JPS5918719B2 - Data compression method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a data compression method useful for data having strong correlation between adjacent sample points, such as audio signals and video signals.

Generally, when an audio signal is sampled, digitized, and recorded, reproduced, or transmitted, it is necessary to quantize it with an accuracy of about 8 bits per sampling at a sampling period of 8 kHz.

This requires an information processing capacity of 64 ld) it per second, so in order to save the capacity of the recording device and narrow the bandwidth of the communication line, various types of data have been used to reduce the amount of information per second. Compression is being attempted. For example, in recent years,
A PARCOR method, etc., which extracts characteristic parameters from an information signal through predetermined information processing and synthesizes these parameters to restore and reproduce the information signal, is attracting attention.

According to the extraction of the above feature parameters, the amount of data is 2k per second.
Although it can be significantly compressed to about 4 bits to 4 ld) it, on the other hand, it requires special signal processing to extract and synthesize the feature parameters, resulting in a considerable complexity in the device configuration. Furthermore, if we want to achieve the naturalness and clarity of the restored and reproduced information, the signal processing becomes extremely complicated.
It would be a big loss, so it was not very practical. others,
Data compression using formant synthesis has also been proposed, but similar problems have arisen. The present invention has been made in consideration of these circumstances, and its purpose is to effectively convert highly correlated information signal data such as audio signals and video signals at nearby sample points through simple signal processing. It is an object of the present invention to provide a data compression method that can compress data in a consistent manner, and can also ensure sufficient fidelity to the original information data.

In other words, the present invention makes effective use of the fact that audio signals, video signals, etc. have strong correlation and periodicity at adjacent sampling points, and that they have selective extraction properties against noise. The aim is to provide a new data compression method that is clear and different from conventional methods.

First, an overview of the data compression method according to the present invention will be explained.

The present invention makes full use of the properties of audio signals, video signals, etc., and the properties are enumerated as follows.

(a) Correlation between adjacent sample points is extremely strong.

(b) The signal data is not random and has a certain degree of periodicity. (e) The data polarity of one sample point and the data polarity of the next sample point are often the same.

(d) There is no problem in information reproduction even if the restored data is not complete.

It has the following properties.

By the way, in the case of electronic computers, even 1-bit data error is often not tolerated, but in the case of audio data, even if there is a bit error, there is no problem because the information is read from the entire audio data. do not have. This is supported by the fact that humans can only hear desired sounds even in a very noisy environment. In other words, when focusing on one piece of desired audio data, the remaining audio data acts like white noise and has no meaning with respect to the desired data.
If the remaining audio data is completely converted into white noise, further improvement in clarity can be expected. Therefore, in the method of the present invention, a plurality of pieces of information data are superimposed and synthesized so that they become random numbers, and this superimposed data is read and reproduced in the reverse procedure of randomization, thereby obtaining desired information from white noise. This extracts data and restores it.

In other words, this method takes advantage of the fact that when the superimposed data is read in a predetermined order, only certain information is found due to its periodicity and other properties, and the remaining components have no meaning due to randomness, etc. . In addition, in the case of such superposition processing, that is, data compression, even if N pieces of data are added randomly, the amplitude will only be J at most. This is used to ensure S/N and improve signal fidelity. We are also striving to improve quality. below,
An embodiment of the system of the present invention will be described with reference to the drawings.

FIG. 1 is a schematic diagram of a signal recording/reproducing apparatus constructed using the data compression method according to the present invention.

An audio signal input from a microphone is sampled at a frequency of, for example, 100 kHz via an input device including a buffer amplifier, a sample circuit, and the like. This sampled signal is led to an A/D converter 1 and quantized with an accuracy of 10 bits per sample. Although the sample frequency and number of quantization bits mentioned above are about 10 times higher than those for a normal audio signal, and the capacity is enormous, the data can be effectively compressed by the data compression process according to the present invention, which will be described later. Ru. Thus, each quantized data at each sample point is guided to a differentiating circuit 2 to obtain change data between each sample point. In this type of digital processing, the differentiating circuit 2 is generally constituted by a differentiator. This difference value is determined to be a sufficiently small value since the sampling frequency is sufficiently high. Also, at this time, it can be determined with high accuracy using a 10-bit 17) A/D converter. The information data of the input audio signal obtained through such signal processing is sequentially written into the storage device 4 which is addressed by the random number address generation circuit 3 in a predetermined random order.

Pseudo-random number generation is performed by P. This is done with a counter having a period of Q. For example, the address M of the Nth data is M=(N+-
remainder of P)×Q+(quotient of N+-P). The above addressing order is reproducible and is initialized by a set signal given to the random number address generation circuit 3.
It has a periodicity that is sufficiently long compared to the audio signal period. Then, in the storage device 4, a predetermined arithmetic processing is performed on a plurality of data specified at the same address, for example, the sum of the plurality of data is determined, and the sum value is stored at the address along with the polarity sign. . Note that if the negative polarity component of the signal data is displayed as a 2's complement, it is convenient for summation calculations, etc. This summation calculation compresses the sample data of the audio signal. Specifically, the sample data for 100 seconds obtained in the above sample period is divided into 1 second units, and is written to the storage device 4 having 105 addresses as described above. When a signal is written by addressing in accordance with the random number order, 100 samples of data will be specified in one address.

These 100 samples of data are subjected to summation calculation processing as described above, and are written together with their polarity codes.
On the other hand, the information data compressed and written to the storage device 4 as described above is read out in the same order as the random number order under addressing control of the random number address generation circuit 3.

At this time, as will be described in detail later, only polarity code data reflecting each data superimposed by data compression is sequentially read out. This randomly read data series corresponds to and reflects the difference data series obtained from the audio signal waveform, which is the original information signal, through the differentiation circuit 2. This read data is input to an integration circuit 5 consisting of an up-down counter, etc., where it is subjected to integration processing and converted into a signal corresponding to the audio signal waveform. Correction circuit 6
performs signal correction by suppressing unnecessary signals such as drift due to errors, and upon receiving the output of this correction circuit 6, the D/A
The converter 7 restores the quantized data to an analog signal. After high-frequency components are removed from the analog restored signal through a filter circuit (not shown), etc., it is supplied from the output device to a speaker, where the audio signal is reproduced. By the way, the compression effect of data written into the storage device 4 by addressing according to a predetermined random number order and its restoration effect will be explained as follows.

Figure 2 schematically shows the effect. In the example mentioned above, there are 100 samples of data for 1 second, but here it is 3 samples, and only 9 unit periods are extracted for simplified representation. . Here, the sampled data that is continuous at a predetermined sampling period is divided into A, B, and so on by one unit time. Add the code C and write the data series as 1
, 2..., 9 are shown with a suffix.

Therefore, the data series of the input audio signal is Al, A2, . . . , A9, Bl, B as shown in Fig. 2a.
2,......,B9,Cl,C2,...
..., shown as C9. These data are randomly assigned and supplied to the addresses of the storage device 4 in accordance with the aforementioned random number order every unit time. For example, for data series A, “1, 2, 3, 4,
The addresses are specified in the order of ``5, 6, 7, 8, 9'', and for data series B, the addresses are specified in the order of ``6, 8, 4, 2, 9, 7''.
, 5, 1'', and for data series C, the order is ``
5, 8, 2, 4, 7, 1, 2, 9'' random number addressing is performed according to a predetermined rule.

In this case, the random number structure is such that data of adjacent sample points are not stored in consecutive addresses. Therefore, the data is distributed as a whole, and the correlation with other data series becomes sufficiently low. Therefore, the data array configuration is as shown in FIG. 2b, and uncorrelated data such as "Al, B, . . . C6" is specified at the same address, for example. Note that for this randomization, it is essential to use a reproducible random number sequence.
Thus, if attention is paid to such a data series arranged in random numbers, data series B and C other than data series A can be regarded as white noise having no random phonemes at all. That is, if we focus on a specific data series, all remaining data series have meaning only as white noise. Therefore, if only a certain data sequence is reproduced by reproducing the data sequence by retracing its random number arrangement order and restoring the signal, all the remaining data will be included in the reproduced signal as white noise, so this is what we want here. It becomes possible to extract only the information signal of the data series. Therefore, predetermined arithmetic processing is now performed on a plurality of data designated at the same address to obtain information S that reflects these data.

For example, information S1 is obtained from data Al, B9, and C6, information S2 is obtained from each address A2, B4, and C3, and information S1 to S9 is obtained for each address. These pieces of information S1 to S are stored in the storage device 4. That is,
When data A, B, and C are positive, they are shown as natural binary numbers, and when they are negative, they are shown as two's complement numbers, and information S
is given as "0" when the polarity is positive or zero, and as "1" when the polarity is negative, by the summation operation of data A, B, and C. In other words, the sign bit data of the total value is the information S.
used as. Also, according to such arithmetic processing,
This effectively suppresses data values from becoming extremely large due to noise components that are unnecessary data. For example, the signal level of the original signal waveform is 1, and when 100 channels are synthesized, the signal level is expected to reach 100V, but according to the arithmetic processing related to this method described above, only positive and negative polarities can be determined. Therefore, extremely large components are automatically saturated, so no inconvenience occurs. In other words, if information "0" is made to correspond to 1v, no matter how many volts the total data level of the data series of 100 channels is, the data will never exceed 1. Note that how many volts correspond to the information "0" and "1" can be determined according to the specifications, and the total level is not affected by an increase or decrease in the number of channels. By the way, when viewed from the data series of each channel, the data written to the storage device 4 as data compression information is given as either "0" or "1".

In addition, the data of another channel is the data “O” of the above information S.
” and “1” is expressed as the probability that the same data “0” and “1” will not appear correctly. According to another expression, the input audio signal (in the device shown above, the differential value shown as a voltage value) is the one in which "1" and "0" appear correctly in the arithmetic processing for obtaining the information S. It has the meaning of probability. Therefore, for example, in an operation to obtain information S1, if the absolute value 1A11 of data A1 among data A, , B, , C6 is large, the above information S1 is drawn into data A1 and set to the polarity sign side of 2A1. The probability of Therefore, in this case, information S1
is the one that best reflects the data A1. Naturally,
Due to the nature of the audio signal, the information S of nearby sample points may be taken into consideration to compensate for this and provide the signal for reproduction. J
Thus, the information Sl obtained by such arithmetic processing
, S2, to, S, are respectively stored at each address of the storage device 4 as the series of data shown in FIG. 2c. Thus, the data stored in the storage device 4 is transferred to the second
It is read out repeatedly as shown in Figure d.

That is, since the data at each address is a convolution of each data constituting the data series of a plurality of channels, it is read out repeatedly four times corresponding to the number of channels. Therefore, for data series A, information Sl, S
Data is read out in a sequence of 2, S3, S4, S5, S6, S7, S8, S, followed by information S9, S4,
S8, S3, S7, S,, S6, S2, S5, then information S6, S3, S7, S4, Sl, S8, S5, S2,
S, will be read out in order. Therefore, if we focus on the above information series Sl, S2, ~, S, this information series reflects data series A, and the data of data series B and C convolved with the same information are distributed respectively. Since the information is converted into white noise, it is possible to ignore this and extract and reproduce only the information of the data series A. Furthermore, regarding the information series of the second channel, it reflects the data system JIB, and the data components of data series A and C are converted into white noise. Similarly, if we look at the information series of the third channel, it reflects data series C, data series A,
The data component of B is converted into white noise. Therefore, data can be reproduced very effectively from the data-compressed information series S. Incidentally, since the address designation according to the random number order of data reading is performed in the same order as when compressing the data, it goes without saying that the random number address generation circuit 3 generates a reproducible random number pattern. In this case, since the same address will be accessed 100 times corresponding to 100 channels, if 17 bits are set as the number of bits for each address specification, the accuracy of 10 bits can be reflected for each channel. However, in reality, 16 bits is sufficient because there is a margin for non-voice periods, etc. By the way, the data stored in the storage device 4 in which 100 channels of data have been compressed is the data of one channel of interest and the remaining 99 channels of random noise data superimposed.

For this reason, the error due to the noise increases in proportion to the % power of the number of data compression channels, and therefore there is a possibility that the number of effective bits of D/A conversion for data reproduction may be exceeded. Therefore, in the correction circuit 6, when the output of the integrating circuit 5 consisting of an up-down counter is positive, it is multiplied by a preset coefficient and then subtracted by the counter, and conversely, when it is negative, it is added. As a result, the correction output is compensated so that it always approaches the zero level, and the error is suppressed. The D/A converter 7 has, for example, a 16-bit structure, and the remaining 6 bits for the 10 bits required for actual conversion are provided to allow for the above-mentioned error drift. It goes without saying that the distortion generated by the correction circuit 6 can be suppressed to a level lower than the error due to the above-mentioned noise. Since the white noise component is a random number of 100 kHz, it is almost inaudible. As described above, according to the device configured by applying this method, 100 seconds worth of data can be compressed into 1 second worth, and the above-mentioned 100 seconds worth of data can be expressed with at most 105 bits. possible.

Therefore, it is possible to significantly reduce the capacity by setting the data rate to 1 kiIt/Sec. Moreover, as mentioned above, the compression algorithm is extremely simple and has the advantage of being excellent in naturalness. Next, the S/N by this method described above
I will explain about it.

The following table shows the possible values of data A, B, and C.
, 0 only to indicate the coincidence and mismatch with the information S generated thereby. When the probability that A=S is 50%, it means that information S does not reflect data A at all; conversely, when the probability that A=S is 100% or 0%, information S greatly reflects data A. It can be said that it reflects well. In other words, it is no exaggeration to say that in this case, data A is completely represented by information S. However, now the information S generated by the data A, B, and C shown in the table above
If you think about it simply, data A, B,
Since the degree of contribution of C is the same, the probability of A-S is as high as 75%.

This also applies to data B and C. However, if we consider this further, if we follow the binomial theorem for 100 channels of data, the probability of A-S will be:

Therefore, the probability of A\S is 45%. This 45%
This probability arises because the data of 99 channels out of 100 channels contributes as randomized noise, and this property is actively used to perform data compression according to the present invention. . So now,
If we consider only the information between 50% and 100% as a signal, then for 100 channels, the signal will be 1/J.
It is below h, that is, 0.1. Also, if the frequency of the audio signal waveform is 4kHz, this signal component is 100
The data sequence represented by the kHz sampling points will be represented by 25 points. These two
The error amount caused by sampling data at five points, that is, the random walk component, is JΣτ=5 dots. On the other hand, since the signal portion is 0.1 out of the 25 points, it becomes 2.5 dots, and therefore the S/N becomes 0.5. However, in reality, the proportion of the data signals of each channel contributing to the information S is not equal, and is generally smaller than the channel of interest, so the S/N ratio can be improved beyond the above-mentioned value. be done. That is, the first thing that can be easily considered is that the silent period included in the audio signal is generally quite long, and the information S of the data focused on is due to this period.
It can be considered that the contribution rate is high. Furthermore, when analyzing the waveform of the audio signal, the waveform amplitude is large at the beginning of each pitch, and gradually becomes smaller. Therefore, it is considered that the amount of information existing in the latter half of the audio pitch is small, and the contribution rate to the information S is correspondingly small. In addition, the high-intensity portion of the waveform of the audio signal has a high contribution rate to the information S and also has a large S/N ratio, so that it is useful for understanding the phoneme as a whole. Considering these characteristics of audio signals, the S/N ratio mentioned above is at most 0.
．． It is not about 5, but easily 1 or more, which is sufficient for signal restoration and reproduction.

Furthermore, the above discussion was conducted on a 4kHz audio signal component, but it is the low frequency of about 100Hz that contributes even more to the determination of phonemes, so if we consider this in the same way, we will find that sampling The number is 103 and its random walk error is 32, while the signal component is 2 × 1
It appears as a clear signal as 03=100. Therefore, it becomes possible to recognize voice signals extremely clearly. In other words, signal reproduction with good S/N is possible for low frequency components. Furthermore, although the S/N ratio is poor for high frequencies, the appearance rate and contribution rate of the signal are also high, so there is no risk that the component will be lost so much. Therefore, effective data restoration and reproduction of audio signals is possible. By the way, in the explanation up to now, an example has been explained in which 100 channels of continuous audio signals, that is, 100 seconds are compressed into a 1 second signal.
Of course, it is also possible to compress data by superimposing 100 independent channels of audio signals onto one channel. That is, as shown in FIG. 3 a to d, data of independent channels is extracted into a basic unit time set to 0.1 seconds, etc.
These are randomly arranged and captured as shown in FIG. Thereafter, a summation operation is performed on each of these data and temporarily stored as shown in FIG. 3c, and this is communicated or stored. Thereafter, data reproduction is performed by reading out the same data according to the aforementioned random number arrangement order as shown in FIG. 3d. In this way, real-time processing of data compression and playback is possible, and the processing time is at most 2 units of time (0.2
(seconds) signal delay. Moreover, it is possible to simultaneously communicate a large amount of information (for 100 channels) with a small signal capacity.

Furthermore, since the signal delay described above is smaller than the delay that occurs when making an international call, there is no risk of causing any practical problems. Furthermore, since the basic unit time can be set even shorter, communication by data compression is possible without causing any inconvenience. Note that the present invention is not limited only to the above embodiments.

For example, the signal to be data compressed may have frequency components separated by frequency from the same waveform. Further, even if the number of data given to the same address is not evenly distributed but is distributed unevenly according to a simple random number order, no inconvenience will occur. Furthermore, it can be applied not only to audio signals but also to video signals and character/pattern information. Furthermore, the form of the information S can be changed in various ways. It is also possible to use binary coded information of 2 bits or more.

Furthermore, the number of channels, the number of bits, etc. may be determined according to specifications, and the differential and integral processing may be omitted and the A/D converted data may be directly compressed. In this case, instead of using only the code, the memory contents can be directly D/A.
Output to. Furthermore, the random number generation method can be modified in various ways, such as taking the central part of multiplication, and in short, it is sufficient as long as it is reproducible when specifying an address. In short, the present invention can be modified in various ways without departing from the spirit thereof, and can be widely applied to and implemented in various devices.

[Brief explanation of drawings]

FIG. 1 is a schematic configuration diagram of a signal recording and reproducing apparatus constructed by applying an embodiment of the present invention, FIGS. 2 a to d are diagrams showing data compression and its restoration operation, and FIGS. 3 a to d FIG. 7 is a diagram showing data compression and restoration operations of another embodiment. 1...A/D converter, 2...Differentiating circuit, 3...Random number address generation circuit, 4...
・Storage device, 5...Integrator circuit, 6...
Correction circuit, 7...D/A converter.

Claims (1)

[Claims] 1. Memory addresses are sequentially specified in a predetermined order for sampled data having a strong correlation at adjacent sampling points, and the plurality of sampled data specified at the same address are A data compression method characterized in that a predetermined arithmetic processing result is written to the specified address, and the arithmetic processing result written to each address of the memory is sequentially read out in the aforementioned order to restore the data. 2. The data compression method according to claim 1, wherein the arithmetic processing result of the plurality of sampled data written to the specified address is a polarity code of the total value of the plurality of sampled data. 3. The sampled data is given to the memory as difference information of the original information data, and the data read out from the memory and restored is obtained by cumulatively adding the calculation results. The data compression method described in scope 1. 4 Claims in which the predetermined order is a pseudo-random number that has a sufficiently long repetition period for a plurality of sampled data sequences, and has randomness for sampled data specified at the same address. The data compression method described in Section 1.