WO2018133043A1 - 量化器与量化方法 - Google Patents

量化器与量化方法 Download PDF

Info

Publication number
WO2018133043A1
WO2018133043A1 PCT/CN2017/071981 CN2017071981W WO2018133043A1 WO 2018133043 A1 WO2018133043 A1 WO 2018133043A1 CN 2017071981 W CN2017071981 W CN 2017071981W WO 2018133043 A1 WO2018133043 A1 WO 2018133043A1
Authority
WO
WIPO (PCT)
Prior art keywords
quantization
threshold
prediction
circuit
source data
Prior art date
Application number
PCT/CN2017/071981
Other languages
English (en)
French (fr)
Inventor
牛凯
董超
郭佳
蒋冰
周德坤
陈广进
孙德福
Original Assignee
华为技术有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 华为技术有限公司 filed Critical 华为技术有限公司
Priority to PCT/CN2017/071981 priority Critical patent/WO2018133043A1/zh
Priority to CN201780056167.1A priority patent/CN109690673B/zh
Publication of WO2018133043A1 publication Critical patent/WO2018133043A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G10—MUSICAL INSTRUMENTS; ACOUSTICS
    • G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
    • G10L19/00—Speech or audio signals analysis-synthesis techniques for redundancy reduction, e.g. in vocoders; Coding or decoding of speech or audio signals, using source filter models or psychoacoustic analysis
    • G10L19/02—Speech or audio signals analysis-synthesis techniques for redundancy reduction, e.g. in vocoders; Coding or decoding of speech or audio signals, using source filter models or psychoacoustic analysis using spectral analysis, e.g. transform vocoders or subband vocoders
    • G10L19/032—Quantisation or dequantisation of spectral components
    • G10L19/035—Scalar quantisation

Definitions

  • the present invention relates to the field of communications technologies, and in particular, to a quantizer and a quantization method.
  • Predictive coding is based on this idea.
  • the predictive coding does not directly encode the source data, but predicts the source data and encodes the difference between the predicted value and the source data.
  • DPCM Different Pulse Code Modulation
  • Its working principle is as follows: predict the value of the currently input source data based on historical source data. And calculating the difference between the predicted value and the actual source data, and quantizing and encoding the difference and then transmitting the digital signal.
  • CPRI Common Public Radio Interface
  • Data compression techniques that reduce the transmission load on the CPRI interface, including time domain compression schemes (such as downsampling rate, nonlinear quantization, quadrature modulated signal data compression, etc.) and frequency domain compression schemes (such as subcarrier compression).
  • the starting point of the frequency domain compression scheme is that the source signal has a large amount of redundancy in the frequency domain; however, the frequency domain compression may result in complex constellation mapping, increase interface logic design and processing complexity, and is not achievable.
  • a representative time domain compression scheme is based on the characteristics of the source signal, which eliminates redundancy by interpolation, low-pass filtering and down-sampling, and also greatly reduces the number of symbols.
  • the data is segmented and scaled to ensure that the same data bit width can represent more signals with larger amplitude.
  • the time domain compression scheme does not make full use of the correlation characteristics between the sources and the distribution law of the source, and the compression ratio has room for improvement.
  • Embodiments of the present invention provide a quantizer and quantization method for compressing a digital signal to reduce a bandwidth required for transmission.
  • a first aspect of the embodiment of the present invention discloses a quantizer, including:
  • each quantization circuit corresponds to a quantization mode
  • An output end of the quantization selection circuit is connected to an input end of the quantization circuit
  • the quantization selection circuit for comparing a first data source e x l L prediction residuals to a threshold to determine whether the segment belongs to L e, e is determined according to the segment belongs to the L e quantizing the target quantization l manner, and the target quantization circuit E l is output to the target for performing the quantization mode; the said X E l l l the predicted value X Difference, said a prediction value obtained by linearly predicting the x l ; the e l reaches the quantization selection circuit via an input of the quantization selection circuit; the e l reaches the output via an output of the quantization selection circuit Target quantization circuit;
  • the target quantization circuit is configured to quantize the e l to obtain a first quantization result u l .
  • the prediction residual is segment-quantized, that is, the matching quantization method is selected for each segment according to the distribution rule of the prediction residual, and the data compression rate can be improved under the premise of satisfying the requirement of the quantized signal to noise ratio. , that is, reduce the bandwidth required for transmission.
  • the quantizer further includes a threshold calculation circuit
  • An output end of the at least two quantization circuits is connected to an input end of the threshold calculation circuit; an output end of the threshold calculation circuit is connected to an input end of the quantization selection circuit;
  • the threshold value calculating circuit according to a second source data x l-1, x l-1 the prediction residuals e l-1, e l-1 of the second quantization result and the target u l-1 quantization noise ratio SNR t determining the threshold, x l-1 of the source data is a time prior to said x l; e l-1 is the x l-1 and the x l-1 of the Predictive value Difference, said a predicted value obtained by linearly predicting the x l-1 ; wherein the x l-1 , the e l-1 , the u l-1, and the SNR t are subjected to the threshold calculation circuit
  • the input arrives at the threshold calculation circuit; the threshold reaches the quantization selection circuit via an output of the threshold calculation circuit.
  • the threshold for predicting the residual difference segment is determined according to the requirements of the target quantized signal to noise ratio to ensure that the quantized data can achieve an ideal signal to noise ratio.
  • the threshold is a first threshold group consisting of n-1 values respectively of ⁇ 1 ⁇ n-1 , and the ⁇ 1 ⁇ n-1 will take the prediction residual
  • the value range is divided into n segments; the first threshold group is an average of n ⁇ threshold groups; wherein an interval between every two thresholds of the threshold group is determined according to a distribution of the prediction residuals; the prediction The distribution of residuals is determined based on prior knowledge;
  • the threshold value calculating circuit according to a second source data x l-1, x l-1 the prediction residuals e l-1, e l-1 of the second quantization result and the target u l-1
  • the quantized signal to noise ratio SNR t determines the threshold, including:
  • the prediction residuals e l-1 and the second quantization result u l-1 calculates the transmission end signal to noise ratio Determining whether the SNR q is greater than the SNR t ; if SNR q > SNR t , increasing each threshold in the threshold group by ⁇ ; if SNR q ⁇ SNR t , decreasing each threshold in the threshold group ⁇ ;
  • N ⁇ obtain results set threshold values, averaging the n ⁇ threshold values to obtain the first set of thresholds.
  • the threshold can be dynamically modified according to the signal-to-noise ratio of the transmitting end to ensure that the quantized data can reach an ideal signal-to-noise ratio.
  • the quantizer further includes a subtractor
  • An output end of the subtractor is connected to an input end of the quantization selection circuit
  • the subtractor is configured to compare the first source data x l with the predicted value of the x l Subtracted to obtain the prediction residual x l e l; the Is the predicted value obtained by the linear prediction x l; wherein said x l and the Via the subtractor reaches the input of a subtracter; e l via said output terminal of said subtractor reaches the quantization selection circuit.
  • the transmission source data is converted into a transmission pair source by using correlation characteristics between source data.
  • the data predicts the residual of the prediction, thereby reducing the data bit width of the transmission and increasing the data compression rate.
  • the quantizer further includes a linear predictor
  • An output of the linear predictor is coupled to an input of the subtractor
  • the linear predictor is configured to calculate a predicted value of the first source data x l according to the second source data x l-1 and the second quantization result u l-1 Wherein the x l-1 and the u l-1 reach the linear predictor via an input of the linear predictor; The subtractor is reached via the output of the linear predictor.
  • the source data is linearly predicted by using correlation characteristics between the source data, and when the data is transmitted, the transmission source data is converted into a prediction residual for predicting the source data, thereby Reduce the data bit width of the transmission and increase the data compression rate.
  • the linear predictor includes:
  • the prediction coefficient calculation circuit is configured to calculate a prediction coefficient, where the prediction coefficient includes a first prediction coefficient, and the first prediction coefficient is a prediction coefficient that performs the linear prediction on the first source data x l ;
  • the calculating the prediction coefficient includes:
  • the first prediction coefficient reaches the linear prediction circuit via an output end of the prediction coefficient calculation circuit
  • the linear prediction circuit is configured to predict a value according to the first prediction coefficient and the second source data And the u l-1 calculation include:
  • K is an order of linear prediction
  • the method of calculating the correlation function by segmenting the input source data is performed, and the correlation function is updated in stages, thereby reducing the data complexity of performing correlation function calculation at the data transmitting end.
  • the quantizer further includes a framing circuit
  • An output end of the quantization selection circuit is connected to an input end of the framing circuit; an output end of the target quantization circuit is connected to an input end of the framing circuit; and an output end of the prediction coefficient calculation circuit is connected to the framing frame The input of the circuit;
  • the quantization selection circuit is further configured to obtain identification information of the position of u l; the location identification information u l e l is the segment belongs; u l position of the identification information and the thresholds by An output of the quantization selection circuit reaches the framing circuit;
  • the u l reaches the framing circuit via an output end of the target quantization circuit
  • the first prediction coefficient reaches the framing circuit via an output end of the prediction coefficient calculation circuit
  • the framing circuit is configured to form the data frame of the u l , the location identifier information of the u l , the threshold, and the first prediction coefficient for transmission.
  • the quantization result, the location identification information, the threshold, and the prediction coefficient are combined into a data frame for transmission.
  • the data receiving end can restore the prediction residual of the source data according to the quantization result, the position identification information and the threshold, and use the prediction coefficient to obtain the predicted value of the source data, and predict the residual and The predicted values are added to restore the source data.
  • a second aspect of the embodiments of the present invention discloses a quantization method, including:
  • e l is a predicted value of the x l and the x l Difference, said a predicted value obtained by linearly predicting the x l ;
  • the e l to the threshold to determine the segment belongs e l, e l of the determination of the segment belongs e l quantizing the target quantization According;
  • the e l is quantized by the target quantization method to obtain a first quantization result u l .
  • the method before the comparing the prediction residual e l of the first source data x 1 with the threshold to determine the segment to which the e l belongs, the method further includes:
  • the threshold is a first threshold group consisting of n-1 values respectively of ⁇ 1 ⁇ n-1 , and the ⁇ 1 ⁇ n-1 will take the prediction residual
  • the value range is divided into n segments; the first threshold group is an average of n ⁇ threshold groups; wherein an interval between every two thresholds of the threshold group is determined according to a distribution of the prediction residuals; the prediction The distribution of residuals is determined based on prior knowledge;
  • the second quantization result e l-1 according to the second source data x l-1 , the x l- 1 , the second quantization result u l-1 of the e l-1 , and the target quantized signal to noise ratio SNR t determining the threshold, including:
  • the prediction residuals e l-1 and the second quantization result u l-1 calculates the transmission end signal to noise ratio Determining whether the SNR q is greater than the SNR t ; if SNR q > SNR t , increasing each threshold in the threshold group by ⁇ ; if SNR q ⁇ SNR t , decreasing each threshold in the threshold group ⁇ ;
  • N ⁇ obtain results set threshold values, averaging the n ⁇ threshold values to obtain the first set of thresholds.
  • the method before the comparing the prediction residual e l of the first source data x 1 with the threshold to determine the segment to which the e l belongs, the method further includes:
  • the method further includes:
  • the method further includes:
  • the prediction coefficient including a first prediction coefficient, the first prediction coefficient being a prediction coefficient for predicting the first source data x l ;
  • the calculating the prediction coefficient includes:
  • Calculating the predicted value of the first source data x l according to the second source data x l-1 and the second quantization result u l-1 include:
  • the method further includes:
  • the location identifier information of the u l is a segment to which the e l belongs;
  • the u l , the location identification information of the u l , the threshold, and the first prediction coefficient are combined into a data frame for transmission.
  • the segment to which the prediction residual needs to be quantized is determined first, and then the target quantization mode for quantifying the prediction residual is determined according to the segment to which the prediction residual belongs, and finally the target residual quantization method is used to perform the prediction residual.
  • Quantization to obtain the quantized result, so that the correlation between the source data and the distribution law of the source data can be fully utilized, and when the expected quantized signal to noise ratio is obtained, the average number of bits required for encoding can be reduced, thereby reducing the data.
  • the bandwidth required for transmission is
  • FIG. 1 is a schematic structural diagram of a quantizer according to an embodiment of the present invention.
  • FIG. 2 is a schematic structural diagram of another quantizer according to an embodiment of the present invention.
  • FIG. 3 is a schematic flowchart diagram of a quantization method according to an embodiment of the present disclosure
  • FIG. 4 is a schematic flowchart diagram of another quantization method according to an embodiment of the present invention.
  • 4a is a schematic structural diagram of a linear predictor according to an embodiment of the present invention.
  • 4b is a schematic diagram of a probability density of prediction residuals according to an embodiment of the present invention.
  • FIG. 5 is a schematic flowchart of calculating an initial value of ⁇ 1 according to an embodiment of the present invention.
  • FIG. 6 is a schematic flowchart diagram of a threshold update method according to an embodiment of the present invention.
  • the embodiment of the invention provides a quantizer and a quantization method, which can fully utilize the correlation characteristics between the source data and the distribution law of the source data, and reduce the average required for encoding when obtaining the expected quantized signal to noise ratio.
  • the number of bits which reduces the bandwidth required for data transmission.
  • FIG. 1 is a schematic structural diagram of a quantizer disclosed in an embodiment of the present invention.
  • the quantizer described in this embodiment includes a quantization selection circuit 10 and a quantization circuit group 20.
  • the quantization circuit group 20 includes at least two quantization circuits, each of which corresponds to a quantization method, wherein:
  • the quantization selection circuit 10 includes an input terminal 11 and an output terminal 12; the quantization circuit group 20 includes an input terminal 21 and an output terminal 22.
  • the output 12 of the quantization selection circuit 10 is connected to the input 21 of the quantization circuit group 20.
  • the source data including the first source data x l and the second source data x l-1 is digitized data, for example, when the data is transmitted, the data is sent.
  • the quadrature modulated signal that needs to be transmitted at the end.
  • the quantization circuit 10 to select the first source data x l e l prediction residuals to a threshold to determine the segment belongs, e l, e l quantizing determined according to the segment belongs e l The target quantization method, and outputs e l to the target quantization circuit for performing the target quantization mode.
  • e l is the predicted value of x l and x l Difference
  • Is the predicted value of the linear prediction is obtained for x l
  • e l quantized selection input of the circuit 10 is 11 reaches the quantization selection circuit 10
  • e l quantized selection target quantization circuit output of the circuit 10 is 12 to the quantization circuit group 20 .
  • the quantization circuit group 20 may be configured with a switching circuit such that the input end of the target quantization circuit is connected to the input end 21 of the quantization circuit group 20, so that the target quantization circuit can receive the quantization selection.
  • the output l of circuit 10 is e l .
  • the target quantization circuit in the quantization circuit group 20 quantizes e l to obtain a first quantization result u l .
  • the threshold may be a set of thresholds pre-stored in the quantization selection circuit 10; or, the quantization selection circuit 10 pre-stores a plurality of sets of selectable thresholds, and selects a set of selectable thresholds every preset time interval.
  • the threshold is updated; alternatively, the threshold is input by the input 11 of the quantization selection circuit 10.
  • the quantization circuit group 20 includes at least two quantization circuits, each of which corresponds to a quantization mode, wherein the optional quantization modes include: uniform quantization, A-rate quantization, and ⁇ -rate quantization, etc., specifically adopting The manner of the present invention is not limited.
  • the embodiment of the invention described in FIG. 1 implements segmentation quantization on the prediction residual, that is, the matching quantization mode is selected for each segment according to the distribution rule of the prediction residual, and the quantization signal to noise ratio requirement can be met.
  • the data compression rate is increased, that is, the bandwidth required for transmission is reduced.
  • FIG. 2 is a schematic structural diagram of another quantizer disclosed in the embodiment of the present invention.
  • the quantizer described in this embodiment can be built on the basis of the quantizer described in FIG. 1, in addition to the quantization selection circuit 10 and the quantization circuit group 20 described in FIG.
  • the threshold calculation circuit 30 is included; wherein the embodiments of the quantization selection circuit 10 and the quantization circuit group 20 can be referred to the embodiment described in FIG.
  • the threshold calculation circuit 30 includes a first input terminal 31, a second input terminal 32, a third input terminal 33, a fourth input terminal 34, and an output terminal 35.
  • the output 22 of the quantization circuit group 20 is connected to the third input 33 of the threshold calculation circuit 30; the output 35 of the threshold calculation circuit 30 is connected to the input 11 of the quantization selection circuit.
  • first input end, the second input end, the third input end, the first output end, and the second output end of each hardware module in the embodiment of the present invention may be multiple in physical implementation.
  • the different split ports may also be a merged port.
  • the specific manner is selected according to the data transmission speed or the hardware implementation difficulty, and is not limited in the embodiment of the present invention.
  • Threshold value calculating circuit 30 in accordance with a second source data x l-1, x l- 1 prediction residuals e l-1, e l- 1 of the second quantization result u l-1 and the target is determined quantization noise ratio SNR t threshold, x l-1 of the source data of a previous x l; e l-1 x l-1 is the predicted value of x l-1 Difference, a predicted value obtained by linearly predicting x l-1 ; wherein x l-1 , e l-1 , u l-1 , and SNR t pass through the first input terminal 31 and the second input terminal of the threshold value calculation circuit 30, respectively. 32.
  • the third input terminal 33 and the fourth input terminal 34 reach the threshold value calculation circuit 30; the threshold value reaches the input terminal 11 of the quantization selection circuit 10 via the output terminal 35 of the threshold value calculation circuit 30.
  • the threshold value of n-1 are values of ⁇ 1 ⁇ ⁇ n-1 consisting of a first set threshold, ⁇ 1 ⁇ ⁇ n-1 prediction residual range division n segments; the first threshold group is an average of n ⁇ threshold groups; wherein the interval between every two thresholds of the threshold group is determined according to the distribution of prediction residuals; and the distribution of prediction residuals is based on prior knowledge determine.
  • the threshold calculation circuit 30 obtains the threshold by:
  • the prediction x l-1, x l- 1 residuals e l-1 and the second quantized result u l-1 calculates the transmission end signal to noise ratio Determining whether SNR q is greater than SNR t ; if SNR q > SNR t , increasing each threshold in the threshold group by ⁇ ; if SNR q ⁇ SNR t , decreasing each threshold in the threshold group by ⁇ ;
  • n ⁇ subthreshold group acquires the calculation result threshold values n ⁇ group, n ⁇ averaged to obtain a set of threshold values a first threshold value group.
  • the quantizer described in FIG. 2 may further include a subtractor 40; the subtractor 40 includes a first input 41, a second input 42 and an output 43; the output 43 of the subtractor 40
  • the input terminal 11 of the quantization selection circuit 10 is connected.
  • the input end 11 of the quantization selection circuit 10 receives both the threshold and the prediction residual e l , which may be the same input or two separate inputs in physical implementation.
  • the specific implementation of the present invention is not limited in the embodiment of the present invention.
  • the subtractor 40 compares the predicted values of the first source data x l and x l Subtracted to obtain the prediction residual x l e l; X l is predicted to be a value obtained by linear prediction; wherein, x l and Respectively via a first input 40 of the subtractor 41 and a second input terminal 42 to the subtracter 40; e l subtracter 40 through the output terminal 43 reaches the quantization circuit 10 to select the input terminal 11.
  • the quantizer described in FIG. 2 may further include a linear predictor 50; the linear predictor 50 includes a first input 51, a second input 52, and an output 53; the output of the linear predictor The terminal 53 is connected to the input end 42 of the subtractor 40;
  • the linear predictor 50 calculates the predicted value of the first source data x l based on the second source data x l-1 and the second quantized result u l-1 Wherein, x l-1 and u l-1 reach the linear predictor 50 via the first input 51 and the second input 52 of the linear predictor 50; The output 53 of the linear predictor 50 reaches the second input 42 of the subtractor 40.
  • the linear predictor 50 may include a prediction coefficient calculation circuit 501 and a linear prediction circuit 502.
  • the prediction coefficient calculation circuit 501 includes an input terminal 5011 and an output terminal 5012.
  • the linear prediction circuit 502 includes a first input terminal 5021, a second input terminal 5022, a third input terminal 5023, and an output terminal 5024.
  • Prediction coefficient calculation circuit 501 calculates the prediction coefficients, the prediction coefficients comprises coefficients a first prediction, the first prediction coefficients for the first source data x l linear prediction coefficients for the prediction.
  • the above calculated prediction coefficients include:
  • the f correlation functions are averaged to obtain an average correlation function, and the relationship between the average correlation function and the prediction coefficients is established by the minimum mean square error criterion to obtain the prediction coefficients.
  • the first prediction coefficient is passed to the first input 5021 of the linear prediction circuit 502 via the output 5012 of the prediction coefficient calculation circuit 501.
  • the linear prediction circuit 502 is based on the predicted values of the first prediction coefficient and the second source data.
  • u l-1 calculation include:
  • the quantizer described in FIG. 2 may further include a framing circuit 60; the framing circuit 60 includes a first input terminal 61, a second input terminal 62, a third input terminal 63, and a fourth input.
  • the end 64 and the output 65; the output of the quantization selection circuit 10 further includes a first output 13 and a second output 14.
  • the first output terminal 13 and the second output terminal 14 of the quantization selection circuit 10 are connected to the first input terminal 61 and the second input terminal 62 of the framing circuit 60; the output terminal 22 of the quantization circuit group 20 is connected to the third of the framing circuit 60 The input terminal 63; the output terminal 5012 of the prediction coefficient calculation circuit 501 is connected to the fourth input terminal 64 of the framing circuit 60.
  • Quantization selection circuit 10 is also configured to obtain the location identification information u l; location identification information is u l e l segment belongs; a first output terminal location identification information and the threshold value u l respectively quantized selection circuit 10 and 13 a second output terminal 14 to the framing circuit 60; u l quantized output circuit group 20 reaches the end 22 of the framing circuit 60; a first prediction coefficient prediction coefficient calculation circuit 501 via the output terminal 5012 reaches the framing circuit 60.
  • the framing circuit combines the position identification information of u l , u l , the threshold value and the first prediction coefficient into a data frame for transmission.
  • the embodiment of the invention described in FIG. 2 implements segmentation quantization on the prediction residual, that is, according to the distribution rule of the prediction residual, the matching quantization mode is selected for each segment, and the quantized signal to noise ratio requirement can be met.
  • the data compression rate is increased, that is, the bandwidth required for transmission is reduced; in addition, the threshold of the prediction residual difference segment is determined according to the requirements of the target quantization signal to noise ratio to ensure that the quantized data can achieve ideal signal noise. ratio.
  • FIG. 3 is a schematic flowchart diagram of a quantization method according to an embodiment of the present invention.
  • the quantization method described in FIG. 3 may include the following steps:
  • the quantizer obtains the prediction residual e l of the first source data x l and a threshold
  • the quantizer may be a circuit module in the terminal device as the data transmitting end.
  • the above e l is a predicted value of x l and x l Difference, The predicted value obtained by linearly predicting x l .
  • the threshold value is n-1 first threshold groups each composed of values of ⁇ 1 to ⁇ n-1 , and the range of the prediction residual is divided into n segments.
  • the first threshold group may be pre-stored in the quantizer, or may be calculated according to data input in real time.
  • the first threshold group is an average of n ⁇ threshold groups; wherein an interval between every two thresholds of the threshold group is determined according to a distribution of prediction residuals; and a prediction residual is distributed according to Prior knowledge is determined.
  • the first threshold group can be calculated in the following manner:
  • the threshold group is calculated: the previous source data according to the first source data x l : the second source data x l-1 , the prediction residual e l-1 of the second source data x l-1 , and the first
  • the second quantization result of the second source data u l-1 calculates the signal to noise ratio of the transmitting end Determining whether SNR q is greater than SNR t ; if SNR q > SNR t , increasing each threshold in the threshold group by ⁇ ; if SNR q ⁇ SNR t , decreasing each threshold in the threshold group by ⁇ ;
  • Loop is executed n ⁇ times and the threshold value set calculation process, the calculation result to obtain threshold values n ⁇ group, n ⁇ averaged to obtain a set of threshold values a first threshold value group.
  • different segments correspond to different quantization modes, wherein the quantized modes that can be used include: uniform quantization, A-rate quantization, and ⁇ -rate quantization; and the same quantization method is used for all quantized residuals.
  • the embodiment of the present invention can better match the distribution rule of the quantized residuals, and further reduce the bit width required for the quantization result.
  • the prediction residual is segment-quantized, that is, the matching quantization method is selected for each segment according to the distribution rule of the prediction residual, which can meet the requirements of the quantized signal to noise ratio.
  • increase the data compression rate that is, reduce the bandwidth required for transmission.
  • FIG. 4 is a schematic flowchart diagram of another quantization method according to an embodiment of the present invention.
  • the quantization method described in FIG. 4 may include the following steps:
  • the quantizer calculates the prediction coefficient by using the source data
  • the quantizer may be a circuit module in the terminal device as the data transmitting end.
  • the source data including the first source data x l and the second source data x l-1 is digitized data, for example, when the data transmission is performed, the data transmitting end needs to transmit.
  • the first prediction coefficient is a prediction coefficient that predicts the first source data x l .
  • the predicted value of the first source data x l can be obtained by linear prediction.
  • the data transmitting end updates the prediction coefficient every time the f*N block data is entered.
  • the N init data is obtained as the initial training sample.
  • k 1 is the order of the correlation function
  • N block is the sample length of the source signal required to calculate the correlation function
  • Table 1 Storage structure diagram of the correlation function of the initial training sample
  • the prediction coefficients need to be transmitted from the transmitting end to the receiving end, the transmission of the prediction coefficients will occupy the transmission bandwidth, so the prediction coefficient update step size cannot be too short.
  • the update of the prediction coefficients needs to be able to keep up with the time-varying characteristics of the correlation characteristics of the source data, so the update step of the prediction coefficients cannot be too long and needs to be compromised.
  • the update step size N update should be less than N init .
  • the process of subsequently calculating the prediction coefficients including:
  • the array is shifted to the left overall, and the result of the new calculation is placed at the far right of the array.
  • the previous R f+1 ⁇ R F is the correlation function calculated from the f+1 to F data of the training sample
  • the following R 1 ⁇ R f are the f group correlations calculated by the subsequent set of data.
  • the function after entering the data, updates the array of related functions in this way.
  • the data transmitting end stores only f correlation functions that are currently updated with prediction coefficients, the hardware overhead is moderate, and the hardware structure is easy to implement.
  • the first source data x l can be predicted by linear prediction.
  • FIG. 4a is a schematic structural diagram of a linear predictor according to an embodiment of the present invention.
  • the product of the previous source data x l (ie, the second source data) x l-1 and the prediction coefficient ⁇ 1 is calculated as an intermediate amount; then the above calculated intermediate amount is repeatedly executed.
  • step 403 is performed to predict the first source data x l and x l Subtracting to obtain the prediction residual e l of x l , and then outputting the prediction residual e l to the encoder for quantization coding.
  • the calculation formula of the prediction residual e l can be:
  • K is the prediction order of the linear predictor, and its selection depends on the correlation characteristics of the source. It is necessary to analyze the relevant characteristics of the source data, that is, calculate the correlation function for the source data shift to select the appropriate interval. When the delay difference between the two sampled data is greater than the interval, the correlation between the two sampled data is lower than the expected threshold, and the interval value is the K value of the demand.
  • the transmitting end needs to quantize the prediction residual.
  • the selection of the quantization scheme depends on the statistical characteristics of the prediction residual. Therefore, the probability distribution of the prediction residual is first counted.
  • the probability density distribution of the prediction residual is shown in FIG. 4b, it can be seen from the figure that the distribution of the prediction residual substantially conforms to the characteristics of the Gaussian distribution, and the probability of occurrence of the small residual is large, and the large residual is large.
  • the probability of occurrence is small.
  • the quantization can be performed with fewer bits, that is, the higher precision can be obtained, and for the residual with larger amplitude, more bits need to be used for quantization to ensure the quantization precision. Therefore, the prediction residual is segmented by the threshold, and each segment selects a quantization method suitable for the segment, which can further reduce the average bit width required for the quantization result.
  • the threshold value is n-1 first threshold groups respectively composed of values of ⁇ 1 ⁇ n-1 , and the value range of the prediction residual is divided into n segments, and each segment is predicted according to the prediction.
  • the distribution of residuals is quantified in different ways.
  • ⁇ 1 is the threshold with the smallest absolute value
  • the interval between the respective thresholds is d i,j
  • FIG. 5 is a schematic flowchart of calculating an initial value of ⁇ 1 according to an embodiment of the present invention.
  • the initial ⁇ 1 can be calculated by the following steps:
  • FIG. 6 is a schematic flowchart diagram of a threshold update method according to an embodiment of the present invention.
  • the threshold update can be performed as follows:
  • step 603. Determine whether the counter count is equal to f. If it is equal to f, execute step 604, otherwise return to step 602.
  • comparison SNR q is greater than SNR t, if the SNR q> SNR t, step 606 is performed; if SNR q ⁇ SNR t, step 607 is executed.
  • step 608. Determine whether k is equal to n ⁇ . If yes, perform step 609. If no, return to step 602.
  • the threshold is periodically adjusted according to the signal-to-noise ratio of the data transmitting end to ensure that the quantization result can meet the preset signal-to-noise ratio requirement.
  • the prediction residual is segmented by the threshold, and each segment selects a quantization method suitable for the segment, which can further reduce the average bit width required for the quantization result.
  • the range of the prediction residual is divided into three segments, wherein the absolute value is smaller than ⁇ 1 and the absolute value is between ⁇ 1 and ⁇ 2 .
  • location identification information u l comprises a segment u l corresponding e l belongs, so as to be contained when u l a data frame to the data receiving end, the data reception terminal belongs based on the threshold value and e l of segment, for determining the data transmission terminal e l quantizing embodiment, recovered by u l e l.
  • the prediction residual is segment-quantized, that is, the matching quantization mode is selected for each segment according to the distribution rule of the prediction residual, which can meet the requirements of the quantized signal to noise ratio.
  • the data compression rate is increased, that is, the bandwidth required for transmission is reduced; in addition, the threshold of the prediction residual difference segment is determined according to the requirement of the target quantization signal to noise ratio to ensure that the quantized data can reach an ideal signal to noise ratio.
  • the signal which has an absolute value greater than ⁇ 2 , is a large signal.
  • the small and medium signals are logarithmically scaled similar to the ⁇ law, and the large signals are uniformly quantized. Refer to Table 4, which is the parameter configuration information for segmentation quantization.
  • the slope of the polyline is determined according to the ⁇ law formula, where the ⁇ law formula is:
  • the 32-line line is used to approximate the ⁇ law, and the slope of the 32-segment polyline of the small signal and the middle signal are respectively obtained.
  • i 1,...,32;
  • the codebook1 and codebook2 are obtained by using the obtained boundary value of the x-axis, namely:
  • Codebook m codebook2*( ⁇ 2 - ⁇ 1 )+ ⁇ 1
  • the small signal and the middle signal are quantized by the above-described quantized codebook, and the large signal is quantized by means of uniform quantization.
  • the prediction residual is quantized, the coded transmission is performed. Because the segmentation quantization scheme is adopted, in order to ensure the correct decoding of the receiving end, the data frame after the framing needs to include the prediction coefficient, the position information of the segment where the prediction residual is located, Threshold, prediction residual.
  • the data transmission process is simulated, and the obtained signal-to-noise ratio is 55.0435dB, and the average bit of the transmitted data is 6.1780.
  • Table 5 for the simulation results obtained under the same conditions as CPRI compression and uniform quantization. Comparison of results:
  • the data receiving end After receiving the data frame, the data receiving end first de-frames to obtain a threshold, calculates a codebook according to the threshold, de-frames the prediction coefficient, and transmits the prediction coefficient to the linear predictor to obtain the predicted value of the source data; Obtaining the position information of the segment where the prediction residual is located, the decoder can be used to decode the quantization result; finally, the decoding result is added to the predicted value obtained by the linear predictor, and the signal sent by the data transmitting end can be recovered. source data.
  • the embodiment of the present invention implements segmentation quantization on the prediction residual, that is, selects a matching quantization mode for each segment according to the distribution rule of the prediction residual, which can satisfy the requirement of the quantized signal to noise ratio.
  • segmentation quantization on the prediction residual, that is, selects a matching quantization mode for each segment according to the distribution rule of the prediction residual, which can satisfy the requirement of the quantized signal to noise ratio.
  • the program can be stored in a computer readable storage medium, when the program is executed
  • the flow of the method embodiments as described above may be included.
  • the foregoing storage medium includes various media that can store program codes, such as a ROM or a random access memory RAM, a magnetic disk, or an optical disk.

Landscapes

  • Physics & Mathematics (AREA)
  • Engineering & Computer Science (AREA)
  • Spectroscopy & Molecular Physics (AREA)
  • Computational Linguistics (AREA)
  • Signal Processing (AREA)
  • Health & Medical Sciences (AREA)
  • Audiology, Speech & Language Pathology (AREA)
  • Human Computer Interaction (AREA)
  • Acoustics & Sound (AREA)
  • Multimedia (AREA)
  • Compression, Expansion, Code Conversion, And Decoders (AREA)

Abstract

一种量化器与量化方法，其中，量化器包括：量化选择电路（10）和至少两个量化电路（1）；每个量化电路（1）对应一种量化方式；量化选择电路（10）的输出端（12）连接量化电路（1）的输入端（21）；量化选择电路（10）用于将第一信源数据x l的预测残差e l与阈值比较，以确定e l所属的分段，根据e l所属的分段确定对e l进行量化的目标量化方式，并将e l输出至用于执行目标量化方式的目标量化电路；e l为x l与x l的预测值（I）的差值，（I）为对x l进行线性预测获得的预测值；目标量化电路用于对e l进行量化以获得第一量化结果u l。本方法可以充分利用信源数据之间的相关特性以及信源数据的分布规律，在获得预期的量化信噪比时，能够降低编码时所需的平均比特数，降低数据传输时所需的带宽。

Description

量化器与量化方法 技术领域
本发明涉及通信技术领域，特别涉及一种量化器与量化方法。
背景技术
在实际的通信场景中，大部分信源属于有记忆的信源，对于有记忆信源，信源输出的信源数据之间存在统计关联，如果对信源数据间的统计关联加以利用，可以降低数据传输中的冗余，起到数据压缩的作用，进而提高带宽利用率。
预测编码就是基于这一思想。预测编码不直接对信源数据进行编码，而是对信源数据进行预测，对预测值与信源数据的差值进行编码。在实际应用中，DPCM(Differential Pulse Code Modulation，差分脉冲编码调制)便是一种常用的基于线性预测的编码架构，其工作原理如下：根据历史信源数据来预测当前输入的信源数据的值，并计算预测值与实际的信源数据的差值，对差值进行量化编码后进行数字信号的传输。
对上述差值进行量化时，可以选择均匀量化的方式，假设量化器的工作范围为(-V，+V)，均匀量化器将(-V，+V)的取值范围均匀分割成M个长度为Δ＝2V/M的区间。获取区间的中点作为量化电平，对上述差值进行量化。然而，均匀量化若获得预期的信噪比，需要占用较多的带宽。
CPRI(Common Public Radio Interface，通用公共无线电接口)是规范基站内部射频控制设备和射频设备之间接口标准的协议。降低CPRI接口传输负载的数据压缩技术，包括时域压缩方案(如降采样率、非线性量化、正交调制信号数据压缩等)和频域压缩方案(如子载波压缩等)。
频域压缩方案的出发点在于，信源信号在频域上存在大量的冗余；但是频域压缩会导致星座图映射复杂，增加接口逻辑设计和处理复杂度，不具备可实现性。
另一种方案是时域压缩方案。一种代表性的时域压缩方案是：基于信源信号特征，通过插值、低通滤波以及降采样的方式消除冗余，同时也大幅减少了符号数量。另外，又针对信号动态范围较大的特性，对数据进行分段缩放，以保证相同的数据位宽能够表征更多幅值较大的信号。然而，时域压缩方案并没有充分利用信源之间的相关特性及信源的分布规律，压缩率还有提升的空间。
发明内容
本发明实施例提供了一种量化器与量化方法，用于将数字信号进行压缩以降低传输时所需的带宽。
本发明实施例第一方面公开了一种量化器，包括：
量化选择电路和至少两个量化电路；每个量化电路对应一种量化方式；
所述量化选择电路的输出端连接所述量化电路的输入端；
所述量化选择电路用于将第一信源数据xl的预测残差el与阈值比较，以确定所述el所属的分段，根据所述el所属的分段确定对所述el进行量化的目标量化方式，并将所述el输出至 用于执行所述目标量化方式的目标量化电路；所述el为所述xl与所述xl的预测值
Figure PCTCN2017071981-appb-000001
的差值，所述
Figure PCTCN2017071981-appb-000002
为对所述xl进行线性预测获得的预测值；所述el经所述量化选择电路的输入端到达所述量化选择电路；所述el经所述量化选择电路的输出端到达所述目标量化电路；
所述目标量化电路用于对所述el进行量化以获得第一量化结果ul。
在该实施方式中，对预测残差进行分段量化，即根据预测残差的分布规律为每个分段选择匹配的量化方式，可以在满足量化信噪比要求的前提下，提高数据压缩率，即降低传输时需要的带宽。
作为一种可选的实施方式，所述量化器还包括阈值计算电路；
所述至少两个量化电路的输出端连接所述阈值计算电路的输入端；所述阈值计算电路的输出端连接所述量化选择电路的输入端；
所述阈值计算电路用于根据第二信源数据xl-1、所述xl-1的预测残差el-1、所述el-1的第二量化结果ul-1和目标量化信噪比SNRt确定所述阈值，所述xl-1为所述xl前一次的信源数据；所述el-1为所述xl-1与所述xl-1的预测值
Figure PCTCN2017071981-appb-000003
的差值，所述
Figure PCTCN2017071981-appb-000004
为对所述xl-1进行线性预测获得的预测值；其中，所述xl-1、所述el-1、所述ul-1和所述SNRt经所述阈值计算电路的输入端到达所述阈值计算电路；所述阈值经所述阈值计算电路的输出端到达所述量化选择电路。
在该实施方式中，用于将预测残差分段的阈值根据目标量化信噪比的要求确定，以保证量化后的数据可以达到理想的信噪比。
作为一种可选的实施方式，所述阈值为n-1个分别为τ1～τn-1的值组成的第一阈值组，所述τ1～τn-1将预测残差的取值范围划分为n个分段；所述第一阈值组为nτ个阈值组的平均值；其中，阈值组的每两个阈值间的间隔根据所述预测残差的分布确定；所述预测残差的分布根据先验知识确定；
所述阈值计算电路用于根据第二信源数据xl-1、所述xl-1的预测残差el-1、所述el-1的第二量化结果ul-1和目标量化信噪比SNRt确定所述阈值，包括：
计算阈值组：
根据所述xl-1、所述xl-1的预测残差el-1和所述第二量化结果ul-1计算发送端信噪比
Figure PCTCN2017071981-appb-000005
判断所述SNRq是否大于所述SNRt；若SNRq＞SNRt，则将所述阈值组内各阈值分别增加δ；若SNRq＜SNRt，则将所述阈值组内各阈值分别减少δ；
获取nτ个阈值组的计算结果，对所述nτ个阈值组求平均以获得所述第一阈值组。
在该实施方式中，阈值可以根据发送端信噪比进行动态修正，以保证量化后的数据可以达到理想的信噪比。
作为一种可选的实施方式，所述量化器还包括减法器；
所述减法器的输出端连接所述量化选择电路的输入端；
所述减法器用于将第一信源数据xl与所述xl的预测值
Figure PCTCN2017071981-appb-000006
相减以获得所述xl的预测残差el；所述
Figure PCTCN2017071981-appb-000007
为对所述xl进行线性预测获得的预测值；其中，所述xl和所述
Figure PCTCN2017071981-appb-000008
经所述减法器的输入端到达所述减法器；所述el经所述减法器的输出端到达所述量化选择电路。
在该实施方式中，利用信源数据之间的相关特性，将传输信源数据转变为传输对信源 数据进行预测的预测残差，从而降低传输的数据位宽，提高数据压缩率。
作为一种可选的实施方式，所述量化器还包括线性预测器；
所述线性预测器的输出端连接所述减法器的输入端；
所述线性预测器用于根据所述第二信源数据xl-1和所述第二量化结果ul-1计算所述第一信源数据xl的预测值
Figure PCTCN2017071981-appb-000009
其中，所述xl-1和所述ul-1经所述线性预测器的输入端到达所述线性预测器；所述
Figure PCTCN2017071981-appb-000010
经所述线性预测器的输出端到达所述减法器。
在该实施方式中，利用信源数据之间的相关特性，对信源数据进行线性预测，在进行数据传输时，将传输信源数据转变为传输对信源数据进行预测的预测残差，从而降低传输的数据位宽，提高数据压缩率。
作为一种可选的实施方式，所述线性预测器包括：
预测系数计算电路和线性预测电路；
所述预测系数计算电路用于计算预测系数，所述预测系数包含第一预测系数，所述第一预测系数为对所述第一信源数据xl进行所述线性预测的预测系数；
所述计算预测系数，包括：
利用所述xl之前的Nblock个信源数据计算相关函数；
获取f个所述相关函数之后对所述f个所述相关函数求平均以获得平均相关函数，通过最小均方误差准则建立所述平均相关函数与所述预测系数的关系以求得所述预测系数；
所述第一预测系数经所述预测系数计算电路的输出端到达所述线性预测电路；
所述线性预测电路用于根据所述第一预测系数、所述第二信源数据的预测值
Figure PCTCN2017071981-appb-000011
和所述ul-1计算所述
Figure PCTCN2017071981-appb-000012
包括：
计算所述
Figure PCTCN2017071981-appb-000013
与所述ul-1的和与所述第一预测系数的积作为中间量；
获取K个所述中间量之后对所述K个中间量求和以获得所述
Figure PCTCN2017071981-appb-000014
其中，所述K为线性预测的阶数；
所述
Figure PCTCN2017071981-appb-000015
和所述ul-1经所述线性预测电路的输入端到达所述线性预测电路；所述
Figure PCTCN2017071981-appb-000016
经所述线性预测电路的输出端到达所述减法器。
在该实施方式中，在计算相关函数时，采用将输入的信源数据分段计算相关函数的方式，并分段进行相关函数的更新，降低在数据发送端进行相关函数计算的数据复杂度。
作为一种可选的实施方式，所述量化器还包括组帧电路；
所述量化选择电路的输出端连接所述组帧电路的输入端；所述目标量化电路的输出端连接所述组帧电路的输入端；所述预测系数计算电路的输出端连接所述组帧电路的输入端；
所述量化选择电路还用于获取所述ul的位置标识信息；所述ul的位置标识信息为所述el所属的分段；所述ul的位置标识信息和所述阈值分别经所述量化选择电路的输出端到达所述组帧电路；
所述ul经所述目标量化电路的输出端到达所述组帧电路；
所述第一预测系数经所述预测系数计算电路的输出端到达所述组帧电路；
所述组帧电路用于将所述ul、所述ul的位置标识信息、所述阈值和所述第一预测系数组成数据帧以进行传输。
在该实施方式中，将量化结果、位置标识信息、阈值和预测系数组成数据帧以进行传 输；数据接收端接收到上述数据帧后，可根据量化结果、位置标识信息和阈值还原出信源数据的预测残差，利用预测系数求得对信源数据的预测值，将预测残差和预测值相加即可还原出信源数据。
本发明实施例第二方面公开了一种量化方法，包括：
获取第一信源数据xl的预测残差el与阈值；所述el为所述xl与所述xl的预测值
Figure PCTCN2017071981-appb-000017
的差值，所述
Figure PCTCN2017071981-appb-000018
为对所述xl进行线性预测获得的预测值；
将所述el与所述阈值比较以确定所述el所属的分段，根据所述el所属的分段确定对所述el进行量化的目标量化方式；
利用所述目标量化方式对所述el进行量化以获得第一量化结果ul。
作为一种可选的实施方式，所述将第一信源数据xl的预测残差el与阈值比较以确定所述el所属的分段之前，所述方法还包括：
根据第二信源数据xl-1、所述xl-1的预测残差el-1、所述el-1的第二量化结果ul-1和目标量化信噪比SNRt确定所述阈值；所述xl-1为所述xl前一次的信源数据，所述el-1为所述xl-1与所述xl-1的预测值
Figure PCTCN2017071981-appb-000019
的差值，所述
Figure PCTCN2017071981-appb-000020
为对所述xl-1进行线性预测获得的预测值。
作为一种可选的实施方式，所述阈值为n-1个分别为τ1～τn-1的值组成的第一阈值组，所述τ1～τn-1将预测残差的取值范围划分为n个分段；所述第一阈值组为nτ个阈值组的平均值；其中，阈值组的每两个阈值间的间隔根据所述预测残差的分布确定；所述预测残差的分布根据先验知识确定；
所述根据第二信源数据xl-1、所述xl-1的预测残差el-1、所述el-1的第二量化结果ul-1和目标量化信噪比SNRt确定所述阈值，包括：
计算阈值组：
根据所述xl-1、所述xl-1的预测残差el-1和所述第二量化结果ul-1计算发送端信噪比
Figure PCTCN2017071981-appb-000021
判断所述SNRq是否大于所述SNRt；若SNRq＞SNRt，则将所述阈值组内各阈值分别增加δ；若SNRq＜SNRt，则将所述阈值组内各阈值分别减少δ；
获取nτ个阈值组的计算结果，对所述nτ个阈值组求平均以获得所述第一阈值组。
作为一种可选的实施方式，所述将第一信源数据xl的预测残差el与阈值比较以确定所述el所属的分段之前，所述方法还包括：
将第一信源数据xl与所述xl的预测值
Figure PCTCN2017071981-appb-000022
相减以获得所述xl的预测残差el；所述
Figure PCTCN2017071981-appb-000023
为对所述xl进行线性预测获得的预测值。
作为一种可选的实施方式，所述将第一信源数据xl与所述xl的预测值
Figure PCTCN2017071981-appb-000024
相减以获得所述xl的预测残差el之前，所述方法还包括：
根据所述第二信源数据xl-1和所述第二量化结果ul-1计算所述第一信源数据xl的预测值
Figure PCTCN2017071981-appb-000025
作为一种可选的实施方式，所述根据所述第二信源数据xl-1和所述第二量化结果ul-1计算所述第一信源数据xl的预测值
Figure PCTCN2017071981-appb-000026
之前，所述方法还包括：
计算预测系数，所述预测系数包含第一预测系数，所述第一预测系数为对所述第一信 源数据xl进行预测的预测系数；
所述计算预测系数，包括：
利用所述xl之前的Nblock个信源数据计算相关函数；
获取f个所述相关函数之后对所述f个所述相关函数求平均以获得平均相关函数，通过最小均方误差准则建立所述平均相关函数与所述预测系数的关系以求得所述预测系数；
所述根据所述第二信源数据xl-1和所述第二量化结果ul-1计算所述第一信源数据xl的预测值
Figure PCTCN2017071981-appb-000027
包括：
计算所述
Figure PCTCN2017071981-appb-000028
与所述ul-1的和与第一预测系数的积作为中间量；
获取K个所述中间量之后对所述K个中间量求和以获得所述
Figure PCTCN2017071981-appb-000029
其中，所述K为线性预测的阶数。
作为一种可选的实施方式，所述方法还包括：
获取所述ul的位置标识信息和所述阈值；所述ul的位置标识信息为所述el所属的分段；
将所述ul、所述ul的位置标识信息、所述阈值和所述第一预测系数组成数据帧以进行传输。
从以上技术方案可以看出，本发明实施例具有以下优点：
本发明实施例中，先确定需要量化的预测残差所属的分段，再根据预测残差所属的分段确定对预测残差进行量化的目标量化方式，最后利用目标量化方式对预测残差进行量化以获得量化结果，从而可以充分利用信源数据之间的相关特性以及信源数据的分布规律，在获得预期的量化信噪比时，能够降低编码时所需的平均比特数，从而降低数据传输时所需的带宽。
附图说明
为了更清楚地说明本发明实施例或背景技术中的技术方案，下面将对本发明实施例或背景技术中所需要使用的附图进行说明。
图1为本发明实施例公开的一种量化器的结构示意图；
图2为本发明实施例公开的另一种量化器的结构示意图；
图3为本发明实施例公开的一种量化方法的流程示意图；
图4为本发明实施例公开的另一种量化方法的流程示意图；
图4a为本发明实施例公开的一种线性预测器的结构示意图；
图4b为本发明实施例公开的一种预测残差的概率密度的示意图；
图5为本发明实施例公开的一种计算初始的τ1的值的流程示意图；
图6为本发明实施例公开的一种阈值更新方法的流程示意图。
具体实施方式
下面结合本发明实施例中的附图对本发明实施例进行描述。
本发明实施例提供了一种量化器和量化方法，可以充分利用信源数据之间的相关特性以及信源数据的分布规律，在获得预期的量化信噪比时，降低编码时所需的平均比特数，从而降低数据传输时所需的带宽。
请参阅图1，图1是本发明实施例公开的一种量化器的结构示意图。如图1所示，本实施例中所描述的量化器，包括量化选择电路10和量化电路组20，量化电路组20包括至少两个量化电路，每个量化电路对应一种量化方式，其中：
量化选择电路10包括输入端11和输出端12；量化电路组20包括输入端21和输出端22。
量化选择电路10的输出端12连接量化电路组20的输入端21。
本发明实施例中，包括第一信源数据xl、第二信源数据xl-1等在内的信源数据，为数字化的数据，举例来说，可为进行数据传输时，数据发送端需要传输的正交调制信号。
本发明实施例中，量化选择电路10将第一信源数据xl的预测残差el与阈值比较，以确定el所属的分段，根据el所属的分段确定对el进行量化的目标量化方式，并将el输出至用于执行目标量化方式的目标量化电路。其中，el为xl与xl的预测值
Figure PCTCN2017071981-appb-000030
的差值，
Figure PCTCN2017071981-appb-000031
为对xl进行线性预测获得的预测值；el经量化选择电路10的输入端11到达量化选择电路10；el经量化选择电路10的输出端12到达量化电路组20中的目标量化电路。
作为一种可选的实施方式，量化电路组20中可配置有开关电路，使得目标量化电路的输入端与量化电路组20的输入端21的连接导通，从而目标量化电路可以接收到量化选择电路10输出的el。
在接收到el之后，量化电路组20中的目标量化电路对el进行量化以获得第一量化结果ul。
本发明实施例中，上述阈值可为量化选择电路10中预先存储的一组阈值；或者，量化选择电路10中预先存储有多组可选阈值，每隔预设时间间隔选择一组可选阈值对阈值来进行更新；或者，阈值由量化选择电路10的输入端11进行输入。具体采用何种方式，本发明实施例不做限定。
本发明实施例中，量化电路组20包括至少两个量化电路，每个量化电路对应一种量化方式，其中，可选的量化方式包括：均匀量化、A率量化和μ率量化等，具体采用何种方式，本发明实施例不做限定。
由此可见，实施图1所描述的发明实施例，对预测残差进行分段量化，即根据预测残差的分布规律为每个分段选择匹配的量化方式，可以在满足量化信噪比要求的前提下，提高数据压缩率，即降低传输时需要的带宽。
请参阅图2，图2是本发明实施例公开的另一种量化器的结构示意图。如图2所示，本实施例中所描述的量化器，可建立在图1所描述的量化器的基础之上，除了图1所描述的量化选择电路10和量化电路组20之外，还包括阈值计算电路30；其中，量化选择电路10和量化电路组20的实施方式可参考图1所描述的实施方式。
阈值计算电路30包括第一输入端31、第二输入端32、第三输入端33、第四输入端34和输出端35。量化电路组20的输出端22连接阈值计算电路30的第三输入端33；阈值计算电路30的输出端35连接量化选择电路的输入端11。
需要说明的是，本发明实施例中的各个硬件模块的第一输入端、第二输入端、第三输入端、第一输出端、第二输出端等端口，在物理实现上可为多个不同的分离端口，也可为合并的一个端口，具体采用的方式根据数据传输速度或硬件实现难度等情况进行选择，本发明实施例不做限定。
阈值计算电路30根据第二信源数据xl-1、xl-1的预测残差el-1、el-1的第二量化结果ul-1和目标量化信噪比SNRt确定阈值，xl-1为xl前一次的信源数据；el-1为xl-1与xl-1的预测值
Figure PCTCN2017071981-appb-000032
的差值，
Figure PCTCN2017071981-appb-000033
为对xl-1进行线性预测获得的预测值；其中，xl-1、el-1、ul-1和SNRt分别经阈值计算电路30的第一输入端31、第二输入端32、第三输入端33和第四输入端34到达阈值计算电路30；阈值经阈值计算电路30的输出端35到达量化选择电路10的输入端11。
作为一种可选的实施方式，上述阈值为n-1个分别为τ1～τn-1的值组成的第一阈值组，τ1～τn-1将预测残差的取值范围划分为n个分段；第一阈值组为nτ个阈值组的平均值；其中，阈值组的每两个阈值间的间隔根据预测残差的分布确定；而预测残差的分布根据先验知识确定。
在上述实施方式中，阈值计算电路30通过如下的方式求得阈值：
首先，通过如下方式计算阈值组：根据xl-1、xl-1的预测残差el-1和第二量化结果ul-1计算发送端信噪比
Figure PCTCN2017071981-appb-000034
判断SNRq是否大于SNRt；若SNRq＞SNRt，则将阈值组内各阈值分别增加δ；若SNRq＜SNRt，则将阈值组内各阈值分别减少δ；
循环执行nτ次阈值组的计算，获取nτ个阈值组的计算结果，对nτ个阈值组求平均以获得第一阈值组。
作为一种可选的实施方式，图2所描述的量化器还可以包括减法器40；减法器40包括第一输入端41、第二输入端42和输出端43；减法器40的输出端43连接量化选择电路10的输入端11。
需要强调的是，量化选择电路10的输入端11既接收阈值，也接收预测残差el，其在物理实现上，既可以是同一个输入端，也可能是分离的两个输入端，其具体的屋里实现方式，本发明实施例不做限制。
减法器40将第一信源数据xl与xl的预测值
Figure PCTCN2017071981-appb-000035
相减以获得xl的预测残差el；
Figure PCTCN2017071981-appb-000036
为对xl进行线性预测获得的预测值；其中，xl和
Figure PCTCN2017071981-appb-000037
分别经减法器40的第一输入端41和第二输入端42到达减法器40；el经减法器40的输出端43到达量化选择电路10的输入端11。
作为一种可选的实施方式，图2所描述的量化器还可以包括线性预测器50；线性预测器50包括第一输入端51、第二输入端52和输出端53；线性预测器的输出端53连接减法器40的输入端42；
线性预测器50根据第二信源数据xl-1和第二量化结果ul-1计算第一信源数据xl的预测值
Figure PCTCN2017071981-appb-000038
其中，xl-1和ul-1经线性预测器50的第一输入端51和第二输入端52到达线性预测器50；
Figure PCTCN2017071981-appb-000039
经线性预测器50的输出端53到达减法器40的第二输入端42。
作为一种可选的实施方式，线性预测器50可以包括预测系数计算电路501和线性预测电路502。
预测系数计算电路501包括输入端5011和输出端5012；线性预测电路502包括第一输入端5021、第二输入端5022、第三输入端5023和输出端5024。
预测系数计算电路501计算预测系数，预测系数包含第一预测系数，第一预测系数为对第一信源数据xl进行线性预测的预测系数。
上述计算预测系数，包括：
利用xl之前的Nblock个信源数据计算相关函数；
获取f个相关函数之后对f个相关函数求平均以获得平均相关函数，通过最小均方误差准则建立平均相关函数与预测系数的关系以求得预测系数。
第一预测系数经预测系数计算电路501的输出端5012到达线性预测电路502的第一输入端5021。
线性预测电路502根据上述第一预测系数、第二信源数据的预测值
Figure PCTCN2017071981-appb-000040
和ul-1计算
Figure PCTCN2017071981-appb-000041
包括：
计算
Figure PCTCN2017071981-appb-000042
与ul-1的和与第一预测系数的积作为中间量；
获取K个中间量之后对K个中间量求和以获得
Figure PCTCN2017071981-appb-000043
其中，K为线性预测的阶数。
Figure PCTCN2017071981-appb-000044
和ul-1分别经线性预测电路502的第二输入端5022和第三输入端5023到达线性预测电路502；
Figure PCTCN2017071981-appb-000045
经线性预测电路502的输出端5024到达减法器40的第二输入端42。
作为一种可选的实施方式，图2所描述的量化器还可以包括组帧电路60；组帧电路60包括第一输入端61、第二输入端62、第三输入端63、第四输入端64和输出端65；量化选择电路10的输出端还包括：第一输出端13和第二输出端14。
量化选择电路10的第一输出端13和第二输出端14连接组帧电路60的第一输入端61和第二输入端62；量化电路组20的输出端22连接组帧电路60的第三输入端63；预测系数计算电路501的输出端5012连接组帧电路60的第四输入端64。
量化选择电路10还用于获取ul的位置标识信息；ul的位置标识信息为el所属的分段；ul的位置标识信息和阈值分别经量化选择电路10的第一输出端13和第二输出端14到达组帧电路60；ul经量化电路组20的输出端22到达组帧电路60；第一预测系数经预测系数计算电路501的输出端5012到达组帧电路60。
组帧电路将ul、ul的位置标识信息、阈值和第一预测系数组成数据帧以进行传输。
由此可见，实施图2所描述的发明实施例，对预测残差进行分段量化，即根据预测残差的分布规律为每个分段选择匹配的量化方式，可以在满足量化信噪比要求的前提下，提高数据压缩率，即降低传输时需要的带宽；除此以外，将预测残差分段的阈值根据目标量化信噪比的要求确定，以保证量化后的数据可以达到理想的信噪比。
请参阅图3，图3为本发明实施例公开的一种量化方法的流程示意图。图3所描述的量化方法可包括以下步骤：
301、获取第一信源数据xl的预测残差el与阈值。
本发明实施例中，量化器获取第一信源数据xl的预测残差el与阈值，并且，该量化器可为作为数据发送端的终端设备中的电路模块。
本发明实施例中，上述el为xl与xl的预测值
Figure PCTCN2017071981-appb-000046
的差值，
Figure PCTCN2017071981-appb-000047
为对xl进行线性预测获得的预测值。而上述阈值为n-1个分别为τ1～τn-1的值组成的第一阈值组，将预测残差的取值范围划分为n个分段。
本发明实施例中，上述第一阈值组可为量化器中预先存储的，也可根据实时输入的数据计算获得。作为一种可选的实施方式，第一阈值组为nτ个阈值组的平均值；其中，阈值组的每两个阈值间的间隔根据预测残差的分布确定；而预测残差的分布根据先验知识确定。 根据上述特征，可通过如下的方式计算第一阈值组：
首先，计算阈值组：根据第一信源数据xl前一次的信源数据：第二信源数据xl-1、第二信源数据xl-1的预测残差el-1和第二信源数据的第二量化结果ul-1计算发送端信噪比
Figure PCTCN2017071981-appb-000048
判断SNRq是否大于SNRt；若SNRq＞SNRt，则将阈值组内各阈值分别增加δ；若SNRq＜SNRt，则将阈值组内各阈值分别减少δ；
循环执行nτ次上述阈值组计算过程，以获取nτ个阈值组的计算结果，对nτ个阈值组求平均以获得第一阈值组。
302、将el与阈值比较以确定el所属的分段，根据el所属的分段确定对el进行量化的目标量化方式。
本发明实施例中，不同的分段对应不同的量化方式，其中，可采用的量化方式包括：均匀量化、A率量化和μ率量化等；与对所有量化残差采用相同的量化方式相比，本发明实施例可以更好地匹配量化残差的分布规律，进一步降低量化结果所需的位宽。
303、利用目标量化方式对el进行量化以获得第一量化结果ul。
由此可见，利用图3所描述的量化方法，对预测残差进行分段量化，即根据预测残差的分布规律为每个分段选择匹配的量化方式，可以在满足量化信噪比要求的前提下，提高数据压缩率，即降低传输时需要的带宽。
请参阅图4，图4为本发明实施例公开的另一种量化方法的流程示意图。图4所描述的量化方法可包括以下步骤：
401、利用信源数据计算预测系数，预测系数包含第一预测系数。
本发明实施例中，量化器利用信源数据计算预测系数，并且，该量化器可为作为数据发送端的终端设备中的电路模块。另外，包括第一信源数据xl、第二信源数据xl-1等在内的信源数据，为数字化的数据，举例来说，可为进行数据传输时，数据发送端需要传输的正交调制信号。第一预测系数为对第一信源数据xl进行预测的预测系数。
本发明实施例中，可以通过线性预测的方式获得第一信源数据xl的预测值，在进行线性预测前，数据发送端每进入f*Nblock个数据更新一次预测系数。
在数据发送端刚开始发送数据时，获取Ninit个数据作为初始训练样本，为了降低计算复杂度，将初始训练样本分为F＝Ninit/Nblock段，以进行相关函数的计算。其中，相关函数的表达式为
Figure PCTCN2017071981-appb-000049
在表达式中，k1为相关函数的阶数，Nblock为计算相关函数所需要的信源信号的样本长度。
之后对每段长为Nblock的数据，利用移位相乘或者快速傅里叶变换的方式，计算出一组相关函数R(0)～R(K)，将求得的相关函数存入发送端的存储器之中，其中，针对初始训练样本计算出的相关函数的存储结构图如表1所示：
R1(0) R2(0) … RF(0)
R1(1) R2(1) … RF(1)
… … … …
R1(K) R2(K) … RF(K)
表1初始训练样本的相关函数的存储结构图
考虑到在DPCM框架中，预测系数需要由发送端传输到接收端，预测系数的传输会占用传输带宽，因此预测系数更新步长不能太短。同时，应当注意到，预测系数的更新需要能够跟上信源数据相关特性的时变特性，因此预测系数的更新步长也不能太长，需要进行折中。在实现中，我们选取更新步长Nupdate。Nupdate应该小于Ninit。
同样，将后续每段Nupdate个数据分为f＝Nupdate/Nblock段，每段长度为Nblock。
后续计算预测系数的过程，包括：
假设当前输入的数据为第一信源数据xl，利用xl之前的Nblock个信源数据计算相关函数，并循环进行相关函数的计算，将求得的f组相关函数存储进如表2所示的相关函数存储结构之中。
Rf+1(0) Rf+2(0) … RF(0) R1(0) … Rf(0)
Rf+1(1) Rf+2(1) … RF(1) R1(1) … Rf(1)
… … … … … … …
Rf+1(K) Rf+2(K) … RF(K) R1(K) … R f(K)
表2相关函数存储结构图
每次计算一组相关函数的时候，数组整体左移，新计算的结果放在数组的最右面。上面表格中，前面的Rf+1～RF是训练样本的第f+1～F段数据计算得到的相关函数，后面的R1～Rf是后续的一组数据计算得到的f组相关函数，之后进入数据都是采用此方式更新相关函数数组。采用如上方式，数据发送端存储的只是当前进行预测系数更新的f个相关函数，硬件开销适中，且硬件结构容易实现。
获取f个相关函数之后，对f个所述相关函数求平均以获得平均相关函数，通过最小均方误差准则建立平均相关函数与预测系数的关系；其中，假设计算出的平均相关函数为R(0)～R(K)，则平均相关函数与预测系数的关系可用如下线性方程组表达：
Figure PCTCN2017071981-appb-000050
因此，根据上述线性方程组，可计算得到预测系数。
402、根据第一预测系数、第二信源数据xl-1和第二量化结果ul-1计算第一信源数据xl的预测值
Figure PCTCN2017071981-appb-000051
本发明实施例中，可以通过线性预测的方式对第一信源数据xl进行预测。如图4a所示，图4a为本发明实施例公开的一种线性预测器的结构示意图。如图4a所示，计算第一信源数据xl前一次的信源数据(即第二信源数据)xl-1与预测系数β1的积作为中间量；之后重复 执行上述计算中间量的过程，获取K个中间量之后对K个中间量求和以获得第一信源数据xl的预测值
Figure PCTCN2017071981-appb-000052
在获得上述
Figure PCTCN2017071981-appb-000053
之后，执行步骤403，将第一信源数据xl与xl的预测值
Figure PCTCN2017071981-appb-000054
相减以获得xl的预测残差el，再将预测残差el输出到编码器中进行量化编码。
403、将第一信源数据xl与xl的预测值
Figure PCTCN2017071981-appb-000055
相减以获得xl的预测残差el。
根据图4a公开的线性预测器的结构示意图，预测残差el的计算公式可为：
Figure PCTCN2017071981-appb-000056
其中，K为线性预测器的预测阶数，其选取依赖于信源的相关特性，需要对信源数据做相关特性的分析，即对信源数据移位计算相关函数，以选取合适的间隔，使得两个采样数据的时延差大于该间隔时，这两个采样数据的相关性低于预期的阈值，则此间隔值即为需求的K值。
404、根据第二信源数据xl-1、xl-1的预测残差el-1、el-1的第二量化结果ul-1和目标量化信噪比SNRt确定阈值。
在线性预测的基础上，发送端需要对预测残差进行量化。量化方案的选取依赖于预测残差的统计特性，因此，首先对预测残差的概率分布进行统计。
举例来说，若预测残差的概率密度分布如图4b所示，从图中可以看出，预测残差的分布基本符合高斯分布的特性，小残差的出现概率较大，而大残差的出现概率较小。对于幅度较小的残差来说，利用较少的比特进行量化，即可以获得较高的精度，而对于幅度较大的残差来说，需要采用较多的比特进行量化，才能保证量化精度；因此，通过阈值对预测残差进行分段，每段选择与该分段相适应的量化方式，可以进一步降低量化结果所需的平均位宽。
本发明实施例中，上述阈值为n-1个分别为τ1～τn-1的值组成的第一阈值组，将预测残差的取值范围划分为n个分段，每段依据预测残差的分布规律采用不同的量化方式。
假定τ1为绝对值最小的阈值，各个阈值之间的间隔为di，j，则τk＝τ1+di，j，k＝2，…，n-1。因此，首先需要确定τ1的值。
请参阅图5，图5为本发明实施例公开的一种计算初始的τ1的值的流程示意图。在数据发送端刚开始发送数据时，可以通过如下步骤计算初始的τ1：
501、获取Ninit个信源数据作为初始训练样本。
502、计算初始训练样本中信源数据的预测残差。
503、采用ΔPCM的方式统计残差功率σ2。
504、构建均值为0，方差为σ2的高斯分布。
505、按照预设的比例要求，根据Q函数计算τ1的值。
在此之后，阈值每输入固定数目的信源数据进行一次更新。请参阅图6，图6为本发明实施例公开的一种阈值更新方法的流程示意图。作为一种可选的实施方式，可以通过如下方式进行阈值更新：
601、初始化参数count＝0，k＝0。同时初始化τ1～τn-1。
602、获取Nblock个信源数据，则计数器count++。
603、判断计数器count是否等于f，如果等于f，则执行步骤604，否则返回步骤602。
604、计算数据发送端的SNRq，同时count归零，计数器k++。
605、比较SNRq是否大于SNRt，若SNRq＞SNRt，则执行步骤606；若SNRq＜SNRt，则执行步骤607。
606、将阈值组内各阈值分别增加δ；其中δ表示每次微调的大小。
607、将阈值组内各阈值分别减少δ；其中δ表示每次微调的大小。
608、判断k是否等于nτ，如果是，则执行步骤609。如果否，则返回步骤602。
609、求nτ个阈值组的平均值以更新阈值。
本发明实施例中，根据数据发送端信噪比的情况定期调整阈值，以确保量化结果可以满足预设的信噪比要求。
405、将el与阈值比较以确定el所属的分段，根据el所属的分段确定对el进行量化的目标量化方式。
举例来说，若预测残差的分布符合高斯分布的特性，则小残差的出现概率较大，而大残差的出现概率较小。对于幅度较小的残差来说，利用较少的比特进行量化，即可以获得较高的精度，而对于幅度较大的残差来说，需要采用较多的比特进行量化，才能保证量化精度；因此，通过阈值对预测残差进行分段，每段选择与该分段相适应的量化方式，可以进一步降低量化结果所需的平均位宽。
406、利用上述目标量化方式对el进行量化以获得第一量化结果ul。
举例来说，若设定两个阈值τ1和τ2，将预测残差的取值范围分为三段，其中绝对值小于τ1的为小信号，绝对值在τ1和τ2之间的为中信号，绝对值大于τ2的为大信号；小信号采用的量化方式为μ＝1的对数压扩量化，中信号采用的量化方式为μ＝30的对数压扩量化，大信号采用均匀量化；若预测残差el为小信号，则采用μ＝1的对数压扩量化。
407、获取ul的位置标识信息。
本发明实施例中，ul的位置标识信息包括ul对应的el所属的分段，以便于将包含ul的数据帧发送至数据接收端时，数据接收端根据阈值和el所属的分段，确定数据发送端对el进行量化的量化方式，并由ul恢复出el。
408、将ul、ul的位置标识信息、阈值和第一预测系数组成数据帧以进行传输。
由此可见，利用图4所描述的量化方式，对预测残差进行分段量化，即根据预测残差的分布规律为每个分段选择匹配的量化方式，可以在满足量化信噪比要求的前提下，提高数据压缩率，即降低传输时需要的带宽；除此以外，将预测残差分段的阈值根据目标量化信噪比的要求确定，以保证量化后的数据可以达到理想的信噪比。
最后，以微波1024QAM正交调制数据为信源数据(整数位宽为3bit)来进行仿真，以具体说明上述的量化方法。
首先，对本发明实施例中涉及的参数进行配置，仿真时的参数配置如表3所示：
Ninit 300000 Nupdate 100000
Nblock 1000 K 20
F 300 f 100
表3仿真参数配置表
之后，设定两个阈值τ1和τ2，将预测残差的取值范围分为三段，其中绝对值小于τ1的为小信号，绝对值在τ1和τ2之间的为中信号，绝对值大于τ2的为大信号。小信号和中信号采用类似于μ律的对数压扩量化，大信号采用均匀量化。请参阅表4，表4为分段量化时的参 数配置信息。
Figure PCTCN2017071981-appb-000057
表4分段量化的参数配置
在本示例中，采用如下方案来确定中小信号的量化方案：
先对中、小信号的量化方式设定不同的μ值，之后分别评估中小信号量化信噪比，通过二分法查找最合适的μ值；
将y值在[0，1]之间平均分成32段，得到33个边界值yi，i＝0，1，...，32，其中y0＝0，y32＝1；
根据μ律公式来确定折线的斜率，其中，μ律公式为：
Figure PCTCN2017071981-appb-000058
考虑到硬件的实现性，利用32折线来逼近μ律，分别求出小信号和中信号的32段折线的斜率
Figure PCTCN2017071981-appb-000059
和
Figure PCTCN2017071981-appb-000060
i＝1，...，32；
利用每段折线的斜率及y值计算对应的x值，计算公式为：
x0＝0
Figure PCTCN2017071981-appb-000061
Figure PCTCN2017071981-appb-000062
利用求得的x轴的边界值求得码本codebook1和codebook2，即：
Figure PCTCN2017071981-appb-000063
Figure PCTCN2017071981-appb-000064
计算正数部分小中信号的量化码本：
codebooks＝codebook1*τ1
codebookm＝codebook2*(τ2-τ1)+τ1
将得到的量化码本取反，即可得到对应的负数部分的量化码本，组成量化码本共64个值。
利用上述量化码本将小信号和中信号进行量化，利用均匀量化的方式对大信号进行量化。对预测残差量化之后便要进行编码传输，由于采用了分段量化的方案，为保证接收端正确译码，组帧后的数据帧需要包括预测系数、预测残差所在分段的位置信息、阈值、预测残差。
根据上述方式仿真数据传输过程，得到的发送端信噪比为55.0435dB，发送的数据的平均比特为6.1780；请参阅表5，为本仿真结果与CPRI压缩和均匀量化在同等情况下，获得的结果的比较：
量化方式 均匀量化 CPRI压缩 本实施例方案
位宽(平均比特) 6bit 6bit 6.18bit
信噪比 22.83 34.87 55.04
压缩率 0 25％ 45％
表5仿真结果比较
由表5中的仿真结果比较可知，与均匀量化和CPRI相比，本发明实施例在量化信噪比和平均比特方面的性能有较大提升，可降低进行数字传输时所需的带宽。
数据接收端接收到上述数据帧之后，首先解帧得到阈值，根据阈值计算出码本；再解帧得到预测系数，将预测系数传递给线性预测器以获取信源数据的预测值；之后解帧得到预测残差所在分段的位置信息，便可利用译码器对量化结果进行译码；最后将译码结果与线性预测器得到的预测值相加，即可恢复出数据发送端发送的信源数据。
综上所述，实施本发明实施例，对预测残差进行分段量化，即根据预测残差的分布规律为每个分段选择匹配的量化方式，可以在满足量化信噪比要求的前提下，提高数据压缩率，即降低传输时需要的带宽。
本领域普通技术人员可以理解实现上述实施例方法中的全部或部分流程，该流程可以由计算机程序来指令相关的硬件完成，该程序可存储于计算机可读取存储介质中，该程序在执行时，可包括如上述各方法实施例的流程。而前述的存储介质包括：ROM或随机存储记忆体RAM、磁碟或者光盘等各种可存储程序代码的介质。

Claims (14)

  1. 一种量化器，其特征在于，包括：
    量化选择电路和至少两个量化电路；每个量化电路对应一种量化方式；
    所述量化选择电路的输出端连接所述量化电路的输入端；
    所述量化选择电路用于将第一信源数据xl的预测残差el与阈值比较，以确定所述el所属的分段，根据所述el所属的分段确定对所述el进行量化的目标量化方式，并将所述el输出至用于执行所述目标量化方式的目标量化电路；所述el为所述xl与所述xl的预测值
    Figure PCTCN2017071981-appb-100001
    的差值，所述
    Figure PCTCN2017071981-appb-100002
    为对所述xl进行线性预测获得的预测值；所述el经所述量化选择电路的输入端到达所述量化选择电路；所述el经所述量化选择电路的输出端到达所述目标量化电路；
    所述目标量化电路用于对所述el进行量化以获得第一量化结果ul。
  2. 根据权利要求1所述的量化器，其特征在于，所述量化器还包括阈值计算电路；
    所述至少两个量化电路的输出端连接所述阈值计算电路的输入端；所述阈值计算电路的输出端连接所述量化选择电路的输入端；
    所述阈值计算电路用于根据第二信源数据xl-1、所述xl-1的预测残差el-1、所述el-1的第二量化结果ul-1和目标量化信噪比SNRt确定所述阈值，所述xl-1为所述xl前一次的信源数据；所述el-1为所述xl-1与所述xl-1的预测值
    Figure PCTCN2017071981-appb-100003
    的差值，所述
    Figure PCTCN2017071981-appb-100004
    为对所述xl-1进行线性预测获得的预测值；其中，所述xl-1、所述el-1、所述ul-1和所述SNRt经所述阈值计算电路的输入端到达所述阈值计算电路；所述阈值经所述阈值计算电路的输出端到达所述量化选择电路。
  3. 根据权利要求2所述的量化器，其特征在于，所述阈值为n-1个分别为τ1～τn-1的值组成的第一阈值组，所述τ1～τn-1将预测残差的取值范围划分为n个分段；所述第一阈值组为nτ个阈值组的平均值；其中，阈值组的每两个阈值间的间隔根据所述预测残差的分布确定；所述预测残差的分布根据先验知识确定；
    所述阈值计算电路用于根据第二信源数据xl-1、所述xl-1的预测残差el-1、所述el-1的第二量化结果ul-1和目标量化信噪比SNRt确定所述阈值，包括：
    计算阈值组：
    根据所述xl-1、所述xl-1的预测残差el-1和所述第二量化结果ul-1计算发送端信噪比
    Figure PCTCN2017071981-appb-100005
    判断所述SNRq是否大于所述SNRt；若SNRq＞SNRt，则将所述阈值组内各阈值分别增加δ；若SNRq＜SNRt，则将所述阈值组内各阈值分别减少δ；
    获取nτ个阈值组的计算结果，对所述nτ个阈值组求平均以获得所述第一阈值组。
  4. 根据权利要求1～3中任意一项所述的量化器，其特征在于，所述量化器还包括减法器；
    所述减法器的输出端连接所述量化选择电路的输入端；
    所述减法器用于将第一信源数据xl与所述xl的预测值
    Figure PCTCN2017071981-appb-100006
    相减以获得所述xl的预测残差el；所述
    Figure PCTCN2017071981-appb-100007
    为对所述xl进行线性预测获得的预测值；其中，所述xl和所述
    Figure PCTCN2017071981-appb-100008
    经所述减法器的输入端到达所述减法器；所述el经所述减法器的输出端到达所述量化选择电路。
  5. 根据权利要求4所述的量化器，其特征在于，所述量化器还包括线性预测器；
    所述线性预测器的输出端连接所述减法器的输入端；
    所述线性预测器用于根据所述第二信源数据xl-1和所述第二量化结果ul-1计算所述第一信源数据xl的预测值
    Figure PCTCN2017071981-appb-100009
    其中，所述xl-1和所述ul-1经所述线性预测器的输入端到达所述线性预测器；所述
    Figure PCTCN2017071981-appb-100010
    经所述线性预测器的输出端到达所述减法器。
  6. 根据权利要求5所述的量化器，其特征在于，所述线性预测器包括：
    预测系数计算电路和线性预测电路；
    所述预测系数计算电路用于计算预测系数，所述预测系数包含第一预测系数，所述第一预测系数为对所述第一信源数据xl进行所述线性预测的预测系数；
    所述计算预测系数，包括：
    利用所述xl之前的Nblock个信源数据计算相关函数；
    获取f个所述相关函数之后对所述f个所述相关函数求平均以获得平均相关函数，通过最小均方误差准则建立所述平均相关函数与所述预测系数的关系以求得所述预测系数；
    所述第一预测系数经所述预测系数计算电路的输出端到达所述线性预测电路；
    所述线性预测电路用于根据所述第一预测系数、所述第二信源数据的预测值
    Figure PCTCN2017071981-appb-100011
    和所述ul-1计算所述
    Figure PCTCN2017071981-appb-100012
    包括：
    计算所述
    Figure PCTCN2017071981-appb-100013
    与所述ul-1的和与所述第一预测系数的积作为中间量；
    获取K个所述中间量之后对所述K个中间量求和以获得所述
    Figure PCTCN2017071981-appb-100014
    其中，所述K为线性预测的阶数；
    所述
    Figure PCTCN2017071981-appb-100015
    和所述ul-1经所述线性预测电路的输入端到达所述线性预测电路；所述
    Figure PCTCN2017071981-appb-100016
    经所述线性预测电路的输出端到达所述减法器。
  7. 根据权利要求6所述的量化器，其特征在于，所述量化器还包括组帧电路；
    所述量化选择电路的输出端连接所述组帧电路的输入端；所述目标量化电路的输出端连接所述组帧电路的输入端；所述预测系数计算电路的输出端连接所述组帧电路的输入端；
    所述量化选择电路还用于获取所述ul的位置标识信息；所述ul的位置标识信息包括所述el所属的分段；所述ul的位置标识信息和所述阈值分别经所述量化选择电路的输出端到达所述组帧电路；
    所述ul经所述目标量化电路的输出端到达所述组帧电路；
    所述第一预测系数经所述预测系数计算电路的输出端到达所述组帧电路；
    所述组帧电路用于将所述ul、所述ul的位置标识信息、所述阈值和所述第一预测系数组成数据帧以进行传输。
  8. 一种量化方法，其特征在于，包括：
    获取第一信源数据xl的预测残差el与阈值；所述el为所述xl与所述xl的预测值
    Figure PCTCN2017071981-appb-100017
    的差值，所述
    Figure PCTCN2017071981-appb-100018
    为对所述xl进行线性预测获得的预测值；
    将所述el与所述阈值比较以确定所述el所属的分段，根据所述el所属的分段确定对所述el进行量化的目标量化方式；
    利用所述目标量化方式对所述el进行量化以获得第一量化结果ul。
  9. 根据权利要求8所述的量化方法，其特征在于，所述将第一信源数据xl的预测残差el与阈值比较以确定所述el所属的分段之前，所述方法还包括：
    根据第二信源数据xl-1、所述xl-1的预测残差el-1、所述el-1的第二量化结果ul-1和目标量化信噪比SNRt确定所述阈值；所述xl-1为所述xl前一次的信源数据，所述el-1为所述xl-1与所述xl-1的预测值
    Figure PCTCN2017071981-appb-100019
    的差值，所述
    Figure PCTCN2017071981-appb-100020
    为对所述xl-1进行线性预测获得的预测值。
  10. 根据权利要求9所述的量化方法，其特征在于，所述阈值为n-1个分别为τ1～τn-1的值组成的第一阈值组，所述τ1～τn-1将预测残差的取值范围划分为n个分段；所述第一阈值组为nτ个阈值组的平均值；其中，阈值组的每两个阈值间的间隔根据所述预测残差的分布确定；所述预测残差的分布根据先验知识确定；
    所述根据第二信源数据xl-1、所述xl-1的预测残差el-1、所述el-1的第二量化结果ul-1和目标量化信噪比SNRt确定所述阈值，包括：
    计算阈值组：
    根据所述xl-1、所述xl-1的预测残差el-1和所述第二量化结果ul-1计算发送端信噪比
    Figure PCTCN2017071981-appb-100021
    判断所述SNRq是否大于所述SNRt；若SNRq＞SNRt，则将所述阈值组内各阈值分别增加δ；若SNRq＜SNRt，则将所述阈值组内各阈值分别减少δ；
    获取nτ个阈值组的计算结果，对所述nτ个阈值组求平均以获得所述第一阈值组。
  11. 根据权利要求8～10中任意一项所述的量化方法，其特征在于，所述将第一信源数据xl的预测残差el与阈值比较以确定所述el所属的分段之前，所述方法还包括：
    将第一信源数据xl与所述xl的预测值
    Figure PCTCN2017071981-appb-100022
    相减以获得所述xl的预测残差el；所述
    Figure PCTCN2017071981-appb-100023
    为对所述xl进行线性预测获得的预测值。
  12. 根据权利要求11所述的量化方法，其特征在于，所述将第一信源数据xl与所述xl的预测值
    Figure PCTCN2017071981-appb-100024
    相减以获得所述xl的预测残差el之前，所述方法还包括：
    根据所述第二信源数据xl-1和所述第二量化结果ul-1计算所述第一信源数据xl的预测值
    Figure PCTCN2017071981-appb-100025
  13. 根据权利要求12所述的量化方法，其特征在于，所述根据所述第二信源数据xl-1和所述第二量化结果ul-1计算所述第一信源数据xl的预测值
    Figure PCTCN2017071981-appb-100026
    之前，所述方法还包括：
    计算预测系数，所述预测系数包含第一预测系数，所述第一预测系数为对所述第一信源数据xl进行预测的预测系数；
    所述计算预测系数，包括：
    利用所述xl之前的Nblock个信源数据计算相关函数；
    获取f个所述相关函数之后对所述f个所述相关函数求平均以获得平均相关函数，通过最小均方误差准则建立所述平均相关函数与所述预测系数的关系以求得所述预测系数；
    所述根据所述第二信源数据xl-1和所述第二量化结果ul-1计算所述第一信源数据xl的预测值
    Figure PCTCN2017071981-appb-100027
    包括：
    计算所述
    Figure PCTCN2017071981-appb-100028
    与所述ul-1的和与第一预测系数的积作为中间量；
    获取K个所述中间量之后对所述K个中间量求和以获得所述
    Figure PCTCN2017071981-appb-100029
    其中，所述K为线性预测的阶数。
  14. 根据权利要求13所述的量化方法，其特征在于，所述方法还包括：
    获取所述ul的位置标识信息和所述阈值；所述ul的位置标识信息包括所述el所属的分段；
    将所述ul、所述ul的位置标识信息、所述阈值和所述第一预测系数组成数据帧以进行传输。
PCT/CN2017/071981 2017-01-20 2017-01-20 量化器与量化方法 WO2018133043A1 (zh)

Priority Applications (2)

Application Number Priority Date Filing Date Title
PCT/CN2017/071981 WO2018133043A1 (zh) 2017-01-20 2017-01-20 量化器与量化方法
CN201780056167.1A CN109690673B (zh) 2017-01-20 2017-01-20 量化器与量化方法

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2017/071981 WO2018133043A1 (zh) 2017-01-20 2017-01-20 量化器与量化方法

Publications (1)

Publication Number Publication Date
WO2018133043A1 true WO2018133043A1 (zh) 2018-07-26

Family

ID=62907500

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2017/071981 WO2018133043A1 (zh) 2017-01-20 2017-01-20 量化器与量化方法

Country Status (2)

Country Link
CN (1) CN109690673B (zh)
WO (1) WO2018133043A1 (zh)

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20070110327A1 (en) * 2005-11-16 2007-05-17 Samsung Electronics Co., Ltd. Method and apparatus for encoding image and method and apparatus for decoding image using human visual characteristics
CN102243876A (zh) * 2010-05-12 2011-11-16 华为技术有限公司 预测残差信号的量化编码方法及装置
CN103533351A (zh) * 2013-09-26 2014-01-22 西安空间无线电技术研究所 一种多量化表的图像压缩方法
CN105144288A (zh) * 2013-04-05 2015-12-09 杜比国际公司 高级量化器
CN106341689A (zh) * 2016-09-07 2017-01-18 中山大学 一种avs2量化模块和反量化模块的优化方法及系统

Family Cites Families (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP4168976B2 (ja) * 2004-05-28 2008-10-22 ソニー株式会社 オーディオ信号符号化装置及び方法
US8977544B2 (en) * 2011-04-21 2015-03-10 Samsung Electronics Co., Ltd. Method of quantizing linear predictive coding coefficients, sound encoding method, method of de-quantizing linear predictive coding coefficients, sound decoding method, and recording medium and electronic device therefor

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20070110327A1 (en) * 2005-11-16 2007-05-17 Samsung Electronics Co., Ltd. Method and apparatus for encoding image and method and apparatus for decoding image using human visual characteristics
CN102243876A (zh) * 2010-05-12 2011-11-16 华为技术有限公司 预测残差信号的量化编码方法及装置
CN105144288A (zh) * 2013-04-05 2015-12-09 杜比国际公司 高级量化器
CN103533351A (zh) * 2013-09-26 2014-01-22 西安空间无线电技术研究所 一种多量化表的图像压缩方法
CN106341689A (zh) * 2016-09-07 2017-01-18 中山大学 一种avs2量化模块和反量化模块的优化方法及系统

Also Published As

Publication number Publication date
CN109690673A (zh) 2019-04-26
CN109690673B (zh) 2021-06-08

Similar Documents

Publication Publication Date Title
CN101010964B (zh) 在可分级视频编码中的使用帧速率上变换技术的方法与设备
KR101497845B1 (ko) 가변장 부호화를 위해 신택스 요소 및 코드워드 사이의 매핑을 결정하는 방법 및 장치
CN114157722A (zh) 一种数据传输方法及装置
CN107343202B (zh) 基于附加码率的无反馈分布式视频编解码方法
CN107666472B (zh) 混合的数字-模拟编解码的方法和设备
CN106331706B (zh) 基于曲线拟合的率失真优化方法以及使用该方法的装置
Wang et al. Distributed image transmission using deep joint source-channel coding
Chen et al. Communication-efficient design for quantized decentralized federated learning
Perić et al. Quasilogarithmic quantizer for Laplacian source: support region ubiquitous optimization task
WO2018133043A1 (zh) 量化器与量化方法
CN108111255A (zh) 一种模拟编码中基于最大后验概率的译码方法
CN113630120B (zh) 一种结合1比特模数转换器的零时延通信方法及其应用
US10390025B2 (en) Method and apparatus for encoding and decoding video signal using improved prediction filter
CN116029345A (zh) 中间层特征压缩传输方法、压缩数据的解码方法及装置
Shin et al. RL-SPIHT: reinforcement learning-based adaptive selection of compression ratios for 1-D SPIHT algorithm
Leinonen et al. Distributed variable-rate quantized compressed sensing in wireless sensor networks
CN113890794B (zh) 一种大规模mimo信道估计方法、装置及存储介质
Jana et al. Optimality of KLT for high-rate transform coding of Gaussian vector-scale mixtures: Application to reconstruction, estimation, and classification
Shirazinia et al. Distributed quantization for measurement of correlated sparse sources over noisy channels
Leinonen et al. Channel-robust compressed sensing via vector pre-quantization in wireless sensor networks
Chen et al. A linear predictive coding based compression algorithm for fronthaul link in C-RAN
CN105706447A (zh) 动画图像编码设备、动画图像解码设备、动画图像编码方法、动画图像解码方法以及程序
Huang et al. MI-based Joint Source Channel Coding for Wireless Image Transmission
CN113711496A (zh) 削波信号的重构
JP2015032924A (ja) 変換量子化方法、変換量子化装置及び変換量子化プログラム

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 17893133

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 17893133

Country of ref document: EP

Kind code of ref document: A1