Data efficient quantization table for a digital video signal processor

Data Quantization Table for Digital Video Processor

The field of the invention relates to compression / decompression networks. In particular, the present invention relates to modifying the characteristics of a compression / decompression table to simplify data throughput and memory efficiency.

The resolution of the encoded characters is very important for accurately structured encoded data. Fixed-length quantization lookup tables allow the compression network to efficiently quantize and dequantize data using minimal processing. A quantized table that uses more bits to represent the output coded characters has better resolution than a table that uses fewer bits to represent the output coded characters. However, more bits require more memory to store data after quantization, and larger bandwidth is required to transmit the data. For a given quantization level, when a plurality of input data points having approximately the same value are compressed into an output value through a quantization table, the quantization table will cause wear. During reconstruction, the same dequantized value represents the same dequantized value of all data points within the resolution of a particular level in the table. The data difference depends on the resolution of the quantization table used to compress and decompress the data. It is known that there are multiple levels specified in the fixed-length quantization table for representing the number of bits of the output encoding character, and all encoding characters are represented by the same number of bits in a given table. For example, a table with a 3-bit output encoding character has 8 levels (2³), And the 4-bit table has 16 levels (2⁴). The average resolution of the table, and the generality of each bit is divided by the number of columns in the table and the number of levels in the table.

In the quantization table, the output quantized data is represented by a plurality of bit characters or symbols. In the article, it is considered that for some types of data, using a quantization table whose symbols are less than the number of dominant bits in each symbol for at least one quantization level can reduce the bandwidth and required memory greatly. According to the principle of the present invention, in addition to applying symbols less than M bits to compress at least one received value level at a known general rate, the compression and decompression table has a relationship with the corresponding bits of the main M bits. N level.

FIG. 1 shows a 4-bit quantization table in the conventional technique.

FIG. 2 shows a 4-bit quantization table in the present invention.

Figure 3 shows a block diagram of a compression / decompression network that can be used with the present invention.

The flowchart of FIG. 4 defines one possible method of designing a table in accordance with the principles of the present invention.

In a representative embodiment, each of the 15 quantization tables is related to a 4-bit output coded character symbol, and only the 7th level is related to the frequency of occurrence of input data. In this example, a 3-bit symbol is used. Each time a short symbol is used, the bandwidth and memory are reserved for other users. For large data sequences, such as data sequences in video data, memory and bandwidth can be reduced considerably. The invention also relates to a dequantization table.

The above type of quantization table is a mixed double-length table. Depending on the number of bits used to represent the coded character, the levels of multiple tables are selected to produce the amount of bandwidth reduction used. No double-length table is used. The mixed table is an N-length table, where N is the number of levels with short coded characters. Moreover, if desired, the number of bits in the short symbol may be two or more. This work requires a more sophisticated state machine to track the number of bits saved.

Generally, there is a fixed length table to produce a fixed and known bit rate / bandwidth savings, and each data value is quantized to have the same number of bits. There are changes to the length table to produce the maximum bit rate, and the minimum bandwidth savings achieved. For example, for a quantization table, both the fixed-length table and the changed-length table generate a certain allowable quantization system to maintain a higher resolution within a reduced bandwidth. An example of this state is inserting excess data into a data string that defines or knows the size / rate / bandwidth of the data.

FIG. 1 shows a fixed-length loss quantization table in the conventional technique. This table has 128 fields (-64 to 63 are included) and reduced input values, such as 7-bit values, to 4-bit output tables. Each symbol represents a dequantized data value. The quantization table parameter includes a quantization level index, which is related to the decision table, the reconstruction level, and the quantized output codeword symbol. The decision point sets the quantization boundary and recognizes the input value of the relevant input symbol (0000... 1110). The input table is equal to the decision point value, but is greater than the value previously determined in the level, and is represented by the relevant symbol during compression, and can be represented by the reconstruction level when decompressed. For example, the decision point for level index 0 is -50, which contains input values from -50 to -64, is represented by the symbol 0000, and can be reconstructed by a value of -57. Because the symbol length is 4 bits in this example, there are 16 (2⁴) Possible level. However, when the table (FIG. 1 and FIG. 2) indicates different pulse wave code modulation (DPCM) prediction errors, the table generally has an odd number of levels (15) to maintain symmetry on the prediction errors. This will be explained later in the DPCM process.

The average resolution of each quantization level is 8.5 (such as 128 columns divided by 15 levels), but any one level can be higher or lower than the average resolution based on other factors unique to one's compression / decompression network. For any given input value in the table column, the corresponding input value for the decision point classification, where the input value is less than or equal to the decision point, but greater than the previous decision point. The index of the decision point then identifies the 4-bit symbol used to represent the data value. When decompressed, 4-bit symbol identification results in indexing of reconstructed values. The reconstruction is equal to or approximately equal to the original input data value.

FIG. 2 design and configure one value of the principle of the present invention. The value has 128 fields (-64 to 63 are included), and the reduced 7-bit input value becomes 14 4-bit output symbols or a 3-bit output symbol (level 7). In a representative 4-bit table, the table has 15 possible quantization levels compared to the maximum 16 possible levels. Regardless of the DPCM processing included in the tables of Figures 1 and 2, one of the digital symbols (000) contains only 3 bits, so the number of possible levels is reduced (from the 2 "maximum). Bits in the table The number of elements minus 1 can reduce the maximum resolution from the table by an average of about 6.3% for each level. For 128 columns, a 16 level table has an average resolution of 8.0 per bit, and a 15 The level table has an average average resolution of 8.5 per bit (for example, 128 columns divided by 15 levels). A 3-bit table will have only 8 possible levels. Use a 15 level table instead of 8 levels Table, 7 levels are added to the 3-bit table, and the average resolution for each level is increased by 87.5%. If a 3-bit symbol (for example, 000 in this example) is arranged in the 15 level table, So it can be accessed frequently. Then the reduction of bandwidth and memory requirements becomes quite meaningful, which is more important than the loss in data resolution. Then, according to the present invention, in a quantization table with N levels (such as 15 levels), each level contains relevant output symbols that are mainly M bits (such as 4 bits), at least one level (such as level 7) and frequently occurs The value is related to the corresponding sign of less than M bits (such as 3 bits).

When extracting data to understand compression, the decompression network must recognize 3-bit and 4-bit symbols. This operation can be simplified by inverting the bit pattern of the 3-bit sign. For example, in FIG. 2, only the 3-bit symbol has the pattern "000". The first 3 bits of all symbols are preserved, so that as long as the "000" bit pattern appears in these bits, the network recognizes the 3 bit symbols, and when dequantizing and reconstructing the 8 bit data values, the network The fourth bit is not processed. For the 4-bit pattern of all symbols, any selected 3-bit pattern will be found twice. This is because the network can only recognize the one-level of 3-bit symbols, and no other information is required. Any 3-bit pattern can be used in a 4-bit table, and only 3-bit symbols have a reserved pattern.

The system used for this table is determined by the designer. A more efficient method is to use two 3-bit symbols instead of one 3-bit symbol in a main 4-bit table. Then 14 or less Or design quantification and dequantization tables. Any two-bit pattern in any bit position, such as "00" will only occur 4 times out of 16 possible symbols. The 4-bit output sign is assigned to the 12 level without the short sign bit pattern. Using a two-bit pattern to identify three-bit symbols allows one bit to be identified between three-bit symbols, which can uniquely identify two different levels. This is the same as the one with a unique 3-bit pattern, where each meta-sign will appear twice within 16 possible symbols. Therefore, 14 of the 16 levels can be used in this configuration.

Locating 3-bit symbols can therefore optimize sign changes from frequently occurring input data values during quantization. Therefore, a specific system needs to be calculated in advance, and a 3-bit symbol can be statistically identified to be placed somewhere. If accurate statistical measurements are required, there can be more than one compression / decompression network in the system. For example, the table in FIG. 2 basically processes the differential pulse code modulation (DPCM) prediction error obtained from a prediction network. In a compression network using DPCM, the prediction network uses the previous data value to predict the next decompression value. Determine the difference between the actual and predicted values. This difference is the prediction error. The absolute value of the prediction error is generally smaller in value than the actual or predicted value, and therefore can be accurately represented with a smaller number of bits.

Generally, the DPCM prediction error value is generated symmetrically around the zero error, and statistically, it usually occurs near a bell curve that is close to the zero error. Before designing the table by obtaining a distribution of all prediction error values, the frequency at which the error values are generated can be measured, where the prediction error values will be entered into the system. Using this information, the 3-bit symbols are placed where they are most likely to be used, so that bandwidth and memory requirements can be minimized while maintaining processing efficiency. The resolution of the level represented by the 3-bit symbol can be adjusted so that the 3-bit symbol can be best used. In FIG. 2, the 3-bit symbol is placed at about zero, which is where the prediction error of the system is most likely to occur. Because the position of the table level is optimized statistically, the special level (level 7) has a resolution of 6 from 3 to -2. In addition, according to the type of the quantization network included, the positioning of the 3-bit symbols is determined based on the statistical state in which the received data values are quantized and dequantized. This statistical positioning can be between different versions of the system without being affected by the short-coded character symbol configuration.

The symbol can be designed to have the advantage of symmetrical input data, such as DPCM input data represented by the prediction error value. The table in Figure 2 is made based on DPCM processing, where a symmetric 0 value produces an input value. Therefore, the unit symbol can be retained as a sign bit. The rightmost (most significant bit) in levels 0 to 6 is "0", and the bit sign in levels 8 to 14 is "1". Then when reconstructing, the network needs to decode 3 of the 4-bit symbols, and a less complicated circuit can assign the correct signal to the reconstructed data value.

FIG. 3 shows a system that has been built using the quantization and dequantization tables of the present invention. The quantization network 12 receives input data as an input 10 from an input network (not shown). The input data is sent to a quantizer 20 and a combiner 22. The required input value passes through the quantizer 20 to the predictor 18, and the predictor generates a predicted value for the value to be quantized. The combiner 22 receives the predicted value and subtracts this value from the original input value related to the predicted value. The difference of the prediction error value is received by the quantizer 20, and the quantizer uses the quantization table designed by the present invention to quantize the prediction error value. The compressed data output symbols from the quantizer 20 are sent to the memory 14.

When no data is needed in the output network (not shown in the figure), the decompression network 16 receives the compressed data from the picture memory 14, and the dequantizer uses 26 a dequantization table designed according to the principles of the present invention to dequantize Compressed prediction error value. The prediction error value is sent to a predictor 24, and the predictor 24 and the predictor 12 are located, and a predicted value is generated. The predicted value is sent back to the dequantizer 26, and a decompressed prediction error value is added, resulting in a corresponding reconstructed original input value, or an approximate value. The reconstructed value is sent to the output network.

The input network may be a signal processor of an MPEG-compatible television receiver that receives audio and video signals encoded and decoded in MPEG format. The receiver decodes and decompresses the received signal and provides an 8-bit 8-bit image picture element (pixel) for input 10. The corresponding output network can be a processor compatible with standard or high-resolution displays. The display processor needs to randomly access pixel blocks in a given image frame to obtain motion compensation information. The picture memory 14 stores image pictures until required by the design processor.

FIG. 4 illustrates a method of designing a quantization and dequantization table for use in the network of FIG. 3 described above. It is not necessary to completely follow the steps in the diagram when designing the form. For example, step 46 may be performed before step 44 and step 48 may be performed at any time. In step 40, the number of levels in the table must be determined. This depends on the number of symbols, the type of data received, the type of network that will process the data, and other specific variables used in the system. In step 42, statistical methods are applied to analyze the data and the system to determine and categorize frequencies where input data values will occur. In step 44, the table level of the frequency of receiving the generated data is assigned to the short symbol. In step 46, the normal length is assigned to other levels. In step 48, a one-bit symbol is reserved as a sign bit, and the sign bit can be separately processed by a less complicated circuit to save the processing amount, and the correct sign is added to the data. At this time, in step 50, the resolution of each level is defined. This level can be adjusted so that some levels have more precise level resolution data than the others and statistical analysis results using the system.

The aforementioned quantization network should not be confused with Huffman coding, which is well known to those skilled in the art. Huffman encoding is a lossless statistical entropy encoding, and the encoding character length is less than or greater than the average output data length. Moreover, each input uses a Huffman-encoded encoder input that uses a unique symbol generated at the output of the encoder. Because Huffman coding is an enthalpy coding, it is impossible to use Huffman coding in a compression system with a fixed bit rate. Huffman coding does not provide the required control over a fixed bit rate.

The method of the present invention can simplify the memory and reduce the frequency bandwidth, even when the frequencies at which the levels occur are approximately equal, because the method and device are used in a wear-out system. The symbol used in the present invention indicates a range of input data determined by a decision point related to a quantization level. Basically, the original data was not completely copied during decompression. Moreover, using Huffman coding instead of the coding described above will result in more complex fixed structures and longer symbol lengths than the average / primary symbol length. This is because the coded characters must have a unique pattern to distinguish between different coded characters. Make a difference. That is, if the 4-bit code is "0101", a coded character with more bits cannot have "0101" in the first 4 bits, otherwise the decoder will misunderstand the bit pattern. In the quantization table of the present invention, for N bits, only the short symbol has a unique bit pattern. After the first N bits, longer bit patterns must be repeated. Compared with this Huffman coding, the complexity of the network has been greatly reduced.