HK1211754B - Arithmetic encoding device and arithmetic decoding device - Google Patents

Description

Arithmetic coding device and arithmetic decoding device

This application is a divisional application of the invention patent application with application date of October 1, 2010, application number 201080045319.6, and invention name “Method and device for arithmetic coding or arithmetic decoding”.

Technical Field

The present invention relates to arithmetic coding and decoding of multimedia data.

Background Art

Arithmetic coding is a method for lossless data compression. Arithmetic coding is based on a probability density function (PDF). To achieve compression, the PDF used for coding must be identical to, or at least similar to, the actual PDF of the data—the closer the better.

If arithmetic coding is based on an appropriate probability density function, significant compression leading to at least nearly optimal codes can be achieved. Therefore, arithmetic coding is a frequently used technique in audio, speech or video coding for encoding and decoding sequences of coefficients, where these coefficients are quantized time-frequency transforms of the values of video pixels or audio or speech signal samples represented in binary.

To further improve compression, arithmetic coding can be based on a set of probability density functions, where the probability density function used to encode a current coefficient depends on the context of the current coefficient. That is, depending on the context in which coefficients with the same quantized value appear, different probability density functions can be used to encode the same quantized value. The context of a coefficient is defined by the quantized values of the coefficient contained in a neighborhood of one or more adjacent coefficients adjacent to the respective coefficient, for example, a subsequence of one or more coded or decoded coefficients that immediately precede the respective coefficient to be coded or decoded in the sequence. Each different possibility in which a neighborhood may appear defines a different possible context that is mapped to a probability density function.

In fact, the compression improvement becomes noticeable only when the neighborhood is large enough. This is accompanied by a combinatorial explosion in the number of different possible backgrounds and a correspondingly huge number of possible probability density functions or correspondingly complex mappings.

An example of a background arithmetic coding scheme can be found in ISO/IEC JTC1/SC29/WG11 N10215, October 2008, Busan, Korea, Proposing a Reference Model for Unified Speech and Audio Coding (USAC). According to this proposal, already coded 4-tuples are considered as background.

Another example of arithmetic coding based on a USAC related background can be found in ISO/IEC JTC1/SC29/WG11 N10847, July 2009, London, UK.

In order to reduce the complexity of high-order conditional entropy coding, US Pat. No. 5,298,896 proposes non-uniform quantization of limited symbols.

Summary of the Invention

The large number of backgrounds to be processed corresponds to a large number of probability density functions, or at least a correspondingly complex mapping from backgrounds to probability density functions, that must be stored, retrieved, and processed. This increases at least one of encoding/decoding latency and storage capacity requirements. Therefore, there is a need in the art for an alternative solution that allows for compression that is nearly as good while reducing at least one of encoding/decoding latency and storage capacity requirements.

In order to address this need, the present invention proposes an apparatus for arithmetically decoding a current spectral coefficient using previous spectral coefficients and an apparatus for arithmetically encoding a current spectral coefficient using previous spectral coefficients according to an embodiment, as described herein.

Characteristic features of further proposed embodiments are specified in the dependent claims.

A device for arithmetically decoding a current spectral coefficient using previous spectral coefficients, the previous spectral coefficients being already decoded, wherein the previous spectral coefficients and the current spectral coefficient are contained in one or more quantized spectra obtained by quantizing sample values of a time-frequency transformed video, audio or speech signal, the device comprising:

-processing unit for processing pre-spectral coefficients;

- a first means for determining a background class from at least two different background classes, said first means being adapted to use the processed front spectral coefficients for determining the background class,

The sum of the quantized absolute values of the front spectrum coefficients is used to determine the background category;

- second means for determining a probability density function, said second means being adapted to use the determined background class and the mapping from at least two different background classes to at least two different probability density functions for determining the probability density function;

- an arithmetic decoder for arithmetically decoding the current spectral coefficient according to the determined probability density function,

wherein the processing means is adapted to non-uniformly quantize the absolute values of the front spectral coefficients for use in the determination of the background class;

The second component is configured to implement the mapping using a lookup table or a hash table.

A device for arithmetically encoding a current spectral coefficient using previous spectral coefficients, wherein the previous spectral coefficients are already encoded, wherein the previous spectral coefficients and the current spectral coefficient are contained in one or more quantized spectra obtained by quantizing sample values of a time-frequency transformed video, audio or speech signal, the device comprising:

-processing unit for processing pre-spectral coefficients;

- a first means for determining a background class from at least two different background classes, said first means being adapted to use the processed front spectral coefficients for determining the background class,

The sum of the quantized absolute values of the front spectrum coefficients is used to determine the background category;

- second means for determining a probability density function, said second means being adapted to use the determined background class and the mapping from at least two different background classes to at least two different probability density functions for determining the probability density function;

- an arithmetic coder for arithmetically coding the current spectral coefficient according to the determined probability density function,

wherein the processing means is adapted to non-uniformly quantize the absolute values of the front spectral coefficients for use in the determination of the background class;

The second component is configured to implement the mapping using a lookup table or a hash table.

The method of arithmetic coding or decoding uses the previous spectrum coefficients for arithmetic coding or decoding of the current spectrum coefficients, respectively, wherein the previous spectrum coefficients have been encoded or decoded respectively. Both the previous spectrum coefficients and the current spectrum coefficients are contained in one or more quantized spectra obtained by quantizing the time-frequency transformation of the sample values of the video, audio or speech signal. The method further includes processing the previous spectrum coefficients, using the processed previous spectrum coefficients to determine a background category as one of at least two different background categories, using the determined background category and the mapping from the at least two different background categories to at least two different probability density functions to determine a probability density function, and arithmetically encoding or decoding the current spectrum coefficients according to the determined probability density function. A feature of the method is that the processed previous spectrum coefficients include the absolute values of the non-uniformly quantized previous spectrum coefficients.

Using background classes instead of backgrounds for determining the probability density function facilitates grouping two or more different backgrounds that yield different but very similar probability density functions into a single background class that is mapped to a single probability density function. This grouping is achieved by using the non-uniform quantized absolute values of the front spectral coefficients for determining the background class.

For example, there are embodiments in which processing the front spectral coefficients includes determining the sum of the quantized absolute values of the front spectral coefficients for use in determining the background class. Similarly, there are corresponding embodiments of the arithmetic encoding device and corresponding embodiments of the arithmetic decoding device, in which the processing means are adapted to determine the sum of the quantized absolute values of the front spectral coefficients for use in determining the background class.

In a further embodiment of the device, the processing means is adapted to cause the processing of the pre-spectral coefficients to further comprise a first quantization of the absolute values of the pre-spectral coefficients according to a first quantization scheme, a variance determination of determining a variance of the absolute values of the pre-spectral coefficients quantized according to the first quantization scheme, using the determined variance to select one of at least two different non-linear second quantization schemes, and a second quantization of the absolute values of the pre-spectral coefficients quantized according to the first quantization scheme according to the selected non-linear second quantization scheme. A further embodiment of the method comprises the corresponding steps. The variance determination may comprise determining a sum of the absolute values of the pre-spectral coefficients quantized according to the first quantization scheme, and comparing the determined sum value with at least one threshold value.

In a further embodiment, the processing means of each device may be adapted to cause the processing to result in a first consequence or at least one different second consequence. Determining the context class then further comprises determining a number of those prespectral coefficients for which the processing results in the first consequence, and using the determined number for determining the context class.

Each device may comprise means for receiving at least one of a mode switch signal and a reset signal, wherein the device is adapted to use at least one of the received signals for controlling the determination of the context category.

The at least two different probability density functions may be determined in advance by using a representative data set to determine the at least two different probability density functions, and the mapping may be implemented using a lookup table or a hash table.

BRIEF DESCRIPTION OF THE DRAWINGS

Exemplary embodiments of the present invention are illustrated in the accompanying drawings and described in more detail in the following description. These exemplary embodiments are only provided to illustrate the present invention, rather than to limit the scope and spirit of the present invention as defined in the claims.

In the attached figure:

FIG1 schematically illustrates an embodiment of an encoder according to the present invention;

FIG2 exemplarily depicts an embodiment of a decoder according to the present invention;

FIG3 exemplarily depicts a first embodiment of a background classifier for determining background categories;

FIG4 exemplarily depicts a second embodiment of a background classifier for determining background categories;

FIG5 a exemplarily depicts a first neighborhood of a previous spectral bin preceding a current spectral bin to be encoded or decoded in frequency domain mode;

FIG5 b exemplarily depicts a second neighborhood of a previous spectral region preceding a current spectral region to be encoded or decoded in a weighted linear prediction transform mode;

FIG6 a exemplarily depicts a third neighborhood of the preceding spectral region preceding the current lowest spectral region to be encoded or decoded in frequency domain mode;

FIG6 b exemplarily depicts a fourth neighboring region of a preceding spectral region preceding a current second lowest spectral region to be encoded or decoded in frequency domain mode;

FIG7 a exemplarily depicts a fifth neighborhood of the preceding spectral region preceding the current lowest spectral region to be encoded or decoded in the weighted linear prediction transform mode;

FIG7 b exemplarily depicts a sixth neighbor of a previous spectral region preceding a current second lowest spectral region to be encoded or decoded in a weighted linear prediction transform mode;

FIG7 c exemplarily depicts a seventh neighbor of the previous spectral region preceding the current third lowest spectral region to be encoded or decoded in the weighted linear prediction transform mode;

FIG7 d exemplarily depicts an eighth neighbor of the preceding spectral region preceding the current fourth lowest spectral region to be encoded or decoded in the weighted linear prediction transform mode;

FIG8 exemplarily depicts a neighborhood of different spectral regions to be encoded or decoded, said different spectral regions being included in a first spectrum to be encoded or decoded after starting encoding/decoding in frequency domain mode or after occurrence of a reset signal; and

FIG9 exemplarily depicts further neighborhoods of different spectral regions to be encoded or decoded in the weighted linear prediction transform mode, said different spectral regions being contained in a second spectrum to be encoded or decoded after the start of encoding/decoding in the weighted linear prediction transform mode or the occurrence of a reset signal.

DETAILED DESCRIPTION

The present invention can be implemented in any electronic device that includes a correspondingly adapted processing device. For example, the device for arithmetic decoding can be implemented in a television, mobile phone, personal computer, MP3 player, navigation system, or car audio system. The device for arithmetic coding can be implemented in a mobile phone, personal computer, active car navigation system, digital camera, digital video camera, or audio recorder.

The exemplary embodiments described below relate to encoding or decoding of quantized spectral regions resulting from time-frequency transformation of quantized multimedia samples.

The present invention is based on the method of using an already transmitted quantized spectral region, for example, a previous quantized spectral region before a current quantized spectral region BIN in a sequence, to determine a probability density function for respectively arithmetic encoding and decoding the current quantized spectral region BIN.

The exemplary embodiments of the methods and apparatus for arithmetic coding or arithmetic decoding each include steps or components for non-uniform quantization. While all steps or components together provide the highest coding efficiency, each step or component individually implements the concepts of the present invention and offers benefits related to encoding/decoding latency and/or storage requirements. Therefore, the detailed description should be understood as describing exemplary embodiments that implement only one of the steps or components, as well as exemplary embodiments that implement a combination of two or more steps or components.

The first step that may, but need not, be included in an exemplary embodiment of the present method is a switching step that determines which general transform mode should be used. For example, in the USAC noiseless coding scheme, the general transform mode may be a frequency domain (FD) mode or a weighted linear predictive transform (wLPT) mode. Each general mode may use a different neighborhood of the coded or decoded spectral region, i.e., a different selection, for determining the PDF.

Thereafter, the background of the current spectral region BIN can be determined in the Background Generation COCL module. Based on the determined background, a background class is determined by background classification, wherein the background is preferably, but not necessarily, processed by non-uniform quantization of the spectral region NUQ1 before classification. The classification can include estimating the variance of the background VES and comparing the variance to at least one threshold. Alternatively, a variance estimate can be determined directly from the background. The variance estimate is then used to control a further quantization NUQ2, preferably, but not necessarily non-linear.

In the encoding process exemplarily depicted in FIG1 , a suitable probability density function (PDF) is determined for encoding the current quantized spectral region BIN. For this purpose, only information that is also known on the decoder side can be used. That is, only the previously encoded or decoded quantized spectral region can be used. This is done in the background classifier block COCL. There, the selected previous spectral region defines a neighborhood NBH for determining the actual background class. The background class can be represented by a background class number. The background class number is used to retrieve the corresponding PDF from the PDF memory MEM1 via a mapping MAP, for example, via a lookup table or a hash table. The determination of the background class may depend on a general mode switch GMS that allows the use of different neighborhoods depending on the selected mode. As mentioned above, for USAC, there may be two general modes (FD mode and wLPT mode). If the general mode switch GMS is implemented on the encoder side, the mode change signal or the current general signal must be included in the bit stream so that the decoder also knows it. For example, in the reference model of Unified Speech and Audio Coding (USAC) proposed by ISO/IEC JTC1/SC29/WG11N10847, July 2009, London, UK, there are Table 4.4 core_mode and Table 4.5 core_mode0/1 proposed for transmitting general modes.

After determining a PDF suitable for encoding the current quantized spectral region BIN by the arithmetic encoder (AEC), the current quantized spectral region BIN is fed into the neighborhood memory MEM2. That is, the current spectral region BIN becomes the previous spectral region. The previous spectral region contained in the neighborhood memory MEM2 can be used by the block COCL to encode the next spectral region BIN. The current spectral region BIN is arithmetically encoded by the arithmetic encoder (AEC) during, before, or after the storage of the current spectral region BIN. The output of the arithmetic encoder (AEC) is stored in the bit buffer BUF or written directly to the bitstream.

The bit stream or the contents of the buffer BUF may be transmitted or broadcasted via, for example, cable or satellite. Alternatively, the arithmetically coded spectral region may be written on a storage medium like a DVD, hard disk, Blu-ray disc, etc. The PDF memory MEM1 and the neighborhood memory MEM2 may be implemented in a single physical memory.

A reset switch, RS, allows for the occasional restart of encoding or decoding at a dedicated frame, known as a decoding entry point, without requiring prior knowledge of the preceding spectrum. If the reset switch, RS, is implemented on the encoder side, the reset signal must be included in the bitstream so that the decoder is also aware of it. For example, in the Unified Speech and Audio Coding (USAC) reference model, as proposed in ISO/IEC JTC1/SC29/WG11N10847, July 2009, London, UK, the arith_reset_flag is included in WD Tables 4.10 and 4.14.

FIG2 exemplarily illustrates a decoding scheme based on a corresponding neighborhood. It contains blocks similar to the encoding scheme. The PDF to be used for arithmetic decoding is determined in the same way as the encoding scheme, ensuring that the determined PDF is identical in both the encoder and decoder. Arithmetic decoding retrieves bits from the bit buffer BUF or directly from the bitstream and uses the determined PDF to decode the current quantized spectral region BIN. The decoded quantized spectral region is then fed into the neighborhood memory MEM2 of the background category number determination block COCL and can be used to decode the next spectral region.

FIG3 exemplarily depicts a first embodiment of a background classifier COCL for determining a background class in more detail.

Before the current quantized spectral region BIN is stored in the spectral memory MEM2, it can be subjected to non-uniform quantization in the block NUQ1. This has two advantages: firstly, it makes the storage of the quantized spectral regions, which are usually 16-bit integer values with symbol values, more efficient, and secondly, it reduces the number of values that each quantized spectral region has. This results in a significant reduction of the possible background categories during the background category determination in the block CLASS. Furthermore, since the symbols of the quantized spectral region may be discarded in the background category determination, the absolute value calculation may be included in the non-uniform quantization block NUQ1. In Table 1, an exemplary non-uniform quantization as can be performed by the block NUQ1 is shown. In this example, after the non-uniform quantization, each spectral region may have three different values. However, in general, the only constraint of non-uniform quantization is to reduce the number of values that a spectral region may take.

Table 1 Example non-uniform quantization steps including absolute value calculation

The non-uniformly quantized/mapped spectral regions are stored in the spectral memory MEM2. According to the selected general mode, GMS is selected, the background class CLASS is determined for each spectral region to be coded, and the selected neighborhood NBH of the spectral region is selected.

FIG5 a exemplarily depicts a first exemplary neighborhood NBH of a spectral region to be encoded or decoded.

In this example, only the spectral regions of the actual or current spectrum (frame) and the spectral regions of one previous spectrum (frame) are defined as the neighborhood NBH. Of course, more than one spectral region of the previous spectrum can be used as part of the neighborhood, which increases complexity but may ultimately provide higher coding efficiency. Note that only spectral regions that have been transmitted relative to the actual spectrum can be used to define the neighborhood NBH, as they must also be accessible at the decoder. Here, and in the following examples, it is assumed that the spectral regions are transmitted in order from low frequency to high frequency.

The selected neighborhood NBH is then used as input to the background category determination block COCL. In the following, the general idea behind background category determination as well as a simplified form is first explained before describing a specific implementation.

The general idea behind background class determination is to obtain a reliable estimate of the variance of the spectral region to be coded. This predicted variance can be used again to obtain an estimate of the PDF of the spectral region to be coded. For variance estimation, it is not necessary to evaluate the symbols of the spectral regions in the neighborhood. Therefore, symbols can already be discarded in the quantization step before storage in the spectrum memory MEM2. A very simple background class determination may look like the following: the neighborhood NBH of the spectral region BIN may look like in FIG5 a, consisting of 7 spectral regions. If the exemplary non-uniform quantization shown in Table 1 is used, each spectral region can have 3 values. This results in 3 ⁷ = 2187 possible background classes.

To further reduce the number of possible background categories, the relative position of each bin in the neighborhood NBH can be discarded. Thus, only bins with values of 0, 1, or 2 are counted, where the sum of the number of 0 bins, the number of 1 bins, and the number of 2 bins is equal to the total number of bins in the neighborhood. In a neighborhood NBH containing n bins that can each assume one of three different values, there are 0.5*( ^n² + 3*n+2) possible background categories. For example, in a neighborhood of 7 bins, there are 36 possible background categories, while in a neighborhood of 6 bins, there are 28 possible background categories.

A more complex but still quite simple background class determination takes into account studies showing that the spectral regions of the same frequency as the previous spectrum are particularly important (the spectral regions depicted with dotted circles in Figures 5a, 5b, 6a, 6b, 7a, 7b, 7c, 8 and 9). For the other spectral regions in the neighborhood, i.e., those depicted as circles with horizontal lines in the respective figures, the relevant positions are less relevant. Therefore, the spectral regions of the same frequency in the previous spectrum are explicitly used for background class determination, while for the other 6 spectral regions, only the number of 0 spectral regions, the number of 1 spectral regions and the number of 2 spectral regions are counted. This results in 3×28=84 possible background classes. Experiments have shown that such background classification is very effective for FD mode.

The background class determination can be extended by controlling the variance estimate VES of the second non-uniform quantization NUQ2. This makes the background class generation COCL better adapted to the higher dynamic range of the prediction variance of the spectral region to be coded. A corresponding block diagram for the extended background class determination is shown exemplarily in FIG4.

In the example shown in FIG4 , the non-uniform quantization is divided into two steps, wherein the first step provides a finer quantization (block NUQ1) and the second step provides a coarser quantization (block NUQ2). This facilitates the quantization to be applied, for example, to the variance of a neighborhood. The variance of the neighborhood is estimated in the variance estimation block VES, wherein the variance estimate is based on the previous finer quantization of the spectral region in the neighborhood NBH in block NUQ1. The estimate of the variance does not need to be exact, but can be very rough. For example, it is sufficient to apply USAC to determine whether the absolute value of the spectral region in the neighborhood NBH after the finer quantization reaches or exceeds a variance threshold, that is, it is sufficient to switch between high and low variance.

2-step non-uniform quantization may look like Table 2. In this example, the low variance mode corresponds to the 1-step quantization shown in Table 2.

Table 2 depicts an exemplary 2-step non-uniform quantization; the second or subsequent steps are quantized differently depending on whether the variance is estimated to be high or low

The final background class determination in block CLASS is the same as in the simplified version of FIG3 . Different background class determinations can be used according to the variance pattern. More than two variance patterns can also be used, which of course results in an increased number of background classes and increased complexity.

For the first spectral region in the spectrum, a neighborhood like that shown in Figures 5a or 5b is not applicable, because for the first spectral region, not all lower frequency spectral regions exist or do not exist. For each of these special cases, a own neighborhood can be defined. In a further embodiment, predetermined values are filled in the non-existent spectral regions. For the exemplary neighborhood given in Figure 5a, the defined neighborhood for the first spectral region to be transmitted in the spectrum is as shown in Figures 6a and 6b. The idea is to extend the neighborhood to higher spectral regions so that the same background category determination function can be used for the rest of the spectrum. This also means the same background category and at least the same PDF can be used. This is not possible if the size of the neighborhood is simply reduced (which is of course also an option).

A reset typically occurs before encoding a new spectrum. As mentioned above, a dedicated starting point is necessary for decoding. For example, if the decoding process begins at a specific frame/spectrum, the decoding process must actually begin at the point of the last reset, decoding the previous frames sequentially until the desired starting spectrum is reached. This means that the more resets occur, the more exit points there are for decoding. However, encoding efficiency is lower in the spectrum after the reset.

After a reset occurs, there is no previous spectrum available for neighborhood definition. This means that only the first spectral region of the actual spectrum can be used in the neighborhood. However, the general process may not change, and the same "tools" can be used. However, the first spectral region must be treated differently than described in the previous section.

An exemplary reset neighborhood definition is shown in Figure 8. This definition can be used in the case of reset in FD mode of USAC.

8 (using quantization resulting in a table with 3 possible quantized values or 6 values if the values after quantization step 1 are used) is as follows: 1 background category is added for the processing of spectral region 1, 6 background categories are added for the processing of spectral region 2 (using the values after quantization step 1), 6 background categories are added for the processing of spectral region 3, and 10 background categories are added for the processing of spectral region 4. If two (high and low) variance modes are additionally considered, the number of such background categories is almost doubled (only for the first spectral region, for which no information is available, and for the second spectral region using the values of the spectral region after quantization step 1).

The resetting process results in 1+6+2×6+2×10=39 additional background categories in this example.

The mapping block MAP uses the background classification determined by the block COCL, for example, the determined background category number, and selects the corresponding PDF from the PDF memory MEM1. In this step, the necessary storage capacity can be further reduced by using a single PDF for more than one background category. That is, background categories with similar PDFs can use a joint PDF. These PDFs can be predetermined during a training phase using a sufficiently large representative dataset. This training can include an optimization phase in which background categories corresponding to similar PDFs are identified and the corresponding PDFs are merged. Depending on the data statistics, this can result in a significantly smaller number of PDFs that must be stored in the memory. In an exemplary experimental form of USAC, a mapping from 822 background categories to 64 PDFs was successfully applied.

If the number of background categories is not too large, the implementation of this mapping function MAP can be a simple lookup table. If the number is getting larger, a hash table search can be applied for efficiency reasons.

As described above, the general mode switch GMS allows switching between frequency domain mode (FD) and weighted linear predictive transform mode (wLPT). Depending on the mode, different neighborhoods can be used. Experiments have shown that the exemplary neighborhoods depicted in Figures 5a, 6a and 6b, and 8 are sufficiently large for FD mode. However, for wLPT mode, larger neighborhoods, as exemplarily depicted in Figures 5b, 7a, 7b, and 7c, and 9, have been found to be advantageous.

That is, an exemplary reset process in wLPT mode is depicted in FIG9 . Exemplary neighborhoods of the lowest, second lowest, third lowest, and fourth lowest spectral regions in the spectrum in wLPT mode are depicted in FIG7 a , 7 b , 7 c , and 7 d , respectively. Furthermore, exemplary neighborhoods of all spectral regions in the spectrum in wLPT mode are depicted in FIG5 b .

The number of background classes resulting from the exemplary neighborhood depicted in FIG5 b is 3×91=273 background classes. The factor 3 is due to the special treatment of a spectral region at the same frequency as a spectral region currently being encoded or decoded. According to the formula given above, for the remaining 12 spectral regions in the neighborhood, there are 0.5*((12*12)+3*12+2)=91 combinations of spectral region numbers having the values 2, 1, or 0. In an embodiment that relies on whether the variance of the neighborhood meets or exceeds a threshold to distinguish between background classes, the 273 background classes are doubled.

The exemplary reset process shown in FIG. 9 may also increase the number of background categories.

In an exemplary test experiment that produced good results in experiments, there were 822 possible background categories broken down in Table 1 below.

Table 1 Possible background categories for decomposition recommended by MPEG USAC CE

In the exemplary test embodiment, these 822 possible background categories are mapped to 64 PDFs. As mentioned above, this mapping is determined during the training phase. The resulting 64 PDFs must be stored in a ROM table, for example, with 16-bit precision for a fixed-point arithmetic coder. Here, another advantage of the proposed scheme is revealed: in the current working draft of the USAC standardization mentioned in the background section, a quadruple (a vector containing four spectral bins) is jointly encoded using a single codeword. Even if the dynamic range of each component in the vector is extremely small, this results in a very large codebook (e.g., each component can have the value [-4, ..., 3] → 8 ⁴ = 4096 possible different vectors). However, scalar encoding allows for an extremely small codebook even with a large dynamic range for each spectral bin. The codebook used in the exemplary test embodiment has 32 entries, providing a spectral bin dynamic range from -15 to +15 and an escape codeword (for the case where the spectral bin value is outside this range). This means that only 64×32 16-bit values must be stored in the ROM table.

The above describes a method for arithmetically encoding a current spectral coefficient using a previous spectral coefficient, wherein the previous spectral coefficient is already encoded and both the previous spectral coefficient and the current spectral coefficient are contained in one or more quantized spectra obtained by quantizing the time-frequency transform of the sample values of the video, audio or speech signal. In one embodiment, the method includes processing the previous spectral coefficient, using the processed previous spectral coefficient to determine a background category as one of at least two different background categories, using the determined background category and the mapping from the at least two different background categories to at least two different probability density functions to determine a probability density function, and arithmetically encoding the current spectral coefficient according to the determined probability density function, wherein the processed previous spectral coefficient includes non-uniformly quantized previous spectral coefficients.

In another exemplary embodiment, an apparatus for arithmetically encoding current spectral coefficients using already encoded previous spectral coefficients comprises a processing component, a first component for determining a background category, a memory for storing at least two different probability density functions, a second component for retrieving the probability density, and an arithmetic encoder.

Then, the processing means is adapted to process the encoded previous spectral coefficients by non-uniform quantization, and the first means is adapted to use the processing result to determine a background category as one of at least two different background categories. The memory stores at least two different probability density functions and a mapping from the at least two different background categories to the at least two different probability density functions for facilitating retrieval of a probability density function corresponding to the determined background category. The second means is adapted to retrieve the probability density function corresponding to the determined background category from the memory, and the arithmetic encoder is adapted to arithmetically encode the current spectral coefficient according to the retrieved probability density function.

There is a corresponding another exemplary embodiment of a device for arithmetically decoding current spectral coefficients using already encoded previous spectral coefficients, the device comprising a processing component, a first component for determining a background category, a memory for storing at least two different probability density functions, a second component for retrieving the probability density, and an arithmetic decoder.

Then, the processing means is adapted to process the encoded previous spectral coefficients by non-uniform quantization, and the first means is adapted to use the processing result to determine a background category as one of at least two different background categories. The memory stores at least two different probability density functions and a mapping from the at least two different background categories to the at least two different probability density functions for facilitating retrieval of a probability density function corresponding to the determined background category. The second means is adapted to retrieve the probability density function corresponding to the determined background category from the memory, and the arithmetic decoder is adapted to arithmetically decode the current spectral coefficient according to the retrieved probability density function.

Claims (2)

1. A device for arithmetic decoding of current spectral coefficients using previous spectral coefficients, wherein the previous spectral coefficients have already been decoded. The preceding spectral coefficients and the current spectral coefficients are included in one or more quantized spectra obtained by quantizing the sample values of the time-frequency transformed video, audio, or speech signal. The device includes: - Processing components for pre-processed spectral coefficients; - A first component for determining a background category from at least two distinct background categories, wherein the first component is adapted to use processed prespectral coefficients to determine the background category. The sum of the absolute values of the quantized prespectral coefficients is used to determine the background category; - A second component for determining the probability density function, the second component being adapted to use the determined background category and the mapping from at least two different background categories to at least two different probability density functions to determine the probability density function; - An arithmetic decoder that arithmetically decodes the current spectral coefficients based on the determined probability density function. The processing unit is adapted to non-uniformly quantize the absolute values of the pre-spectral coefficients for use in determining the background category; The second component is configured to implement the mapping using a lookup table or a hash table. 2. A device for arithmetic encoding current spectral coefficients using previous spectral coefficients, wherein the previous spectral coefficients are already encoded. The preceding spectral coefficients and the current spectral coefficients are included in one or more quantized spectra obtained by quantizing the sample values of the time-frequency transformed video, audio, or speech signal. The device includes: - Processing components for pre-processed spectral coefficients; - A first component for determining a background category from at least two distinct background categories, wherein the first component is adapted to use processed prespectral coefficients to determine the background category. The sum of the absolute values of the quantized prespectral coefficients is used to determine the background category; - A second component for determining the probability density function, the second component being adapted to use the determined background category and the mapping from at least two different background categories to at least two different probability density functions to determine the probability density function; - An arithmetic encoder that arithmetically encodes the current spectral coefficients based on the determined probability density function. The processing unit is adapted to non-uniformly quantize the absolute values of the pre-spectral coefficients for use in determining the background category; The second component is configured to implement the mapping using a lookup table or a hash table.