because the human eye is not that sensitive to noise, some alteration or loss of information during the compression de-compression process is acceptable.

The system should provide sufficient performance with respect to he data rates provided by and accepted by the devices with which the data compression and de-compression systems are communicating. The rate at which data can be compressed is determined by the input data processing rate into the compression system, typically in millions of bytes per second (megabytes/sec). Sufficient performance is necessary to maintain the data rates achieved in present day disk, tape and communication systems which rates typically exceed one megabyte/sec. Thus, the data compression and de-compression system must have enough data bandwidth so as to not adversely affect the overall system. The performance of data compression and de-compression systems is typically limited by the computations necessary to compress and de-compress and the speed of the system components such as, random access memory (RAM), and the like, utilized to store statistical data and guide the compression and de-compression process. Performance for a compression device is characterized by the number of processor cycles required per input character under the compressor. The fewer the number of cycles, the higher the performance.

Another important criteria in the design of data compression and de-compression systems is compression effectiveness, which is characterized by the compression ratio. The compression ratio is the ratio of data size in uncompressed form divided by the size in compressed form. In order for data to be compressible, the data must contain redundancy. Compression effectiveness is determined by how effectively the compression procedure uses the redundancy in the input data. In typical computer stored data, redundancy occurs both in the nonuniform usage of individual symbology, example digits, bytes, or characters, and in frequent recurrence of symbol sequences, such as common words, blank record fields and the like.

General purpose data compression procedures are also known in the prior art, three relevant procedures being the Huffman method, the Tunstall method and the Lempel-Ziv method. The Huffman method is widely known and used, reference thereto in article of D. A. Huffman entitled "A Method For Construction Of Minimum Redundancy Codes", Proceedings IRE, 40, 10 pages 1098-1100 (September 1952). Reference to the Tunstall algorithm may be found in Doctoral thesis of B. P. Tunstall entitled "Synthesis of Noiseless Compression Codes", Georgia Institute of Technology (September 1967). Reference may be had to the Lempel-Ziv procedure in a paper authored by J. Ziv and A. Lempel entitled "A Universal Algorithm For Sequential Data Compression", IEEE Transactions on Information Theory, IT-23, 3, pages 337-343 (May, 1977).

One of the first general purpose data compression procedures developed is the Huffman method. Briefly described, the Huffman procedure maps full length segments of symbols into variable length words. The Huffman data compression procedure suffers from two limitations. Firstly, the Huffman procedure operates under the constraint that the input data to be compressed be parsed into fixed length segments of symbols. Although the Huffman procedure provides the best compression ratio that can be obtained under these constraints, when the constraint is relaxed it is possible to obtain significantly better compression ratios by utilizing other procedures. Secondly, Huffman coding requires full knowledge of the statistical characteristic of the source data. The Huffman procedure operates under the assumption that the probability with which each fixed length input segment occurs is known. This requirement of the Huffman procedure can in practice, be satisfied by the use of an adaptive version of the procedure which accumulates the necessary statistics during processing of the data. This, however, is cumbersome, and requires considerable working memory space and performs sub-optimally during adaptation.

The Tunstall algorithm, which maps variable length segments of symbols into fixed length binary words, is complimentary to the Huffman procedure with the fixed length constraints now applied to the output segments instead of the input segments. Like the Huffman procedure, the Tunstall procedure requires a foreknowledge of the source data probabilities. Again this foreknowledge requirement can be satisfied to some degree by utilizing an adaptive version which accumulates the statistics during processing of the data.

The Lempel-Ziv procedure maps variable length segments of the symbols into variable length binary words. It is asymptotically optimal when there are no constraints on the input or output segments. In this procedure the input data string is parsed into adaptively grown segments, each segment consisting of an exact copy of an earlier portion of the input string suffixed by one new symbol from the input data. The copy which is to be made is the longest possible and is not constrained to coincide with any earlier parsed segment. The code word which replaces the segment in the output contains information consisting of a pointer to where the earlier copied portion begins, the length of the copy, and the new symbol.

It would appear that Huffman or Shannon-Fano coding is the perfect means of compressing data. However, this is not the case. As mentioned above, these coding methods are optimal when and only when the symbol probabilities are integral powers of 1/2, which is usually not the case.

The technique of arithmetic coding does not have this restriction: It achieves the same effect as treating the message as one single unit (a technique which would, for Huffman coding, require enumeration of every single possible message), and thus attains the theoretical entropy bound to compression efficiency for any source.

In arithmetic coding, one decision after another is encoded to define successfully smaller, lesser included intervals along a number line. Additional information on arithmetic coding can be found in "An Introduction To Arithmetic Encoding"; by G. G. Langdon, Jr., IBM Journal of Research and Development, Vol. 28, n. 2, March 1984, 135-149; and "Arithmetic Compression Code Control Parameters Approximation"; by D. R. Helman, G. G. Langdon, Jr., and J. J. Rissanen, Vol. 23, n. 11, April 1981, 5112-5114 and U.S. Pat. No. 4,905,297, "Arithmetic Coding Encoder And Decoder System", Langdon, Jr. et al. all incorporated herein by reference.

As noted in the above articles, arithmetic coding provides that each decision has a plurality of possible exclusive outcomes "or events". Each outcome or event is represented in data by a symbol. In the imaging environment, for example, each decision may correspond to whether or not a given pixel is black. The decision outcome being represented by a Y (or YES) symbol if the pixel is black or an N (or NO) symbol if the pixel is not black. Thus, a plurality of decisions may then be represented by a sequence of symbols, e.g., YNNY . . . .

In accordance with prior arithmetic coding teachings, a probability line has a current interval defined thereon. The first current interval is 0 to 1. The current interval is divided into segments in which segment corresponds to one possible outcome for the next decision. Where there are only two possible outcomes for each decision, the current interval is divided into two segments. The length of each segment is based on its respective associated probability. The respective probabilities may remain fixed or may adapt as decision data is entered.

It is the correlating of large segments to symbols which occur with greater frequency that leads to the compression effect. In the former cited article ("An Introduction To Arithmetic Encoding"), a 4-symbol arithmetic coding example is set forth in which each decision can result in an "a" event (having a 50% probability), a "b" event (having a 25% probability), a "c" event (having a 12.5% probability), or a "d" event (having a 12.5% probability). Representing the four events in binary form would require two bits for each decision, where the events would be represented respectively by 00, 01, 10, 11. For three decisions such as "aab" which is highly likely, the straight forward encoding data would be 00 00 01; requiring six bits. However, as observed in the article at page 137, the arithmetic coding approach permits the sequence "aab" to be represented by the value 0.001. Instead of six bits, the information can be represented in three bits. This conservation of bits results as successive events having relatively high associated probabilities occur.

The conservation deteriorates if numerous events occur for which there are low probabilities and relatively short line segments. With the above noted probabilities, a sequence of events "dd" would be represented with encoded data as 11 11 whereas, by arithmetic coding, the "dd" events would be represented by 111111. Provided that the larger segments in fact correspond to events that occur with correspondingly greater frequency, the additional bits needed for less probable symbols are outweighed by the conservation achieved when more probable symbols occur.

Arithmetic coding adapts to data throughout the entire compression run and never forgets the past. This is counter to many LZ based schemes which are constantly losing the contents of their dictionaries. LZ rebuilds the dictionary, thus adapting to the next section of data. For LZ schemes, a 1 KB of data compresses just as well as a 100 KB of data. Arithmetic coding continues to improve it's probabilities throughout the run, which means that it will not do as well on 1 KB of data because it has not had the chance to optimize its adaptations. However, arithmetic coding compresses the 1 KB better than LZ.

Arithmetic coding uses a probability table to store statistics on the image. As each bit is compressed, the table is accessed to determine how the bit should be handled. The larger the table, the slower it moves to a final/optimal state. The larger the table, however, the better the compression ratios for larger images because more information is available for each bit. Simple experiments have shown the break point for benefiting from a larger table to be around 10 KB--after that point, larger tables yield a significantly better compression ratio.

In the operation of the printers, several different types of images are created. Some images are 100 KB to 200 KB in size, other images may be as small as 400 B or even smaller. A major contributor these small images is a font cache. Each unique character is first created and stored in the font cache before printing begins. Depending on the page, the amount of space used by the font cache may be an issue for successful rendering of the page.

Font usage becomes more critical if the same printer is used in a LAN environment. Several different users can send print jobs with their own preference for fonts and point sizes. With each new job, the printer determines if the requested character set already resides in the font cache from previous jobs. If it does not exist, then the time is taken to render the character. As the printer's memory is populated with font cache characters, there comes a point where those characters not immediately in demand must be removed from the font cache to make room for the other characters. Thus, the longer font cache characters stay around--especially across several jobs--the faster users receive their printouts. It is not uncommon for the wait for the font cache creation to be a noticeable amount of time. This is most often noticed when printing the first page after a power-on.

In order to accomplish the present invention there is provided a method for compressing and decompressing data. The method is accomplished by selecting a context model based on a count of an amount of the data processed. An Arithmetic compressor compresses, or decompresses, a portion of the data using the context model. A counter counts the amount of the data processed. The counter is incremented after the step of compressing. A first context model is selected if the count is less than a predefined amount and a second context model if the amount is more than the predefined amount. Any number of context models may be defined and appropriate decision points chosen.

There is also provided an apparatus for compressing, or decompressing data, the apparatus is made from an arithmetic compressor, where the arithmetic compressor includes a probability table, a first context model, and a second context model. There is a shift register connected to the arithmetic compressor. The shift register receives the data. A counter is connected to the arithmetic compressor. The counter signals the arithmetic compressor to use the first context model to compress a portion of the data when the portion is less than predefined amount, in the alternative, the arithmetic compressor uses the second context model to compress the portion.

A better understanding of the invention may be had from the consideration of the following detailed description taken in conjunction with the accompanying drawings in which:

FIG. 1 is a block diagram of an arithmetic compressor.

FIG. 2 graphically represent the growing size of the probability table used.

FIG. 3 shows how a given context model can be used to dynamically adjust the sized of the probability table used.

FIG. 4 is a flow diagram showing the logical operation of the preferred embodiment.

FIG. 5 is a block diagram of an arithmetic compressor in accordance with the present invention.

The present invention is not limited to a specific embodiment illustrated herein. Referring to FIG. 1 a block diagram of a hardware implementation of the preferred embodiment of the present invention is shown. As arithmetic compressor 1116 attempts to compress bit 101, data from image/shift register 1107 is passed to context model 1115. The context model 1115 maps the data from the image/shift register 1107 into the probability table 1113. The compressor, along with the probability table 1113 and context model 1115 compresses bit 101. The compressed data is then written out, generally to a memory device not shown. Bit 101 is shifted into the image/shift register 1107 and a new bit from the image is shifted into 101.

Decompression is performed in generally the same manor as compression. However, during decompression, the arithmetic compressor reads in the compressed data, using the probability table 1113 and context model 1115 decompresses and stores bit 101. As before, once bit 101 is decompressed the data in image/shift register 1107 is shifted left. Data exiting image/shift register 1107 is stored, generally, in a memory device.

The compressor 1114 uses the probability table 1113 to store statistics on the image as it is compressed or decompressed. As each bit is compressed/decompressed, the probability table 1113 is accessed to determine how the bit should be handled. The larger the probability table 1113, the slower the final/optimal state is reached. However, the larger the probability table 1113 the better the compression ratios will be for larger images because more information is available for bit 101.

Context model 1115 develops the index into probability table 1113 by looking at bits of data surrounding the bit being compressed. The important aspect to the indexing is the addressed location contains useful information concerning the value of the bit to be encoded/decoded. More specifically, the bits used to index the probability table should be able to provide significant information so that a reliable prediction of the bit to be encoded/decoded can be made. The more reliable the prediction, the better the compressibility of the image.

The present invention rapidly tracks the statistics at the beginning of the compression run, slowing down to a traditional pace as the image increases in size. This is accomplished by limiting the size of the probability table at the start of each image. As more data passes through the compressor, the size of the table is expanded. The size of the probability is controlled by gradually opening the context window for the first 5.1K Bytes of the image.

Referring to FIG. 3, the two dimensional context model 300 expands as more data is compressed. Each pixel is connected to an address line of the probability table. The numbers on the pixels indicate which particular address bit is controlled by that pixel. Referring to FIG. 3a, for the first 100 bytes of data, address bits 9, 8, 7, 6, 1 and 0 are forced to zero, which effectively reduces the size of the probability table as shown in FIG. 2a. The next 1K bytes of data, address bits 9, 8 and 0 are forced to zero as shown in FIG. 3b, effectively reducing the size of the probability table as shown in FIG. 2b. For the next 4K bytes of data, address bit 0 is forced to zero as shown in FIG. 3c, effectively reducing the size of the probability table as shown in FIG. 2c. The context model of FIG. 3d, and the entire probability table as shown in FIG. 2d, is then used to compress the remaining data.

Two aspects are important in this initial context model of FIG. 3a. First, those pixels around bit 301 should be used over those pixels further away. Thus, address bits 5, 4, 3, and 2 are used. Second, in almost all images, zeros are more common than ones. So, forcing the masked bits to a zero allows this aspect to benefit the compression ratios after the mask is lifted. In the next block (FIG. 3b) the pixels closer to bit 301 are unmasked. This process continues until the final stage (FIG. 3d) where none of the address bits are masked.

The block diagram of FIG. 1 can be modified to accomplish the present invention by the addition of a counter. As shown in FIG. 5, counter 1102 counts the number of bits processed. The output of counter 1102 signals the context model 1115 which bits to mask. As additional bits are processed, the context model 1115 is instructed to mask fewer bits as described above.

A flowchart of the preferred embodiment is shown in FIG. 4. As each new file or image is processed, the counter is first reset 401. The bit to be compressed is retrieved from the data 403. If count is less than X 405, then the small context model 407 is used to compress the bit 419. In a similar manner, if the count is greater than X and less than Y 409, then the medium context model 411 is used to compress the bit 419 and if the count is greater than Y and less than Z 413, then the large context model 415 is used to compress the bit 419. Once more than Z bytes have been processed, the full context model 417 is used to compress the bit 419. The compressed data is stored 421. The count is incremented 423 and a check is performed to determine if all the data has been processed. Before processing the next bit, the data in the shift register must be shifted 426. One skilled in the art will understand that the order of operations given here may be varied while still implementing the underlying idea. Also, while the above describes small, medium, large and full context models, any number of context models may be defined and appropriate values for X, Y, Z, . . . chosen.

The present invention provides improved compression ratios for font cache data (i.e., small files) with only a minor impact of the compression ratio of larger images. In the particular case for Latin characters, the compression of font cache characters improved by 40%. That is to say that 40% more compressed characters will fit in the font cache as compared to the traditional arithmetic coding compression techniques. The complexity for masking the address bits is very small with only a noticeable amount of logic added to the design.

Although the preferred embodiment of the invention has been illustrated, and that form described, it is readily apparent to those skilled in the art that various modifications may be made therein without departing from the spirit of the invention or from the scope of the appended claims.