JP2004229184A - Method and device for encoding time-series signal - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a method and a device for encoding a time-series signal in which more highly efficient compression can be executed by selecting the optimum method as needed in accordance with fluctuations of signal characteristics in the time-series signal. <P>SOLUTION: An estimated error is calculated by a plurality of estimating calculation formulas, which are prepared beforehand, for each sample of the time-series signal consisting of sample sequences in time-series (S1). The estimated error to become an object for encoding is selected by utilizing an accumulated error in each estimating calculation expression (S2). After updating the accumulated error (S3), the accumulated error is re-set whenever the number of the samples exceed a prescribed number (S4, S5). The estimated error, which is optimum for each sample, can be obtained by executing the processing repeatedly. <P>COPYRIGHT: (C)2004,JPO&NCIPI

Description

[0001]
[Industrial applications]
The present invention relates to the field of music production, such as music production, storage of audio data materials, and relay of location materials, particularly to the field of producing high-definition audio with higher quality than CDs, and audio recording using digital recording media such as CDs and DVDs. The present invention relates to a reversible data compression technique suitable in a field where reproduction of data, a transmission of a biological signal in telemedicine, and the like where data modification is disliked, and the like are performed.
[0002]
[Prior art]
Conventionally, various methods have been used for compressing an acoustic signal. MP3 (MPEG-1 / Layer3), AAC (MPEG-2 / Layer3) and the like have been put to practical use as a technique for compressing and encoding an audio signal. With such a compression encoding method, it is possible to treat an audio signal as small data, which contributes to the efficiency of data recording and transmission.
[0003]
MP3, AAC, and the like as described above are all called lossy coding schemes, and can be efficiently compressed. However, decoding involves a considerable deterioration in quality and completely reproduces the original signal. It is not possible. For this reason, in the music production field such as music production, material storage, and location material relay, these encoding methods cannot be applied, and although inefficient, non-compressed storage / transmission methods are used. In particular, recently, the number of productions that handle high-definition audio has increased, the material capacity has become enormous, and this has become a problem in managing work disks.
[0004]
Recently, in order to solve the above problem, as a method of reversibly compressing and encoding an audio signal, a method using prediction encoding and combining prediction error data with encoding processing according to the frequency of appearance has been proposed. (For example, see Patent Document 1).
[0005]
In addition, the present applicant also performs a difference operation between channels and between frames on a sample sequence of an audio signal to reduce the value of each sample, and then compresses data using predictive coding. is suggesting. (See Patent Document 2).
[0006]
[Patent Document 1]
JP 2002-278600 A [Patent Document 2]
Japanese Patent Application No. 2002-231150 [0007]
[Problems to be solved by the invention]
However, in the technologies proposed in Patent Documents 1 and 2, the prediction method is limited to one. For example, the technique described in Patent Document 2 proposes predictive coding based on the past two samples. The genre of contents to which predictive coding based on the past two samples can be applied is the widest, but contents that change slowly are predicted based on the past one sample, and contents that change rapidly are predicted based on the past three or more samples. It has been confirmed that the error is reduced. Therefore, there is a method that prepares multiple formulas and selects an appropriate one according to the content, but if the entire content does not have uniform properties, the compression effect can not be obtained, especially vocal and mixed music Then, there is a problem that the compression efficiency is reduced.
[0008]
Further, in the content proposed in Patent Document 2, a method is employed in which a quantization noise component having a fixed bit length which is set in advance and is difficult to compress is separated from a variable length coding target. At this time, the fixed bit length differs for each content depending on the bit precision of the A / D converter, the conditions at the time of audio recording, and the like, and therefore, it is necessary to set an optimal value as needed. Further, it has been found that the optimum fixed bit length changes due to a change in the signal amplitude even in the content, and there is a problem that the compression effect is reduced if the processing is performed uniformly.
[0009]
In view of the above points, the present invention is capable of performing more efficient compression by selecting an optimal method at any time in accordance with a change in signal characteristics in a time-series signal. It is an object of the present invention to provide a time-series signal encoding method and apparatus capable of completely decoding an original time-series signal at the time of decoding.
[0010]
[Means for Solving the Problems]
In order to solve the above problem, the present invention provides a coding method for compressing the amount of information so that all the sample sequences can be reproduced with respect to a time-series signal composed of a time-series sample sequence. Prediction error calculating step of calculating a plurality of prediction error values based on a plurality of prediction calculation formulas from a temporally past sample sequence with respect to each sample, and a prediction calculation formula of a prediction error value of a past sample sequence An encoding target error selecting step of selecting one as the encoding target prediction error value from a plurality of prediction error values calculated in the prediction error calculating step based on the cumulative error that is another accumulated value; A variable length coding step of performing coding with a variable length code on an error sample having a prediction error value selected as the coding target, the prediction error calculation step, a coding target error selection step, a variable length coding step; Mark Characterized in that the reduction step to be executed for all samples.
[0011]
According to the present invention, for each sample, a plurality of prediction error values are calculated based on a plurality of prediction calculation formulas, and a prediction error of a calculation formula in which the cumulative prediction error for each calculation formula with respect to a past sample sequence is minimized is calculated. The prediction error is selected as the prediction error of the sample, so that the prediction error is calculated based on the optimal calculation formula according to the change of the signal waveform with the change of time. It is possible to do.
[0012]
BEST MODE FOR CARRYING OUT THE INVENTION
Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.
(Configuration of time series signal)
First, a time-series signal to be encoded in the present invention will be described. Here, a case of an audio signal having a plurality of channels as a time-series signal will be described as an example. First, an analog audio signal that is a time-series signal is digitized. This can be done by using a conventional general PCM technique, sampling this analog audio signal at a predetermined sampling frequency, and converting the amplitude into digital data using a predetermined number of quantization bits. In the present embodiment, a description will be given below on the assumption that positive and negative signs are recorded at a sampling frequency of 44.1 KHz and a quantization bit number of 16 bits. Sampling at a sampling frequency of 44.1 KHz results in a sample sequence composed of 44100 samples per second. FIG. 1 schematically shows a predetermined number of samples from the top. In FIG. 1, the horizontal axis represents time t, and the vertical axis represents amplitude x (t). In FIG. 1, the numbers t-4 to t + 5 indicate the sample numbers of the respective samples. When sampling at 44.1 KHz as described above, these 10 samples correspond to 1/4410 seconds. Here, the sample number and the time are used synonymously. This is because the sample numbers are assigned in ascending order as the time changes, and the value obtained by multiplying the sampling frequency by the time becomes the sample number. In FIG. 1, the line segment extending from each sample number indicates an amplitude value. When the amplitude value is 16 bits and a positive / negative sign is recorded as described above, the amplitude is −32768 to 32767. Value.
[0013]
(Outline of processing)
Next, the encoding process for the time-series signal as described above will be described. First, adaptive linear prediction coding is performed. The processing outline of the adaptive linear prediction coding is shown in the flowchart of FIG. In FIG. 2, first, a linear prediction error corresponding to each prediction formula is calculated using a plurality of prediction formulas prepared in advance (step S1). Specifically, the following [Equation 1] to [Equation 4] are prepared as prediction calculation equations for calculating the prediction error of the sample number t.
[0014]
[Formula 1]
e1 (t) = x (t) -x (t-1) -e1 (t-1) / 2
[0015]
[Formula 2]
e2 (t) = x (t) −2 × x (t−1) + x (t−2) −e2 (t−1) / 2
[0016]
[Equation 3]
e3 (t) = x (t) −3 × x (t−1) + 3 × x (t−2) −x (t−3) −e3 (t−1) / 2
[0017]
[Equation 4]
e4 (t) = x (t) −4 × x (t−1) + 6 × x (t−2) −4 × x (t−3) + x (t−4) −e4 (t−1) / 2
[0018]
In the above [Equation 1] to [Equation 4], e1 (t) to e4 (t) are prediction errors in the sample at time t by the respective prediction calculation expressions, and x (t) to x (t-4) are time It is an amplitude value from t to t-4.
[0019]
“2 × x (t−1) −x (t−2)” in the above “Formula 2” and “3xx (t−1) −3 × x (t−2) + x ( t−3) ”, and“ 4 × x (t−1) −6 × x (t−2) + 4 × x (t−3) −x (t−4) ”in the above [Equation 4] is the past 2 Linear prediction component based on ~ 4 samples. The prediction error at time t using this linear prediction component and the prediction errors “e1 (t−1) / 2” to “e4 (t−1) / 2” (error feedback components) calculated in the immediately preceding sample. e1 (t) to e4 (t) are calculated.
[0020]
Next, a linear prediction error that minimizes the cumulative error, which is the accumulation of the absolute values of the prediction error values for each of the above-described prediction formulas, is selected as the prediction error of the sample (step S2). Here, the concept of accumulated error is used. Specifically, the cumulative values of the prediction errors calculated by the prediction calculation formulas [Formula 1] to [Formula 4] for the past samples are set as R1 to R4. Then, a prediction error corresponding to the smallest one of the accumulated errors R1 to R4 is selected. For example, it is assumed that R2 is the minimum among R1 to R4. In this case, the prediction error e2 (t) calculated by [Equation 2] is selected as the prediction error e (t) to be encoded. The selected prediction error e (t) is replaced with the original value x (t) of the sample, and the subsequent processing is performed. A sample in which the prediction error e (t) is recorded to be distinguished from the original sample will be referred to as an error sample.
[0021]
Subsequently, the absolute values of the prediction errors e1 (t) to e4 (t) are added to the accumulated errors R1 to R4 (step S3). Specifically, as shown in the following [Equation 5], the variables R1 to R4 serving as the accumulated error values are updated. At the same time, every time the processing of each sample is performed, the processing of incrementing the counter by one is performed.
[0022]
[Equation 5]
R1 ← R1 + | e1 (t) |
R2 ← R2 + | e2 (t) |
R3 ← R3 + | e3 (t) |
R4 ← R4 + | e4 (t) |
[0023]
Subsequently, it is determined whether the counter has exceeded a predetermined number of times (step S4). In the present embodiment, the predetermined number is set to 100 times. That is, it is determined whether the counter has exceeded 100.
[0024]
As a result, if the counter exceeds 100, the accumulated error is halved (step S5). Specifically, as shown in the following [Equation 6], variables R1 to R4 that are cumulative errors are divided by two. At the same time, the counter is reset to zero. That is, R1 to R4 are not cumulative errors in a pure sense, but are moving averages of the cumulative errors. In the present embodiment, up to the immediately preceding maximum of 100 samples are accumulated, but the previous samples are processed so as to be halved. Thereby, the influence of the samples separated in time is reduced.
[0025]
[Equation 6]
R1 ← (R1) / 2
R2 ← (R2) / 2
R3 ← (R3) / 2
R4 ← (R4) / 2
[0026]
By executing the processing of steps S1 to S5 over all samples at all times in the time-series signal, the values of all samples are replaced from the original amplitude value x (t) with the target error e (t). become.
[0027]
Subsequently, positive / negative processing is performed on each error sample (step S6). The value of each sample is replaced from the amplitude value to the prediction error by the processing of the above steps S1 to S5, but the bit format of each sample remains unchanged. Normally, when operated by a computer such as a computer, each data is processed in units of 32 bits and is expressed using a two's complement representation. This is converted into a positive / negative signed absolute value expression, and the absolute value portion is shifted one bit higher, and the positive / negative sign bit is shifted to the LSB (least significant bit). FIG. 3 schematically shows how the bit configuration is converted in step S6. FIG. 3A shows a bit configuration before processing, and FIG. 3B shows a bit configuration after processing. The reason why the positive / negative sign bit is shifted to the LSB is to make it easier to detect the bit length of each error sample in the processing of step S7 and later (variable length coding) described later.
[0028]
From here on, the process of converting each error sample into a variable length is performed. The bit length conversion in the present embodiment employs a method generally called Golomb coding. More specifically, the bit components forming one sample are divided into upper bit components and lower bit components, the lower bit components are left unchanged, and the upper bit components are the numerical values obtained by converting only the upper bits into decimal numbers. Bits “0” are arranged, and a separator bit “1” is added at the end. For example, consider an 8-bit bit component "00101000". At this time, if the lower bit component is 4 bits, the lower bit component is “1000”. Since the upper bits are "0010", "2" pieces of "0" obtained by converting this number into decimal numbers are arranged, and are converted into "001" obtained by adding "1" at the end. As a result, the 8-bit bit string “00101000” is converted into a 7-bit bit string “00111000”. In the present embodiment, the bit length of the lower bit component that makes the bit component invariable before and after the conversion is made variable in each sample.
[0029]
Hereinafter, a specific description will be given. FIG. 4 is a flowchart showing an outline of the variable length coding. First, an average bit length Bf, which is a moving average of the bit lengths of past samples, is calculated (step S7). The average bit length Bf is obtained by dividing the accumulated bit length RB, which is the accumulated value of the past bit length, by a counter C based on the number of past samples. That is, it is calculated by Bf = RB / C. Since the accumulated bit length RB is 0 in the initial state, when processing a sample at t = 1, the bit length Bd (t) of the sample at t = 1 is set as an initial value. Also, the initial counter C is set to 1.
[0030]
Subsequently, the bit length Bd (t) of the sample at time t is calculated (step S8). For the samples after t = 2, after calculating the average bit length Bf, the bit length Bd (t) of the sample is calculated. The bit length Bd (t) is easily calculated by performing the conversion of the bit configuration as in step S6. As a result of the conversion into the bit configuration as shown in FIG. 3B, the bit length starts from the first bit "1" in the bit configuration of each sample. Next, the bit length Bv of the changing unit is calculated (Step S9). This is calculated by subtracting the average bit length Bf from the bit length Bd (t) of the error sample. Subsequently, the code output of the data is performed (step S10). Specifically, after outputting “0” by the numerical value obtained by converting the upper Bv bits into a decimal number, a separator bit “1” is output, and the lower Bf bits are output as an invariable part. The code output is performed as recording on an external storage device such as a hard disk or a CD-R. Next, the bit length Bd (t) is added to the accumulated bit length RB (step S11). At the same time, each time the processing of each error sample is performed, the processing of adding the counter C one by one is performed. Subsequently, it is determined whether the counter C has exceeded a predetermined number (step S12). As the predetermined number, about 100 is set here as well. Therefore, it is determined whether the counter has exceeded 100. As a result, if the counter exceeds 100, the cumulative bit length RB is halved (step S13). Specifically, the variable RB that is the accumulated bit length is divided by two. At the same time, the counter C is halved.
[0031]
As described above, encoding with a variable bit length is performed for each sample. The variable-length error samples obtained by the encoding are recorded on a target recording medium as code data. Note that the processing of steps S1 to S13 has been described separately for adaptive linear prediction of FIG. 2 and the variable length coding of FIG. 4 for convenience of description, but actually, the processing of steps S1 to S13 is performed for each sample. Are performed in parallel. That is, as shown in FIG. 1, instead of returning to step S1 after the processing of steps S4 and S5, the processing returns to step S1 after performing the processing of steps S6 to S13.
[0032]
(Combination with other compression methods)
The present invention can sufficiently enhance the compression effect even with only the above description, but a higher effect can be obtained by combining with another compression method. Hereinafter, preferred combinations will be described. Here, a description will be given assuming that processing is performed on a stereo sound signal as shown in FIG. FIG. 5A shows a two-channel stereo sound signal, in which an L (left) signal is recorded in Ch1 and an R (right) signal is recorded in Ch2. 5A to 5D, the left end is the start time and the right end is the end time. Compared to FIG. 1, the time intervals on the horizontal axis are shown in a condensed form. FIG. 6 is a flowchart showing an outline of the entire processing when another method is combined with the present invention. From now on, description will be made according to FIG.
[0033]
(Signal flat part processing method)
First, processing of a signal flat portion is performed on a sample sequence that is a digital acoustic signal (step S21). The signal flat portion refers to a portion where the same signal level is continuous. In particular, it often occurs in a silent part where the signal level is “0” and in a saturated part where the absolute value of the signal level is the maximum. Silence occurs when the sound is actually silent or when the sound is not recorded very small, while saturation occurs during the signal recording and A / D conversion. Regardless of whether a silent portion, a saturated portion, or the other same signal level continues, the same signal level is continuously recorded for a predetermined time (a predetermined number of samples) in the signal flat portion. For this reason, this part is data that can be easily compressed. Specifically, the three values of the head time position of the signal flat portion, the number of samples following the same signal level, and the signal level (sample value) are recorded as signal flat portion data separately from the sample sequence of each channel. I do. The signal flat portion is deleted from the sample sequence of each channel. This is schematically shown in FIGS. 5B and 5C. FIG. 5B shows a sample sequence before the signal flat portion processing. In FIG. 5B, a shaded portion indicates a signal flat portion. By the processing in step S21, the signal flat portion is deleted from the original sample sequence, and the result is as shown in FIG. However, in order to restore the original state at the time of decoding, the separated signal flat part is recorded in a format as shown in FIG.
[0034]
As described above, the signal flat portion data is recorded for each signal flat portion with three attributes of the start time (sample number), the number of samples, and the sample value. Here, the head time is the time from the start position of the signal, and in the example of FIG. 5E, it is recorded with the sample number from the head. As described above, if the sample number is divided by the sampling frequency, it will be converted to time. The sample number is information indicating how continuous the sample value continues. The end time of the signal flat portion may be recorded instead of the number of samples. The sample value indicates the digitized signal level. Here, since quantization is performed with 16 bits, the maximum value is “32767” and the minimum value is “−32768”. That is, “0” indicates a silent part, and “32767” and “−32768” indicate a saturated part. However, the signal flat portion is not unconditionally processed. Here, since the purpose is to compress data, it is meaningless if the signal flat portion data is larger than the reduction of the sample sequence. Therefore, only when the sample which becomes the signal flat portion continues for a predetermined number or more, the signal flat portion data is created and separated from the sample sequence of each channel.
[0035]
Subsequently, a conversion process from the original sample value to the prediction error is performed on each sample (Steps S1 to S5). This converts the value of each sample into a prediction error by executing the processing shown in the flowchart of FIG. When there are a plurality of channels, processing is performed on the sample sequence of each channel.
[0036]
(Calculation method between channels)
Next, a difference calculation between channels is performed on the error sample sequence of each channel in which the prediction error value is recorded (step S22). This is performed by simply taking the difference between the error sample values at the same time. The result of the difference operation is given as an error sample sequence of one channel, and the value of the error sample sequence of the other channel is left as it is. Specifically, in the case of a two-channel stereo sound signal as shown in FIG. 5C, the value of the L signal is recorded as it is for Ch1, and the RL difference value is given to Ch2. Generally, in a stereophonic signal, each data at the same time has a correlation, and the difference between the two data at each time is a smaller value than the original value. This is the same when predictive coding is performed by linear prediction. For this reason, in the example of FIG. 5D, the value of each error sample in Ch2 becomes small, and there is more room for subsequent compression.
[0037]
(Inter-frame operation method)
Subsequently, a frame having a predetermined section length is set for the error sample sequence of each channel on which the inter-channel operation has been performed, and the operation between the set frames is performed (step S23). The similarity of the error sample sequence constituting each frame is obtained, and similar frames are selected. Here, the frame length is fixed over the entire section from the start time to the end time of the error sample sequence. Specifically, one frame has 256 samples. 256 samples from the beginning of the sample sequence are extracted as one frame, and the similarity of each frame is determined. Since the similarity between frames means that the correlation between both signals is obtained, various methods for performing the correlation calculation can be used. Here, the difference between the 256 samples corresponding to each frame is calculated. Calculate and calculate the maximum value of each absolute value. Here, a difference absolute value that maximizes each of the 100 subsequent frames to the basic frame is calculated, and a frame whose maximum value is equal to or less than a predetermined value is selected as a correlation frame, and one group is formed with the basic frame. Will be. This process is performed over the entire section of the error sample sequence. FIGS. 7A to 7C show how the error sample sequence changes in the process of step S23. Although only one channel is shown in FIG. 7 unlike FIG. 5, the other channels are processed similarly. First, as shown in FIG. 7A, the error sample sequence framed to a fixed length is divided into frames F1, F2, F3,... Fm,.
[0038]
Subsequently, differences are calculated for a plurality of frames subsequent to one basic frame. First, a difference is calculated for each sample in the first frame F1 and the next frame F2. In this example, 256 difference values are obtained for each sample time. The maximum absolute value of the obtained difference value is recorded as an index value indicating the correlation with the F1 frame in the F2 frame. Similarly, the maximum value of the absolute difference between the F3 frame and the F1 frame is determined, and the frame having the smallest maximum value is selected as a correlation frame candidate. For example, when the frame F1 is a basic frame, the maximum difference absolute value of the frame F3 is the smallest, so that the frame F3 is a correlation frame candidate. Then, when the maximum value of the absolute value of the difference is smaller than or equal to the maximum value of the absolute value of each sample value of the frame F3 before taking the difference, the frame F3 is determined to be a correlation frame. , And a group A with the frame F1 which is a basic frame. At this time, the frame F1 remains unchanged, but each sample of the frame F3 is updated to a difference value from the frame F1. In order to indicate the difference value, the processed frame is represented by a frame “F3-F1”. Further, the same processing is performed on the subsequent frames. For example, the frame Fn is determined as a correlation frame with respect to the basic frame Fm, a group G is formed, and a difference process is performed on the frame Fn to obtain a frame “Fn−Fm”. After all, the basic frame in the group remains as it is, and the difference from the basic frame is recorded in the correlation frame in the group.
[0039]
In step S23, frame structure data, which is a relationship between frames, is recorded in parallel with the above-described difference calculation processing. Specifically, information on which frames have been grouped is recorded. The recording of a frame is performed by recording the frame number of each frame. Here, an example of the frame structure data is shown in FIG. As shown in FIG. 7D, the frame structure data is recorded by a group number and each ID number of a basic frame and a correlation frame belonging to the group. This frame structure data is required to faithfully restore the original signal during decoding. In step S23, similar frames are selected, and the correlation frame of each group is recorded by the difference from the basic frame. Since the difference value between similar frames is small, it is possible to represent the difference value with a small number of bits when changing the number of bits to be recorded in the processing described later.
[0040]
Thereafter, the processing of the positive / negative polarity as shown in FIG. 3 is performed (step S6). As shown in FIG. 3, the two's complement representation is converted to a signed absolute value representation, and the sign is shifted to the least significant bit. After that, variable length coding is performed (steps S7 to S13). In this case, the bit length of each sample is coded in a variable length by executing the processing shown in the flowchart of FIG.
[0041]
(Example of prediction error)
Next, the calculation of the prediction error performed in step S1 will be described. Here, a description will be given of a case where [Equation 2], which is predicted by two past samples, is used as a representative. For example, consider a case where the sample value x (t) is in a state as shown in FIG. In FIG. 8A, the horizontal axis represents time (sample number), and the vertical axis represents sample value x (t). The line segment at each time indicates the magnitude of the sample value x (t) at each time. The procedure of calculation using [Equation 2] will be described with reference to FIG. 8. First, each prediction error eo (t) is calculated without adding an error feedback component. As shown in FIG. 8B, when calculating the prediction error eo (t) at the time t, the sample value x (t-1) at the immediately preceding time t-1 and the sample value x (t-1) at the immediately preceding time t-2 are calculated. The prediction error eo (t) is calculated based on the difference between the value taken by the prediction line connecting the value x (t−2) at the time t and the sample value x (t) at the time t (shown by a thick dotted line in the figure). You. The same operation is performed after time t + 1 to calculate the prediction error eo (t + 1). The calculated prediction error eo (t) is as shown in FIG. As can be seen from a comparison between FIG. 8A and FIG. 8C, the range in which the value fluctuates is greatly narrowed, and data compression becomes more convenient.
[0042]
Subsequently, based on [Equation 2], 50% of the prediction error e2 (t-1) obtained by adding the correction at the immediately preceding time t-1 to the prediction error eo (t) is subtracted, and the error feedback processing is performed. FIG. 8D shows the added result. Compared with FIG. 8C, the reduction of the prediction error at the times t + 1 and t + 2 is remarkable. Conversely, the prediction error increases at times t + 3 and t + 4, but the prediction error is reduced on average, and the range in which the value fluctuates is further narrowed as compared with FIG. 8A, and the data compression effect is improved.
[0043]
(Device configuration for realization)
It goes without saying that the above encoding method is actually realized by a computer equipped with dedicated software. By installing a program that causes the computer to execute the above-described processing, the computer reads a time-series signal such as a digital acoustic signal, and then executes the processing of steps S1 to S13 to obtain compressed code data. .
[0044]
【The invention's effect】
As described above, according to the present invention, in performing lossless compression of a time-series signal composed of a time-series sample sequence, a plurality of samples in the sample sequence are sampled from a temporally past sample sequence. Calculates a plurality of prediction error values based on the prediction calculation formula of, and calculates an encoding target Is selected as a prediction error value of the target, and each sample is coded by performing coding with a variable length code on the prediction error value selected as a coding target, thereby calculating a prediction error, and calculating a coding target error. Since the selection and the variable length coding are performed for all the samples, it is possible to perform the compression with higher efficiency.
[Brief description of the drawings]
FIG. 1 is a diagram schematically showing a sample sequence of a digital time-series signal to be processed in the present invention.
FIG. 2 is a flowchart illustrating an outline of an adaptive linear prediction process according to the present invention.
FIG. 3 is a diagram showing a state of positive / negative polarity processing in step S6.
FIG. 4 is a flowchart showing an outline of variable-length coding according to the present invention.
FIG. 5 is a diagram illustrating a state of a signal flat portion process and an inter-channel calculation process.
FIG. 6 is a flowchart showing an outline of a process when another compression process suitable for the present invention is combined.
FIG. 7 is a diagram illustrating a state of an inter-frame calculation process.
FIG. 8 is a diagram showing a state of a prediction error calculation process using [Equation 2].
[Explanation of symbols]
R1 to R4: cumulative error Bv: changing unit bit length Bd (t): error sample bit length Bf: average bit length RB: cumulative bit length

Claims (12)

An encoding method for compressing the amount of information so that all the sample sequences can be reproduced for a time-series signal composed of a time-series sample sequence,
For each sample of the sample sequence, a prediction error calculation step of calculating a plurality of prediction error values based on a plurality of prediction formulas from a temporally past sample sequence,
Based on the cumulative error that is the cumulative value of the prediction error value of the past sample sequence for each prediction calculation formula, from among the plurality of prediction error values calculated in the prediction error calculation step, the prediction error value to be encoded An encoding target error selecting step of selecting one as
A variable-length encoding step of performing encoding with a variable-length code on an error sample having a prediction error value selected as the encoding target,
Encoding the time-series signal by performing variable-length encoding on all samples of the time-series signal by performing the prediction error calculating step, the encoding target error selecting step, and the variable-length encoding step on all the samples. Method. In claim 1,
A method for encoding a time-series signal, wherein a plurality of prediction calculation formulas in the prediction error calculation step are provided depending on a difference in the number of past samples to be referred to. In claim 1 or claim 2,
In the variable length encoding step, among the bit components of each of the error samples, the lower bit component is encoded with the same bit component, and the encoding is performed by changing the bit component with respect to the remaining upper bit components. A time-series signal encoding method, characterized in that: In claim 1 or claim 2,
The variable length encoding step determines an invariant bit length based on a moving average of the bit lengths of the past error samples, and the lower invariant bit length of the bit components of each error sample is used as it is. A coding method for a time-series signal, wherein coding is performed using a component, and coding is performed by changing a bit component with respect to the remaining higher-order bit components. In any one of claims 1 to 4,
A polarity processing step of shifting the entirety of the absolute value of the prediction error value selected in the encoding target error selecting step by one bit higher and inserting a positive / negative bit into the least significant bit is performed before the variable length coding step. A method for encoding a time-series signal, which is provided as stages. In any one of claims 1 to 5,
The encoding method of a time-series signal, wherein the accumulated error is calculated by using the accumulation of absolute values of a predetermined number of prediction error values in the past. In any one of claims 1 to 6,
In the sample sequence, a signal flat portion in which the value of the sample is continuously the same is extracted, separated from the sample sequence, and the leading time position of the separated sample, the number of samples, and the sample value A time-series signal encoding method, characterized in that a signal flat portion processing step of coding three values as signal flat portion data is performed as a preceding stage of the prediction error calculation step. In any one of claims 1 to 6,
When the sample sequence is composed of a plurality of channels having a plurality of values at the same time, a predetermined operation is performed on the error sample sequence between the channels to update the error sample sequence of any of the channels. A method for encoding a time-series signal, wherein an inter-operation step is performed between the encoding target error selecting step and the variable length encoding step. In any one of claims 1 to 6,
A plurality of frames composed of a predetermined number of error sample sequences are extracted from the error sample sequence, a predetermined operation is performed between the extracted frames, and each prediction error value of one frame is calculated as a calculated value. A method for encoding a time-series signal, wherein an inter-frame operation to be updated is performed between the encoding target error selecting step and the variable length encoding step. A coding apparatus that compresses the amount of information so that all the sample sequences can be reproduced with respect to a time-series signal including a time-series sample sequence,
For each sample of the sample sequence, a prediction error calculation unit that calculates a plurality of prediction error values based on a plurality of prediction calculation formulas from a temporally past sample sequence,
Cumulative error calculation means for calculating a plurality of cumulative prediction error values corresponding to the plurality of prediction error values,
Encoding target error selecting means for selecting one as a coding target prediction error value from among a plurality of prediction error values calculated by the prediction error calculating means, based on the cumulative prediction error value;
A variable-length encoding unit that performs encoding with a variable-length code on an error sample having a prediction error value selected as the encoding target,
An encoding device for a time-series signal, comprising: A recording medium on which code data encoded by the encoding method according to claim 1 is recorded. For a time-series signal composed of a time-series sample sequence, an encoding program that compresses the amount of information so that all the sample sequences can be reproduced,
On the computer,
For each sample of the sample sequence, a prediction error calculating step of calculating a plurality of prediction error values based on a plurality of prediction formulas from a temporally past sample sequence, each of the prediction error values of the past sample sequence An encoding target error that selects one as a prediction error value to be encoded from among a plurality of prediction error values calculated in the prediction error calculation step based on a cumulative error that is a cumulative value for each prediction calculation formula. A selecting step, performing a variable length coding step of performing coding with a variable length code on the prediction error value selected as the coding target, and calculating the prediction error, a coding target error selecting step, and a variable length coding step. A program that causes the encoding step to be performed on all samples.

Images