KR100194205B1 - Improved Hoffman decoding method, decoding time table and its creation method - Google Patents

Abstract

The present invention relates to an improved decoding method of a Hoffman code, particularly comprising taking L bits of index bits from a Hoffman coded bitstream; Reading the root subtable of the decoding timetable; Searching for an item having the same index information as the taken index bit with reference to the read sub table; Determining the class of the item based on the attribute information of the searched item; Outputting the value of the item as a decoded value of a single 1-bit Hoffman codeword when determined to be Class A by the determination; Simultaneously outputting a value of an item as decoding values of a plurality of consecutive 1-bit Huffman codewords when determined to be a C class by the determination; Reading a new index bit from the bitstream according to the attribute information of the item if it is determined that the class is B by the determination; Reading a sub table corresponding to the value of the class B item and returning to the item search step; And outputting the decoded values of the A and C classes, shifting the index bits by the predetermined number of bits below the L by the attribute information to form the L bits index bits for the next item search, and then proceeding to the item search step. And returning.

Description

Improved Hoffman decoding method, decoding time table and its creation method

The present invention relates to a Hoffman code decoding method, and more particularly, to an improved Hoffman decoding method, a decoding time table, and a method for creating the table, which are efficient in terms of memory when executed in software and can perform decoding at high speed.

Huffman codes that can increase coding efficiency by allocating a small number of bits from a code having a high probability according to a frequency of occurrence, that is, a probability of occurrence, are widely used in various fields.

There are two types of decoding methods of the Hoffman code, a tree searching method and a table look-up method. In tree search, the Hoffman tree is stored in memory, and the decoder searches step by step from the root of the tree until it reaches a leaf. This method is effective in terms of memory capacity, but has a disadvantage in that the decoding speed is very slow.

In addition, the table lookup method stores a large table in memory so that any codeword can be decoded in one search. However, the table lookup method has a disadvantage of occupying a lot of memory because the table size becomes very large.

Many algorithms have been proposed for fast Hoffman decoding execution. However, most are more inappropriate for applications where software execution is not adequate and Hoffman tables are not fixed.

In 1993, a new Hoffman decoding algorithm was proposed for hardware execution, but it is similar to the tree search method, except that it does not shift one bit in one search, but shifts several bits in one search, that is, four bits. There is this.

However, the problem with this algorithm

1. It cannot be used to run the software because no automatic table creation method is provided.

2. Sometimes it is not fast enough and improvement is needed.

SUMMARY OF THE INVENTION An object of the present invention is to provide an improved Hoffman code decoding method capable of high-speed decoding and excellent memory efficiency in software execution in order to solve the problems of the prior art.

Another object of the present invention is to provide an improved decoding time table.

It is still another object of the present invention to provide a method for creating an improved decoding time table.

In order to achieve the above object, the decoding method of the present invention provides a decoding method using a multi-bit search method of Hoffman code, comprising: taking L bits of an index bit from a Hoffman coded bitstream; Reading the root subtable of the decoding timetable; Searching for an item having the same index information as the taken index bit with reference to the read sub table; Determining the class of the item based on the attribute information of the searched item; Outputting the value of the item as a decoded value of a single 1-bit Hoffman codeword when determined to be Class A by the determination; Simultaneously outputting a value of an item as decoding values of a plurality of consecutive 1-bit Huffman codewords when determined to be a C class by the determination; Reading a new index bit from the bitstream according to the attribute information of the item if it is determined that the class is B by the determination; Reading a sub table corresponding to a value of the class B item and returning to the item search step; And after outputting the decoding values of the A and C classes, shifting the index bits by the predetermined number of bits below the L by the attribute information to form the L bits index bits for the next item search, and then the item searching step. It characterized in that it comprises the step of returning to.

In addition, the decoding method of the present invention is a decoding method using a multi-bit search method of Hoffman code, comprising: taking L bits of an index bit from a Hoffman coded bitstream; Reading the root subtable of the decoding timetable; Searching for an item having the same index information as the taken index bit with reference to the read sub table; Determining the class of the item based on the attribute information of the searched item; Outputting the value of the item as a decoded value of a single 1-bit Hoffman codeword when determined to be Class A by the determination; Simultaneously outputting a value of an item as decoding values of a plurality of consecutive 1-bit Huffman codewords when determined to be a C class by the determination; Reading a new index bit from the bitstream according to the attribute information of the item if it is determined that the class is B by the determination; Recognizing a code not defined on the Hoffman code table and performing an exceptional operation if it is determined that the class is D; Reading a sub table corresponding to a value of the class B item and returning to the item search step; And after outputting the decoding values of the A and C classes, shifting the index bits by the predetermined number of bits below the L by the attribute information to form the L bits index bits for the next item search, and then the item searching step. It characterized in that it comprises the step of returning to.

An improved decoding time table of the present invention for achieving the above another object is a decoding time table for use in decoding by a multi-bit search method of a Hoffman code, wherein the decoding time table includes a plurality of Huffmans having a predetermined number of bits or less. A code field consisting of codes; An index field indicating an arrangement order of the plurality of codes; An attribute field having attribute information such as a class for classifying the Hoffman codes, the number of consecutive identical codewords, the number of shifts of index bits for search, or the number of index bits to be read; And a value field composed of decoding values corresponding to the respective Hoffman codes.

According to another aspect of the present invention, there is provided a method of creating a decoded time table in the multi-bit search method of a Hoffman code, comprising: first global data for storing an original Hoffman table and a decoding timetable. Defining second global data for storing data; Defining a level variable indicating a level of a sub table being created and an entry variable indicating a number of codewords registered in the Hoffman code table; Creating a root subtable according to a given table size while referring to the Hoffman code table; Checking the created root subtable to determine which subtable needs to be expanded; Creating a subtable by allocating a base address of the subtable and repeatedly performing the subtable creating step and the determining step when the subtable needs to be expanded; And returning if no expansion is necessary in the determining step.

1 is a table showing one of the original Hoffman code tables.

2 is a table showing an example of a root sub table of a decoding time table of a conventional multi-bit search method.

3 is a table showing an example of a sub table of a decoding time table of a conventional multi-bit search method.

4 is a table showing an example of the root sub table of the decoding time table of the improved multi-bit search method according to the present invention.

5 is a diagram for explaining a class of an item of a sub table according to the present invention;

6 is a flowchart for explaining an improved Hoffman decoding method according to the present invention.

7 is a flowchart for explaining a method of creating an improved decoding table according to the present invention.

Hereinafter, with reference to the accompanying drawings will be described in more detail the present invention.

First, the conventional multi-bit search method will be described in order to help understanding of the present invention.

Every Hoffman decoding algorithm has its own decoding time table which is convenient to use during decoding. That is, a Hoffman tree exists for tree search, and a large table exists for table lookup.

In the multibit search algorithm, there is a hierarchical table consisting of a series of subtables. These tables are formed in a tree-like structure.

The first subtree is a sub table located at the root of the tree. Therefore, the first subtree is referred to as the root subtable or table 0.

All subtables that extend directly from the entry of table 0 are at level 1 and are called table 1X (11, 12, etc.). Subtables that extend from entries in Table 1X exist at level 2 and are called Table 2X (21, 22, etc.). In this way, as the level increases, the subtables are formed hierarchically.

Here, each subtable has a size of 2n (n = L0, L1X, L2X, ...): L0 is the number of bits of the Hoffman bitstream used for search in table '0'. The same is true for L1x and L2x. These bits are called index bits). Each entry in a subtable is called an item. Therefore, if L0 = 4, the size of table 0 has 24, i.e. 16 items.

In the multibit search algorithm, there are two kinds of items, class A and class B. Class A items are referred to as leaf items. This is because it is at the top of the tree-like structure, and no subtable extends further from the leaf item. If a leaf item is encountered in the Huffman decoding process, decoding of the current codeword is terminated. Class B items are also called subroot items because they are not at the top of the tree-like structure, and there is one subtable or one subtree extending from it. If a Class B item encounters a bit stream, then decoding of the current codeword does not end, and more bits will be needed to trace the next subtable extended from that item.

Here, each item has two fields, one of the two fields is called an attribute field, and one bit of this attribute field is used to indicate an A or B class of the item. The other three bits of the attribute field in items of class A are used to indicate the number of index bits to be shifted out. In the class B item, all index bits are shifted out. These three bits are used to indicate how many index bits are read for the next search. The index bits (3 bits) of the attribute field will be shifted out, and these 3 bits are used to indicate the index bits to be read during the next search. In class A items, the other field is used to store the encoded value, and the other field of the class B item is used to store the address of the extended subtable.

For example, the Hoffman table structure used in the multi-bit search algorithm is described below.

Assume that the Hoffman code table has six codewords: 0, 10, 110, 1110, 11110, and 11111. The original Hoffman table is stored in an array structure by executing the following program.

struct table {char length; / * length of the codeword * /

(type) codeword; / * the codeword itself * /

(type) value; } / * the encoded value * /

The type of the codeword field and the encoded value depends on the application. Figure 1 shows the six codeword structures used to store the original Hoffman code table. 16-bit integers are used for codeword storage. The valid codeword bits are filled from the MSB of 16 bit integers. The sign type can be a char.

Each item in the decoding time table is stored by execution of the next program.

struct tab {char attribute;

(type) value; }

Here, the value field is an 8-bit char. The root subtable of the decoding time table created by the dl program is shown in FIG.

In FIG. 2, 'cod' is a codeword bit ('1111' is 4 bits from the MSB of Hoffman codewords '11111' and '11110'), and 'idx' is an index bit used to search within the table. 'Attr' is the item's attribute field, and 'val' field is the item's value field.

3 shows Table 11 of size 2 extended from Table 0. FIG.

Since the longest Hoffman codeword is 5-bit, we can take LO = 4, so Table 0 has 16 items. In the attribute field, three bits from the LSB are used to indicate how many index bits are shifted out after the codeword is decoded. The MSB can be used to indicate the type of this item, where '0' represents a leaf item (Class A item) and '1' represents a sub-root item (Class B item). In the last item of table 0, the signed value field is 16, because table 0 and table 11 are stored as one large array, and the reference address of table 11 has offset 16 from the beginning of the entire table.

Each of the four bits of codewords from the MSB has a leaf item corresponding to the sixteenth item of table zero. 11110 and 11111 belong to a sub-table extended from item 1111 of table 0. The size of table 11 is two.

The first 8 items in table 0 all match the same codeword '0', and their index bits are 0, 1, 10, 100, 101, 110, 111. Since LO = 4, four bits will be read at the beginning of Huffman decoding, and these four bits will be used as index bits for searching Table 0. If the first bit is 0, the initial codeword is 0 regardless of the other 3 bits, so all 8 items are leaf items and all correspond to the same codeword 0. Therefore, after these codewords are decoded, it can be seen that the number of index bits shifted out by the attribute field is one. The first bit is then shifted out and another one bit is read to form a four bit index for decoding the next codeword.

For example, assuming that the bit stream is 01111010 ......, decoding the first codeword by searching for item 111 in table 0 and obtaining the sign value '0' by the first four bits being read and the four bits read. Ends. Since the attribute field in item 111 is '1', if only the first '0' of the bitstream is shifted out, 1111 is obtained. Searching on table 0 using 1111 as an index results in a search for the subroot item 1111. The decoding of this codeword cannot be terminated, thus shifting out all four bits. In the attribute field of Table 0, you can see that the number of bits read for the next search is '1'. One bit read is 0, the sixth bit of the bitstream. Use this as an index in Table 11 to search for the corresponding item to get the decoding value '-2'. Since this is a leaf item, the decoding of this codeword is also terminated. Then, one read bit is shifted out, another L0 (= 4) bits are read, and decoding of another codeword is started. Typically, all items belonging to sub-tables of size less than 16 are leaf items.

Each value of L0, L1x, L2x is called the length of the sub-table, and these values are typically small. Typical values of these are L0 = 4, L1x <4, L2x <4, ... Each sub-table with a length of 40,000 does not have a sub-table extending from it. If the length of the sub-table has a representative value, the size of the entire decoding time table is comparable to the Hoffman tree.

Multiple bit search algorithms have the advantage of being able to shift out more than one bit, typically four bits at a time. Therefore, although the operation required for each search is more complicated than tree search, it is faster than tree search. However, this is not always the case, and as is known, short codewords occur more often than long codewords in the Hoffman bit stream. In other words, when the average length is very small, the multiple bit search algorithm cannot be seen as a distinct advantage. However, if implemented in hardware, it is always faster than tree search, but sometimes slower than tree search in software.

For example, if the Hoffman table contains a 1-bit or 2-bit codeword (this codeword often appears in the bit stream), the performance of the multiple bit search algorithm is very poor. In the Hoffman bitstream, this case is poultry-prone, requiring improvement of this problem.

If there is a 1-bit or 2-bit codeword, there is a lot of redundancy in the first subtable. As described above, eight leaf items corresponding to one bit, '0000', '0001', '0010', '0011', '0100', '0101', '0110', and '0111' exist. do.

Therefore, the present invention intends to assign a C class item in a sub table. C class items are also referred to as multiple leaf items. In the Hoffman decoding procedure, encountering this kind of item allows multiple codewords to be decoded simultaneously. These codewords are all the same and appear more than once in succession. For example, 0, 1, 10, 11 would be C class items.

In class C items, the sign value field is the same as the sign value field of class A, but the attribute field is different. Two bits are used to indicate the form (currently there are more than one kind of item) and the other three bits represent the number of codewords. 4 represents 0, and 3 represents 1. The other three bits are used to indicate how many index bits can be shifted out. The one-bit codeword is equal to the number of codewords for the codeword and twice the number of codewords for the two-bit codeword.

After introducing the C class item in the decoding time table, it can be confirmed that the algorithm of the present invention is faster than the tree search method.

In theory, every Hoffman code table should be a complete tree structure, but sometimes for some unusual reason some Hoffman tables cannot be made a complete tree structure. In this case, we introduce another kind of item, namely a class D item, because non-existent riffs may also correspond to some items. Class D items are referred to as no match items. This means that the pattern of the bitstream does not match the Hoffman code table. If the decoding program encounters this kind of item, the decoding process stops, and an exceptional operation progresses.

When the modified decoding time table is applied to the above-described example, it appears as shown in FIG. 4.

In the attribute field of FIG. 4, the 3 bits of the LSB are used to indicate the number of index bits shifted out (class A and C class items) or to indicate the number of bits read for the next search ( B class items). As shown in Fig. 5, two bits from the MSB are used to indicate the class type, which is represented by 00--A class, 10--B class, 01--C class, and 11--D class. . The other three bits are used to indicate the number of codewords in the C class.

Table 11 does not change in this case.

Referring to FIG. 6, the decoding method according to the present invention takes L bit index bits from the Huffman coded bitstream (step 10) and reads the root subtable of the decoding time table (step 12). The item having the same index information as the taken index bit is searched for by referring to the read sub table (step 14). The attribute information of the searched item determines whether the item class is A class (step 18), B class (step 20), C class (step 22), or D class. If it is determined that the class A is determined by the determination, the value of the item is output as a decoding value of a single 1-bit Hoffman codeword (step 26). If it is determined that the class C is determined by the determination, the value of the item is simultaneously output as decoding values of a plurality of consecutive 1-bit Huffman codewords (step 28). If it is determined that the class is B, the new index bit is read from the bitstream according to the attribute information of the item (step 32). If it is determined that the class D is determined by the determination, the Huffman code table is recognized as an undefined code and an exceptional operation is performed (step 24). The sub table corresponding to the value of the class B item is read and the process returns to the item search step (step 34). After outputting the decoding values of the A and C classes, the index bits are shifted by a predetermined number of bits below the L by the attribute information to form L bit index bits for the next item search, and then the item searching step is performed. Return (step 30).

To explain the decoding operation by the modified multi-bit search method according to the present invention, a simple example will be described.

If the bitstream starts with '000101111100 ...', L0 = 4, so the first 4 bits are read, and this index is '0001', and referring to the attribute field of the corresponding index in table 0 of FIG. (Decimal) or '10 -111-011 '(binary). The MSB 2 bits indicate that it is a C class item, and the next 3 bits indicate that there are three codewords that appear continuously. Therefore, the corresponding decoding value '0' is read from the code value field, and three '0's are decoded and output. The last 3 bits indicate that 3 bits are shifted out. Accordingly, it can be seen that another 3 bits of the bitstream are read so that the current index becomes '1011'. Referring to the attribute field of the item corresponding to this index, it can be seen that the decoding value is '2', so this item is a leaf item, and two bits should be shifted out. Also, the corresponding code value of the decoding value field is '1'. Subsequently, two bits are shifted out so that a new two bits of the bitstream are read so that the index is '1111'. Since the item of the index is a sub-root item, the decoding process is performed by referring to Table 11 as described above.

The performance of the multi-bit search algorithm has been greatly improved by introducing C class items and D class items. However, the table building algorithm needs to be properly formed for software execution and for those cases where the Hoffman table is not fixed.

Cyclic procedures have been developed for the preparation of decoding time tables.

This algorithm will be used in the pseudo C language.

Referring to FIG. 7, first, two global data structures necessary for creating a table are defined (step 40). One of the two global data structures concerns the storage of the original Hoffman table.

struct table {char length; / * length of this codeword * /

(type) codeword; / * the codeword itself * /

(type) value; } / * the encoded value * /

The type of codeword field and sign value field depends on the application.

The other relates to the storage of the decoding time table.

struct tab {char attribute;

(type) value; }

Another two global variables are needed (step 42).

level: represents the level of the current subtable to be created, the level value of the root subtable is 0, and the level value of the subtable extended from the root subtable has one and the same value.

entry: Number of codewords in Hoffman code table:

The function is a Build-table (int root, int tab). Here, root defines the bitstream in which the subtable is extended. Tabl also defines the length of the subtable.

Each time this function is executed, a subtable is created. Starting from the root of the Hoffman tree, a root subtable (level = 0, root = 0, tabl = 4) is created (step 44). First, after the sub table is created, the sub table is scanned to check whether another sub table needs to be expanded from the sub table (step 46). If necessary, the base address of this subtable is assigned and written by a recursive call. If there is no subtable to be extended, it will return.

In each table creation, the original Hoffman table is scanned from beginning to end. In each Hoffman codeword, by checking the level value and using the independent variable root, it will check if it is already placed in the decoding time table, and if it does not belong to this subtable.

If the codeword belongs to this subtable, it will determine whether it is a leaf or should be expanded further, and if it is a leaf, all leaf items are matched to it and a mark (set table [i] .length to 128) is made. If it is not a leaf item, it creates a matching subroot item (just create an attribute field, the next field will be specified later).

The create-table function looks like this:

Build-table (int root, int tabl)

{int i, clen, flag, tbl;

if (level = 0) root = 0;

for (i = 0; ientry; i ++)

{flag = (table [i] .codeword (level1 * 4-4)) 12;

/ * assume 16 bits are used for codeword field of the table structure, otherwise 12 should be changed * /

if (level = 0) {flag = 0;}

if (table [i] .length! = 128 (table [i] .length 5) = level flag = root)

{clen = tab [i] .length 0xlf;

if (clen tabl) {make leaf items;

table [i] .length = 128;}

else {if (clen8) attribute = 128 + 4;

else {if (clen-4 attribute) attribute = 128 + clen-4;

table [i] .length = table [i] .length-tabl;

table [i] .length + = 32; }

}

for (i = 0; i (1tabl); I ++)

{if i th item is a subroot item

{tbl = attribute 15;

assign a value for the second field of this item;

level = level + 1;

Build-table (i, tbl);}

}

level-= 1;

return;

}

After create-table 0 is run, a scan is performed on all tables to add class D items. Another scan is performed on the root subtable to add the C class.

As described above, in the present invention, in order to improve the problem of the multi-bit search method, information indicating that the same 1-bit codes are continuously formed is formed in each item to form a C class item, and a pattern that does not exist in the Hoffman code table is a D class. By forming the item, it is possible to improve the operation speed during decoding and improve the efficiency of memory usage.

In addition, by providing a method of creating a decoding table, it can be used in software even in applications where the Hoffman table is not fixed.

Claims (5)

A decoding time table for use in decoding a Hoffman code by a multi-bit search method, the decoding time table comprising: a code field comprising a plurality of Hoffman codes having a predetermined number of bits or less; indicating an arrangement order of the plurality of codes An attribute field having attribute information such as a class for classifying the Hoffman codes, the number of consecutive identical codewords, the number of shifting of the index bits for search, or the number of index bits to be read; And a value field consisting of decoding values corresponding to the decoding time table. In the decoding method using the multi-bit search method of Hoffman code, Taking L bits of index bits from the Hoffman coded bitstream; Reading the root subtable of the decoding timetable; Searching for an item having the same index information as the taken index bit with reference to the read sub table; Determining the class of the item based on the attribute information of the searched item; Outputting the value of the item as a decoded value of a single 1-bit Hoffman codeword when determined to be Class A by the determination; Simultaneously outputting a value of an item as decoding values of a plurality of consecutive 1-bit Huffman codewords when determined to be a C class by the determination; Reading a new index bit from the bitstream according to the attribute information of the item if it is determined that the class is B by the determination; Reading a sub table corresponding to a value of the class B item and returning to the item search step; And After outputting the decoding values of the A and C classes, the index bits are shifted by a predetermined number of bits below the L by the attribute information to form L bit index bits for the next item search, and then the item searching step is performed. And returning the Hoffman code. In the decoding method using the multi-bit search method of Hoffman code, Taking L bits of index bits from the Hoffman coded bitstream; Reading the root subtable of the decoding timetable; Searching for an item having the same index information as the taken index bit with reference to the read sub table; Determining the class of the item based on the attribute information of the searched item; Outputting the value of the item as a decoded value of a single 1-bit Hoffman codeword when determined to be Class A by the determination; Simultaneously outputting a value of an item as decoding values of a plurality of consecutive 1-bit Huffman codewords when determined to be a C class by the determination; Reading a new index bit from the bitstream according to the attribute information of the item if it is determined that the class is B by the determination; Recognizing a code not defined on the Hoffman code table and performing an exceptional operation if it is determined that the class is D; Reading a sub table corresponding to a value of the class B item and returning to the item search step; And After outputting the decoding values of the A and C classes, the index bits are shifted by a predetermined number of bits below the L by the attribute information to form L bit index bits for the next item search, and then the item searching step is performed. And returning the Hoffman code. A method of creating an improved decoding time table in a multi-bit search method of Hoffman code, Defining first global data for storing the original Hoffman table and second global data for storing the decoding timetable; Defining a level variable indicating a level of a sub table being created and an entry variable indicating a number of codewords registered in the Hoffman code table; Creating a root subtable according to a given table size while referring to the Hoffman code table; Checking the created root subtable to determine which subtable needs to be expanded; Creating a subtable by allocating a base address of the subtable and repeatedly performing the subtable creating step and the determining step when the subtable needs to be expanded; And And returning if no expansion is necessary in the determining step. 5. The method of claim 4, wherein in the step of creating the root subtable, a scan is performed on all tables to add a D class item, and then another scan is performed on a root subtable to add a C class. An improved decoding time table creation method.