JP2019525502A - Digital filter with trust input - Google Patents

Abstract

The digital filter has an assigned filter function and has an assigned filter coefficient (bK), an input that receives an input sample (xk), another input that receives a confidence value (cK), and an output (out). Have. Each input sample value (xk) is associated with an input confidence value (cK), and the filter output (out) depends on the input sample, the input confidence value, and the filter coefficients. The filter includes a plurality of accumulators (acc01, acc02, acc03, acc04), and output samples are generated after a predetermined number of sample values with associated confidence values are input to the filter.

Description

  The present disclosure relates to a digital filter, specifically, a digital filter for noise suppression.

  Sampling a signal (significantly) faster than its actual information content change to sense an analog signal for processing in a digital device takes advantage of information redundancy to enhance the digitized signal. It is a general practice that makes it possible. Examples of such devices include capacitive touch sensing or touchless position and gesture sensing systems, digital voltmeters, thermometers, or pressure sensors.

Exemplary capacitive sensing systems that can be subject to significant noise are both from Microchip Technology Inc., the assignee of the present application. The application note AN1478 “mTouch ^™ Sensing Solution Acquisition Capacitative Voltage Divider” and AN1250 “Microchip CT Capit Pac Description”, which are available from and incorporated herein by reference in their entirety. .

  Another exemplary application is a touchless capacitive 3D gesture system, also known as GestIC® technology, manufactured by the applicant of the present application.

  Sensor signals are typically disturbed by various noise types such as broadband noise, harmonic noise, and peak noise. The latter two can arise from, for example, a switching power supply, which is addressed in an electromagnetic immunity standard test, such as IEC 61000-4-4.

  Signal collection can be interrupted in a planned or deterministic manner, for example when multiplexing several sensors in time, or by irregular events such as data transmission failures. Such discontinuities or missing samples can cause unwanted phase jumps in the signal. For digital filters that are designed for regular sampling intervals, this can corrupt filter timing and have a significant impact on their noise suppression performance.

  Similar to messages erased in the context of channel coding in digital communications (Blahut, 1983; Bossert, 1999), we are missing samples and samples that do not carry useful information due to, for example, peak noise. Is called an erasure.

  FIG. 1a shows a system 100 that performs a basic procedure for estimating a noisy actual value baseband signal. The analog / digital converter (ADC) 110 samples the signal at a (significantly) higher rate than the information change. The digital signal is then input to the low-pass filter 120 and decimate at the rate R by the R decimator 130. The downsampled results are further processed or simply displayed on a numeric display 140, for example as shown in FIG. 1a. Among them, the low-pass filter 120 can attenuate higher frequency components of broadband noise, but will not completely suppress noise peaks.

  The problem of peak noise suppression occurs in many applications such as image processing (T. Benazir, 2013), seismology, and medicine (B. Boashhash, 2004). The standard approach to combat peak noise is to apply a median filter or variant.

  An approach that suppresses peak noise but still smoothes the input signal excludes samples that are identified as noise peaks or outliers, or excludes, for example, n maximum and n minimum samples. , A filter that averages a portion of a sample in a time window (selective arithmetic average (SAM) filter or “sigma filter” (Lee, 1983)). Clearly, a SAM filter is a time-varying filter with a finite impulse response (FIR) that matches the time domain characteristics of its input signal.

  However, while excellent in the presence of noise peaks (ie, with erasures), the noise suppression characteristics of such a SAM averaging filter are not illustrated for a window length of 32 samples without the peaks. As shown in 1b, for example, other state-of-the-art filters that use a Hamming window as the impulse response, or filters whose frequency response is designed using the least squares method. For the filter amplitude response, the least squares filter solid curve and the Hamming filter dashed curve show improved sidelobe attenuation compared to the averaging filter (dotted curve) with a rectangular impulse response.

  There is a need for improved methods and systems for processing signals subject to noise. The present application is not limited to any of the sensor systems described above, but can be applied to any type of signal that is subject to noise and requires evaluation.

  According to an embodiment, the digital filter may comprise an assigned filter function, having assigned filter coefficients, an input for receiving input samples, another input for receiving confidence values, and an output, each input sample A value is associated with the input confidence value, each input sample is weighted with its associated confidence value, the filter output depends on both the input sample and the input confidence value, and the filter has a predetermined number of confidence values. An accumulator configured to accumulate weighted input samples, associated confidence values, confidence values weighted with assigned filter coefficients, and confidence weighted input samples further weighted with assigned filter coefficients .

  According to a further embodiment, the filter receives an input confidence value that is weighted with a coefficient from the coefficient set, receives an input confidence value that is weighted with a coefficient from the coefficient set, and generates a first cumulative value. A first branch having one accumulator, a second branch having a second accumulator for receiving an input confidence value and generating a second cumulative value, weighted by a coefficient from the coefficient set and the input confidence value A third branch having a third accumulator that receives the received input sample value and generates a third cumulative value; and a fourth accumulator that receives the confidence weighted input value and generates a fourth cumulative value. And a fourth branch. According to a further embodiment, the first cumulative value is subtracted from the constant value, and the result of the subtraction is divided by the second cumulative value, multiplied by the fourth cumulative value, and added to the third cumulative value. The first, second, third and fourth accumulators are subsequently cleared.

  According to another embodiment, the digital filter may comprise an assigned filter function and has first and second filter coefficient sets, an input for receiving input samples, another input received confidence information value, and an output. , Each input sample value is associated with an input confidence value, the filter output depends on both the input sample and the input confidence value, and the digital filter is input confidence weighted with coefficients from the first coefficient set. A first branch having a first accumulator for receiving a value and generating a first accumulated value; an input confidence value weighted with a coefficient from a second coefficient set; and a second accumulated value A second branch having a second accumulator to generate, and a third accumulator that receives an input sample value weighted with a coefficient from the first coefficient set and an input confidence value and generates a third cumulative value. And a fourth branch having a fourth accumulator that receives a weighted input value with a coefficient from the second coefficient set and an input confidence value and generates a fourth cumulative value. And.

  According to a further embodiment of the digital filter described above, the first cumulative value is subtracted from the constant value, and the result of the subtraction is divided by the second cumulative value, multiplied by the fourth cumulative value, Added to the accumulated value, the first, second, third and fourth accumulators are subsequently cleared.

  According to a further embodiment of any of the above digital filters, multiple instances of the filter are operated in parallel, each instance having a dedicated coefficient for a subset of input values and associated confidence values. Can be operated with. According to a further embodiment of any of the above digital filters, the confidence value is represented by a digital logic value. According to a further embodiment of any of the above digital filters, the constant value is the sum of all coefficients. According to a further embodiment of any of the above digital filters, the assigned filter function is a low pass filter function. According to a further embodiment of any of the above digital filters, the low pass is obtained from converting a high pass or band pass into an equivalent low pass region. According to a further embodiment of any of the above digital filters, the assigned filter function has only positive value coefficients or only negative value coefficients. According to a further embodiment of any of the above digital filters, the assigned filter function has at least one non-zero value coefficient having a magnitude different from another non-zero coefficient. According to a further embodiment of any of the above digital filters, the DC gain of the digital filter is constant or nearly constant.

  According to yet another embodiment, the filter system is first and second digital filters, each comprising an assigned filter function, assigned filter coefficients, an input for receiving input samples, a confidence value. , Each input sample value is associated with an input confidence value, each input sample is weighted with its associated confidence value, and the filter output is the input sample and Depending on both the input confidence values, the filter is a confidence-weighted input sample, an associated confidence value, a confidence value that is weighted by the assigned filter coefficients, and a confidence-weighted input sample that is further weighted by the assigned filter coefficients A first and second digital filter comprising an accumulator configured to accumulate Receiving the items, it may include a demultiplexer for generating an input sample for the first and second digital filters.

  According to a further embodiment of the filter system, the system receives a first outlier detector that receives the input samples for the first digital filter and generates an associated confidence value; and the second digital A second outlier detector that receives the input samples for the filter and generates an associated confidence value. According to a further embodiment of the filter system, the input samples for the first digital filter are high samples and the input samples for the second digital filter are low samples.

  According to yet another embodiment, the filter system comprises first and second digital filters, each comprising an assigned filter function, first and second filter coefficient sets, an input for receiving input samples. Each of the input sample values is associated with an input confidence value, and the filter output depends on both the input sample and the input confidence value, Each receives an input confidence value weighted with a coefficient from a first coefficient set, a first branch having a first accumulator that generates a first cumulative value, and a coefficient from a second coefficient set A second branch having a second accumulator that receives a weighted input confidence value and generates a second cumulative value, and a weighted input with a coefficient from the first coefficient set and the input confidence value. A third branch having a third accumulator for receiving a sample value and generating a third cumulative value; receiving an input value weighted with a coefficient from the second coefficient set and an input confidence value; And a fourth branch having a fourth accumulator for generating a cumulative value of 4, the system includes a demultiplexer that receives the input signal and generates input samples for the first and second digital filters. It has more.

  According to a further embodiment of the above filter system, the system receives a first outlier detector that receives the input samples for the first digital filter and generates an associated confidence value, and the second A second outlier detector that receives the input samples for the digital filter and generates an associated confidence value. According to a further embodiment of the filter system described above, the input samples for the first digital filter are high samples and the input samples for the second digital filter are low samples.

  According to yet another embodiment, the digital filter may comprise an assigned filter function, having assigned filter coefficients, an input for receiving input samples, another input received confidence information value, and an output, each input The sample value is associated with the input confidence value, the filter output depends on the input sample, the input confidence value, and the filter coefficients, the filter includes a plurality of accumulators, and the output sample is a predetermined number of The confidence value generated and associated after the sample value has been entered into the filter.

  According to yet another embodiment, a method of filtering digital input samples includes receiving a digital input sample value and an associated input confidence value, accumulating input confidence values weighted with coefficients from a coefficient set, Generating a cumulative value of 1; accumulating an input confidence value; generating a second cumulative value; accumulating an input sample value weighted with a coefficient from the coefficient set and the input confidence value; Generating a cumulative value of 3, and accumulating a confidence weighted input value to generate a fourth cumulative value.

  According to a further embodiment of the method, the method is to subtract a first cumulative value from a constant value, the result of the subtraction being divided by a second cumulative value and multiplied by a fourth cumulative value; Adding to the third cumulative value and subsequently clearing the first, second, third, and fourth accumulators. According to a further embodiment of the method, the constant value is the sum of all coefficients. According to a further embodiment of the method, the input confidence value is binary.

  According to yet another embodiment, a method of filtering digital input samples includes receiving a digital input sample value and an associated input confidence value, and accumulating an input confidence value that is weighted with a coefficient from a first coefficient set. Generating a first cumulative value; accumulating an input confidence value weighted with a coefficient from the second coefficient set; generating a second cumulative value; and Accumulating input sample values weighted with coefficients and input confidence values to generate a third cumulative value, and accumulating input values weighted with coefficients and input confidence values from the second coefficient set. Generating a fourth cumulative value.

  According to a further embodiment of the method, the method is to subtract a first cumulative value from a constant value, the result of the subtraction being divided by a second cumulative value and multiplied by a fourth cumulative value; Adding to the third cumulative value and subsequently clearing the first, second, third, and fourth accumulators. According to a further embodiment of the method, the constant value is the sum of all coefficients. According to a further embodiment of the method, the input confidence value is binary.

FIG. 1a shows an exemplary collection of analog signals with analog / digital conversion and conventional noise suppression. FIG. 1b shows the amplitude response of different low-pass filters. FIG. 2 shows typical tap weights for a low-pass filter with a finite impulse response. FIG. 3 shows an exemplary block diagram of a data source as an input source for a digital filter and associated trust generation. FIG. 4 shows an exemplary implementation of a digital filter with a trust input. FIG. 5 shows a system with an external trust generation controller. FIG. 6 shows an exemplary shift register implementation of a digital filter with a trust input. FIG. 7 shows an example of redistribution of erasure coefficient weights according to various embodiments. FIG. 7A illustrates another example of redistribution of erasure coefficient weights according to various embodiments of a two-level signal. FIG. 8 shows an example of peak noise suppression performance. FIG. 9 shows a comparison of filter coefficients and magnitude spectra with and without erasure. FIG. 10 shows an example of redistribution of erasure coefficient weights in a high-pass filter embodiment. FIG. 11 illustrates an embodiment of a non-contact gesture detection system that uses an alternating quasi-static electric field. FIG. 12 shows an example of a two-level measurement. FIG. 13 shows an implementation for packet data processing with minimum buffer requirements. FIG. 14 shows an implementation similar to that of FIG. 13, but with binary confidence ck = 0, 1 implemented with switches. FIG. 15 shows an implementation similar to FIG. 13 but with a differential stage after decimation. FIG. 16 is similar to FIG. 13, but shows an implementation with a general redistribution function. FIG. 17 shows an implementation with a binary confidence input implemented using a general redistribution function and a switch. FIG. 18 illustrates the splitting of the original filter impulse response into two. Figure 19 shows an implementation using two filters with confidence input for the ADC output samples r _k amplitude modulated.

  According to various embodiments, if the input signal is oversampled and noisy, a real value baseband signal, eg, a reliable estimate of the demodulated and downsampled GestIC® signal is obtained. Can be done. GestIC® devices, such as MGC3030 or MGC3130, or newer designs are available from the Applicant. For example, FIG. 11 illustrates an exemplary embodiment in which the controller 740 represents a GestIC® device. General description and design guides such as “GestIC® Design Guide” published online on January 15, 2015 are available from Microchip Technology Inc. Available from and incorporated herein by reference.

  A 3D gesture detection system 700 shown in FIG. 11 provides a transmission electrode 720 that can be formed by a frame structure as shown in FIG. However, the entire rectangular area under the receiving electrodes 710a..d can be used as a transmission electrode, or such an electrode can be divided into a plurality of transmission electrodes. The transmission electrode 720 generates an alternating electric field. The gesture controller 740 receives signals from the receiving electrodes 710a · d, and the signals may represent capacitance between the receiving electrodes 710a · d and the system ground and / or transmission electrodes 720. Gesture controller 740 may evaluate the signal and provide human device input information to processing system 730. This information may be 3D movement coordinates similar to 2D movement information generated by a computer mouse and / or may include commands generated from detected gestures.

  The problem faced in such applications is that the noise introduced into the sensor signal is both wideband and peak noise, and there is no known state-of-the-art approach to address both problems simultaneously. In GestIC® applications as well as other applications, some samples of the input signal may be lost or cannot be generated for various reasons. Although the adverse effects of input noise are obvious, irregularities in the input sampling interval lead to filter timing corruption and have a significant impact on noise suppression performance. Digital filters are typically designed for regular sampling intervals, others lead to undesirable phase jumps in the input signal from the filter perspective. The location of noise peaks and missing samples in the signal is determined by some other means, such as a peak noise detection system or a deterministic noise indicator. As mentioned above, there are standard approaches to combat wideband noise, ie apply a frequency (low pass) filter. And there is another standard approach to counter the peak noise, i.e. to apply a median filter over the window of signal samples.

  Such problems are particularly relevant with GestIC® systems, as described above, but these scenarios can be applied not only to GestIC® systems, but also to other sensor systems. May be related. Hence, the proposed measures can be applied to various signal sources.

  Because of the proposed filtering method, each input sample is associated with a confidence value. This confidence value indicates whether the relevant sample is an erasure, that is, whether it is actually a missing sample or a sample that is known not to carry useful information . The confidence value is assumed to be known by some other means. Such means include, for example, deterministic input or outlier detection methods such as Grubbs test (Grubbs, 1950), generalized extreme studentized deviation (GESD) test, or Hampel identifier (Hampel, 1974). be able to. In the context of image processing, confidence values are used, for example, as weights in least squares regression for improved alpha matting (J. Horentrup, 2014).

According to various embodiments, the following should be observed for suppressing broadband noise and ignoring undesirable, eg, noisy or missing samples (both issues The state-of-the-art approach to deal with was not known):
1. Erasures (eg, detected noise peaks) should not affect the filter output.
2. A constant filter gain should be provided at DC (for a constant input signal, the filter output signal level must also be constant).
3. In the absence of erasures, a gradual adaptation to several erasures should be provided while preserving the fault filter characteristics.

FIG. 2 shows the filter coefficients of a typical low-pass filter function (or “windowing” function), where an exemplary normalized Hamming window of length 8 is shown. Each filter coefficient defines its associated tap weight in the tap delay line implementation also shown in FIG. 2, with each tap aligned below its associated coefficient in the bar plot. We therefore use the terms “filter coefficients” and “tap weights” interchangeably. In this example, the tap delay line includes seven consecutive delay stages z ⁻¹ and eight tap weights. An exemplary input sample is also shown in FIG. Other sample structures can be applied with fewer or more steps.

According to various embodiments, given an input signal with a filter function and confidence information for each sample, the weights of the filter taps corresponding to the input samples with less confidence maintain the DC gain of the filter. Can be reduced. FIG. 2 shows a typical weight / coefficient distribution of the low-pass filter, formed by the low-pass filter of this example, eg, seven consecutive delay stages z ⁻¹ . Other sample structures can be applied with fewer or more steps.

  In the following, input samples that do not carry any useful information, and equivalently missing samples, can be considered erasures, and the corresponding samples are said to have zero confidence. Information about whether a sample is erased is assumed to be obtained from any other source or algorithm, for example, by comparing the sample to a threshold. The impulse response of each digital filter will be referred to as a “filter function”.

FIG. 3 shows a block diagram of an exemplary digital filter 300 with a trusted input and its input signal source. The data source 320 generates a signal x having samples x _k at discrete time k. The signal x is input to a peak noise (or “outlier”) detector 330, which associates the confidence value c _k with x _k to make each sample x _k “noisy” or “ Classified as “noisy”, for example, c _k = 1 means “no noise” or “full confidence in x _k ” and c _k = 0 means “noisy” or “in x _k ” It means “no trust”. That is, sample x _k with association c _k = 0 is an erasure. According to other embodiments, the trust information may also arise from some external means we refer to as external indicator 310. This external indication can be regarded as a deterministic trust input.

FIG. 5 shows a system 500 with a 2D capacitive touch detection and finger tracking system consisting of several sensor electrodes “2D electrode pattern” 520 and controller unit “2D touch controller” 510, which can be, for example, on a touch panel or touch display used. Around the 2D electrode pattern, four additional electrodes A, B, C, D are arranged, which are used with the 3D gesture controller 530 to form a capacitive 3D gesture detection system. When the 2D touch detection system 510, 520 is active, it interferes with the received signal of the 3D gesture detection system, ie the received data of the 3D gesture detection system is noisy and unusable. . In order to yield a functional 2D-3D capacitive sensor system 500, the 2D touch controller 510 is rarely active when no touch is detected, and while it is active, this is the 3D gesture controller 530 (dashed line). 3D gesture controller 530 then knows that its current received value is noisy and has an associated zero confidence. That is, when sample x _k is generated while 2D system 510, 520 is interfering with the received signal, the external indicator can set c _k = 0, otherwise c _k = 1 can be set. x _k and c _k are input to a digital filter with a confidence input in the 3D gesture controller 530.

  An example of a simple peak noise detector or outlier detector is the following: The last M samples at each time k

Average, and

Standard deviation of

Calculate

If, set c _k = 0, otherwise set c _k = 1.

(1. Main approach)
A standard digital finite input response (FIR) filter of order N is considered using the time-invariant filter function b = [b ₀ ; b ₁ ; ...; b _N ], and b _i , i = 0, 1,. Is a filter coefficient. For a given input signal x with samples x _k , the filter output signal y is

k is a discrete time index. The sum of all filter coefficients b _i is the direct current (DC) filter gain. For simplicity, the following is assumed without loss of generality.

Assume that for each input sample x _k , an associated confidence value c _k is provided. First, the confidence value is binary, c _k ε {0; 1}, c _k = 0 means “sample x _k is unreliable”, and c _k = 1 is “ Assume that it means “full trust”. From the time-invariant filter function b with the coefficients b _i , the time-varying filter function w (k) with the coefficients w _i (k) depending on the confidence value c _k will be calculated. The filter function w (k) is equal to the function b when the latest N + 1 input samples are all with full confidence, i.e. c _k-i = 1 for i = 0,. Is desired. However, one or more erasure, i.e., if there is an input sample _{x k-i} with associated _{c k-i = 0, x} k-i should not affect the output value _{y k.}

This is achieved by multiplying each filter coefficient b _i in (1.1) by the confidence value c _k−i of its associated input sample x _k−i . However, the modified filter coefficients

DC filter gain

Is no longer guaranteed to be constant.
As a result, the erasure filter weight must be distributed over the other filter coefficients. The preferred approach to do so is to use the following erasure weight:

The remaining non-erasing factor

Which results in a linear time-varying (LTV) filter with the following coefficients:

here,

And

And represents the relative confidence associated with the sample _xk-i at time k.

This is equivalent to replacing the erasure input samples by the average of the non-erasure input samples at each time instance k and setting c _k−i = 1 for all i. This method of algorithm implementation is similar to that performed with windowing and DC value calculation to estimate a finite set of DC values for successive samples, as well as a one-time or “per block” of each input sample. Of particular interest for processing.
Proof:

Since this procedure suggests overwriting the input data, it is not applicable for continuous filtering where each input sample affects multiple output samples, and calculating the output value is the input divided by the filter length Performed at a rate higher than the rate, we define the filter length as (N + 1), ie, the filter order plus one.

The redistribution of erasure filter coefficient weights is visualized in FIG. At the top, the 8 most recent input samples at time k are shown, of which _xk-4 and xk _-1 are erasures. The first plot shows the original filter, ie, the length 8 Hamming window coefficient bi, aligned with the lower shift register implementation. In the second plot, the values of coefficients b ₁ ′ (k) and b ₄ ′ (k) are set to zero because the corresponding input samples x _k−4 and x _{k−1 at} time k are erasures. Is done. The sum of the elimination coefficients is also shown on the right end side of the second plot. In the third plot, the erasure weights as shown to the right of the second plot are evenly redistributed over the coefficients assigned to the non-erasure input samples, yielding w _i (k). The added weights are shown with different diagonal lines in the third plot. The coefficient weights are shown in the lower shift register filter diagram as w ₀ -w ₇ in this embodiment.

When there is the next input sample at time k + 1, the samples and their corresponding confidence information move to the right within the filter's shift register, and the redistribution is again performed for the erasure shift pattern and is different The filter coefficient w _i (k + 1) is generated.

  An example of the noise suppression performance of the filter is shown in FIG. The top plot shows the filter input signal, a slowly varying information signal with additional Gaussian noise, with some noise peaks starting at the sample index 250 and with additional 60 Hz sinusoidal noise. The second plot shows that conventional low pass filtering with a length 64 Hamming function reduces higher frequency noise, but only blurs the noise peaks present in the input signal. However, since noise peaks are identified, they are completely suppressed according to various embodiments. The bottom plot of FIG. 8 demonstrates the benefit of using a Hamming cancellation filter function instead of a moving average. Compared to simple averaging over non-peak samples, ie selective arithmetic average filtering, Hamming cancellation filtering results in better suppression of broadband noise contained in the input signal and a smoother output.

  FIG. 9 shows how the erasure affects the frequency response of the filter. Here, the left side shows a typical low-pass filter using a rectangular window and a Hamming window and its associated magnitude spectrum. On the right side, the same filtering is shown using two erased samples. It can be observed that the spectrum of the Hamming cancellation filter resembles a rectangular cancellation filter, depending on the location of the cancellation.

(2. Generalization)
(2a) Non-binary reliable input)
Up to this point, the confidence input was binary, that is, the associated input sample was either used or not used to calculate the filter output value. However, considering the above note, it is easier to generalize the confidence input to take an actual value between 0 and 1, ie, c _k ∈ [0, 1], and the larger c _k is, We have confidence in the relevant sample x _k. Apart from the definition of _ck , equation (1.2) remains the same.

(2b) General redistribution function)
For binary confidence inputs, in equation (1.2), the erasure weights are evenly distributed to the other coefficients. The two terms in (1.2) can be interpreted as two parallel filter branches whose outputs are summed. The filter function in the first term is calculated from b and the confidence input, and the second term is a time-varying averaging filter. The latter can be generalized by introducing another FIR filter function g with a coefficient g _i resulting in:

It is also applicable with non-binary trust inputs c _k ε [0, 1]. G-weighted relative confidence associated with sample x _k-i at time k

Represent as

  Therefore, the filter output is given as:

A possible implementation of this filter is shown in FIG. 4, where the block designated “B” refers to a standard FIR filter with a filter function b, and the analog of the block designated “G” and the filter function g And

Represents a division by 1 by the input data of the block, that is, the output of this block is the inverse of the input multiplication (reciprocal). In this implementation, the filter coefficients (“B” and “G”) of the four filter blocks are constant. Of course, confidence-weighted input data values

As shown in the shift register implementation of FIG. 6 for a filter order N = 7 that highlights both delay lines of the confidence value c _k , the same input data, ie, _ck or

Filters “B” and “G” that process can share the same buffer. Here, the adaptability is included in the input signal of the filter block. Still, the implementation is equivalent to a single FIR filter with adaptive filter coefficients w _i (k).

The characteristic of the tap delay line implementation of the FIR filter of FIG. 6 is that the tap delay line of the confidence value TDL-C and the tap delay line of the confidence weighted input data TDL-XC are the same, that is, they have the same number of delays Are the same tap weights b ₀ , b ₁ ,... And g ₀ , g ₁ ,. Of course, one or the other delay line can be simplified depending on the input variable type, eg, binary confidence input. Even when g ₀ = g ₁ = g ₂ ..., the delay line or associated computation block can be simplified. Furthermore, since the constant coefficients can be compensated outside the tap delay line, the weights b ₀ , b ₁ ,... In the TDL-C become weights b ₀ , b ₁ ,. - and the constant multiple different whether, also, weight _{g 0, g} 1 in the TDL-C, ··· is, weight _{g 0, g} 1 in the TDL-XC, even if ... and if a constant multiple different problems do not become.

  For example, if gi = 1/8, each tap weight can also be set to 1, thus saving the multiplication and summing at the end of the tap delay line before the (1 / x) block Only divided by 8, which can also be done using a bit shift operation.

For a confidence value from a binary confidence input or a finite set of values, the computation of the confidence weighted input data value q _k realized by the multiplication of FIG. 6 is realized by a conditional statement, eg, IF / ELSE or SWITCH statement. For example, q _k is set to 0 if c _k = 0, and q _k is set to 1 if c _k = 1. Instead of before the delay line, a conditional statement can also be assigned to each tap of the delay line. The tap weight input value b _i · x _k−1 or g _i · x _k−1 is then added to the total output value of each delay line only if the associated c _k−1 is 1. In this case, the sample x _k can be input directly to the TDL-XC and does not need to be multiplied by c _k in advance. The same is true for TDL-C.

  The filter orders with impulse responses b and g need not necessarily be equal. Without loss of generality, the filter is defined to have an equal order N, where N is at least as large as the maximum of the filter order with b and g, and the unused coefficients are zero. It is assumed that there is.

  Selecting g = b results in another undesirable approach for redistributing erasure coefficient weights. The non-erasing filter coefficients are scaled by the same multiple, which is recalculated at each discrete time instance k, i.e.

It is.

(3. Exception handling)
With the denominator in (1.2) or (1.3), it is clear that the output value cannot be calculated if all the latest N + 1 input samples have zero confidence. A possible exception in such a case can be to repeat the most recent valid output sample, or the exception can be advanced to a subsequent processing stage.

(4. Windowing and DC value calculation for signals with more than one expected signal level in particular)
For a symmetric filter or “windowing” function b, taking a snapshot when convolving the input signal with the function b is equivalent to weighting the input samples with b and summing over a point-by-point product. Thus, when interested in the DC value of the window signal, the above concept can be applied equally. The main difference between windowing with DC computation and continuous filtering is that for the former, typically each input sample affects only a single output value, ie it is the input sample. This is a one-time process or a process for each block.

  In many applications, the measurement signal typically alternates between two different levels with additional noise. We refer to these levels as “high” and “low” signal levels. FIG. 12 shows exemplary measurements with such high and low levels. An example is amplitude modulation (AM) with synchronous sampling of the analog received signal at twice the carrier frequency, and the information is contained in the difference between the “high” and “low” signal levels. The method is applied, for example, with capacitive touch detection system or GestIC® technology. The measurement (or “received”) signal of such an AM sensor system can be demodulated, for example, by multiplying it alternately by +1 and −1, then a DC (zero frequency) value, That is, low pass filtering is performed to estimate the “average” difference between the actual information, “high” and “low” signal levels.

  Here, such a signal with two levels is considered, and in a standard application such a signal is input to a low-pass filter, and the terms “high” and “low” samples are two different signals. It is retained to represent a set of samples corresponding to either one of the levels. Deviations from their respective signal levels will be assumed to be caused by noise.

  When a “low” sample is detected as useless, eg, due to detected peak noise, the weight of its corresponding coefficient in the filter is set to zero and an erasure weight is placed on the other coefficients. I want to redistribute. However, in order to maintain the expected value of the filter output, the redistribution must only be for coefficients assigned to other “low” samples. Otherwise, the filter output will be closer to the “high” level than it should be because the coefficients assigned to the “high” samples will gain additional weight.

  In general, the samples of the input signal must be classified into a set of samples with the same expected value, and digital filtering with a confidence input, i.e. the redistribution of coefficient weights, is the whole of the coefficients assigned to each set. Must occur so that the target weight remains constant, which is most easily achieved when redistributing weights that are eliminated within one set within the same set.

  FIG. 7A shows an exemplary windowing and DC calculation of a signal with two levels. The handling of the erased values is done individually for the samples at the “high” and “low” signal levels, respectively, as shown in FIG. 7A. Every other filter coefficient is assigned to measurements from “high” and “low” signal levels. Graph a) in FIG. 7A shows the original filter coefficients. Graph b) separates the “high” level coefficients, and graph c) separates the “low” level coefficients. Graph d) in FIG. 7A shows that coefficient 3 corresponds to the sample at the “low” signal level and is erased. The weight is redistributed to other coefficients assigned to the samples with “low” signal levels. Graph e) shows the redistribution for “low” signal level coefficients. The lower graph f) shows the redistributed “low” and “high” signal level factor combinations. Therefore, the expected output value is retained. The method can be implemented with two data branches, one for samples with a “high” signal level and one for samples with a “low” signal level, and finally summing the outputs of the branches To do. Again, since this is a time-varying filter, the redistribution of filter coefficients is updated for each output sample.

(5. Reliable output)
The ability to process input confidence values immediately raises the question of whether confidence values can also be provided for each output sample. Such an output confidence measure should be independent of the input sample values, but should be a function of only the input confidence values and the filter coefficients, ie, for an integer M,

It is.

  A measure that satisfies (1.4) and is readily available is the sum of the filter coefficients weighted by their corresponding input confidence values, ie,

And
Bi ≧ 0 for all i,

If d _k ∈ [0, 1] is also applied, especially when all input samples have full confidence, d _k = 1 and all input confidence values are c _k−1 = 0. At some point, d _k = 0.

  Using such a confidence output, various proposed filters can be cascaded. This output can also be used for high-level control, for example “if the output confidence is too low, do not trigger a touch event”.

(7. Design rules)
The proposed approach is applicable to any low pass FIR filter. However, all filter coefficients should have the same sign and should be positive, for example. Primarily, the constant DC filter gain requirement can also be met when tap weights (some of them) are negative values. However, this will introduce the risk of creating undesirable filter characteristics, such as high pass characteristics, for some reliable input configurations.

  Furthermore, when input samples assigned to coefficients with large values are eliminated, the more similar the filter coefficient values are, the less problematic. Specifically, the square window, triangular window, hamming, and Hann window coefficients follow these rules.

  Apart from the original filter coefficient selection and exception handling, there are no further parameters to consider.

  This approach can be extended to high-pass filters. FIG. 10 shows an example of a high-pass filter in which the original filter coefficient weight repeats a positive sign and a negative sign as shown in the first plot from the top. Thus, according to an embodiment, first, the high pass filter coefficients are demodulated using alternating codes, as shown in the second plot from the top. The same method, similar to the low pass filter shown in FIG. 2, is then applied as shown in the third and fourth plots from the top. The modified weight, as shown in the fourth plot from the top, is then remodulated using the reverse alternating code. This results in a distributed weight as shown in the fifth plot from the top. Instead of remodulating the filter coefficients, the input signal of the filter can also be demodulated and filtered using an equivalent low pass with the coefficients from the top four plots before modulating the signal again. Can be done. If the input signal would be directly filtered using a high-pass filter or demodulated and filtered using an equivalent low-pass filter, then the high-pass filter It is important that the samples of the input signal will have a single expected value.

(8. Uses and use cases)
As mentioned above, the proposed concept can be applied to any filtering system, the input signal being sampled faster than the actual information change, the better the sampling rate. In particular, such systems include 3D capacitive sensor systems such as Applicants' GestIC system and 1D / 2D capacitive touch solutions. The filtering method can be further applied to other sensor signals and is not limited to capacitive sensor systems.

(9. Properties)
When initializing the filter with arbitrary but zero confidence data, it provides an estimate of the input signal from the turn-on time. Thus, the filter does not show a typical step response when filtering a signal with a non-zero average, and the filter condition is initialized at zero.

(Numerical example)
In the following, a numerical example of a calculation for an output value for a filter with a confidence input is given. The table below describes the input sample value x _{k at} time k and the associated confidence value c _k , the coefficient b _i of the original filter function b. Here, g is a constant, i.e., the erasure coefficient weights will be uniformly redistributed.

When k = 5 and k = 8, there is an erasure, ie c ₅ = c ₈ = 0. For example, since it has been detected value x ₅ and x ₈ are noise peaks, reliability is set to zero.

For filter initialization, all trust information in the filter's memory is set to zero (eg, zero trust before entering the first sample-trust pair (x ₀ , C ₀ )). By entering N samples with). This is indicated in the table by c _k = 0 for k <0. When k = 0, a sample x ₀ = 7 with confidence c _k = 1 is input to the filter. According to the above equation, the modified coefficient b _i ′ (k = 0) and coefficient w _i (k = 0) are calculated as follows:

That is, x ₀ = 7 is transferred directly to the output, ie y ₀ = x ₀ = 7.
When k = 9, when two erasures x ₅ and x ₈ are in the buffer of the filter, b _i ′ (k = 9) and wi (k = 9) are calculated as follows:

13-17 illustrate yet another implementation of an erasure filter for packet data processing. These embodiments use different but very memory efficient solutions.

  In packet data processing, each input sample affects only a single output value, so this implementation can be used instead of a tap delay line (buffer), as shown in the embodiment according to FIGS. 13-17. It can be done using the accumulator acc01. For each new data packet, according to some embodiments, these accumulators are set to zero, and then each input value is continuously added to them.

  Furthermore, an intermediate version with a shorter delay line than other embodiments with an accumulator is possible when the output data rate is lower than the input data rate by some decimation factor.

Hereinafter, the length of the data packet is represented as L, the order of the low-pass filter is represented as N, and N = L−1. Let x _k , k = 0, 1,... L−1 represent L samples of a data packet of length L. Each such sample is associated with a binary confidence value c _kε [0,1], ie 0 ≦ c _k ≦ 1, and c _k = 0, where x _k is, for example, peak noise or anomalous value Means that the detector will not carry any useful information and will not affect the filter output, c _k = 1 means that x _k will completely affect the filter output To do.

FIG. 13 shows an implementation for packet data processing with minimum buffer requirements, in other words, without any redundant buffers, and with only four accumulators acc01, acc02, acc03, and acc04. Since each data packet of length L is processed, L input samples x _k , k = 0, 1,. , L−1 are fed into the filter along with their associated confidence c _k . The corresponding filter coefficient b _k can be stored in the flash. According to some embodiments, after each data packet, accumulator acc0x needs to be reset to zero. Note that the operation on the right side of the decimation block represented by the box “L ↓” should be updated only once for each data packet. Thus, the box “L ↓” only forms a gate that, when the accumulator accumulates all input values of a packet having length L, transfers the accumulator value at that input to its output. Thus, after accumulating the L value, the box “L ↓” will output the accumulated value. FIG. 13 also includes a multiplicative inversion represented by (.) ⁻¹ , ie, the output value of the block is 1 divided by the input value.

k = 0, 1,. For symmetric filter impulse response is a _{b N-k} = _b k with respect to N, _{b N-k} in FIG. 13 can also be replaced by _{b k.}

For binary trust inputs with c _kε {0,1}, multiplication with c _k , ie weighting with confidence values, is implemented using switches (see FIG. 14) or conditional increments of variables. Can also be done.

  According to other embodiments, although not a typical use case, if the accumulator is not reset after each packet L, the filter is no longer a FIR filter but can be an IIR filter.

  Specifically, “confidence weighted” data refers to multiplying the data by a confidence value and also refers to opening or closing a data path using a switch. Similarly, while “weight” and “weighted” refer to “multiply” or “multiplied” respectively, multiplication can also occur by turning the data path on or off. .

14 for a pseudo-code software implementation for the calculation of the output for a single data packet using a symmetric filter impulse response and binary confidence values, samples x_k and associated confidence values c_k, filter coefficients b_k. Is shown below.
float acc01 = 0;
uint16 acc02 = 0;
float acc03 = 0;
float acc04 = 0;
float aux = 0; for (k = 0: N) aux + = b_k; // constant
for k = 0: N
if (c_k == 1) {
acc01 + = b_k;
acc02 ++;
acc03 + = b_k ^* x_k;
acc04 + = x_k;
}
}
float out = (aux-acc01) ^* (1 / acc02) ^* acc04 + acc03;
Instead of employing accumulators acc01, acc02, acc03, and acc04, which have input L samples and need to be reset to zero after calculating the final output value, the sum over the L input values is the current Before entering data for a packet, store the values of acc01, acc02, acc03, and acc04, and after entering the data for the current packet, from the respective values of acc01, acc02, acc03, and acc04 It can also be done by subtracting these stored values. This can be achieved, for example, by employing a primary CIC filter, which is described, for example, in IEEE Transactions on Acoustics, Speech, and Signal Processing, Vol. ASSP-29, no. 2, published in April 1981, pages 155-162. Known from Hogenauer, “An Economic Class of Digital Filters for Decimation and Interpolation” (which is a moving sum filter using decimation). This is shown in FIG. 15, with an additional differential stage that subtracts the previous accumulated and decimation value from the current value after the accumulator and decimation stage.

(General redistribution function)
The general redistribution function g can be employed for packet data processing as well as for continuous data processing. 13 and 14 change to FIGS. 16 and 17, respectively.

(Sample value with multiple expected values)
In many applications, such as capacitive sensing, the actual information is amplitude modulated, the ADC output data needs to be demodulated prior to low pass filtering, and both the ADC output value and the demodulated ADC output value are both Both have two expected values. However, digital filters with confidence inputs can only be applied directly to input samples with a single expected value. Thus, the demodulated ADC output values need to be divided into two sets with a single expected value for all samples in each set.

  Typically, the ADC alternately outputs samples at two different expected signal levels, which are represented as low and high samples, respectively, assigned a low sample to an even time index k, and to an odd time index. High samples are assigned.

Symmetric filter impulse response, i.e., b _i = b _N−i , i = 0, 1, 2,. Assume N. x _k ^(L) = x _2k and x _k ^(H) = x _{2k + 1} are introduced, and the same is true for the filter coefficients b _i ^(e) = b _2i and b _i ^(o) = b _{2i + 1} .

The original low-pass filter impulse response is split in two by separating the coefficient b _i with even i and the coefficient b _i with odd i. This is illustrated in FIG. 18 for an example of a true filter impulse response, which is a Hamming window of length L = 16.

Using these exemplary filter impulse responses, FIG. 19 shows how demodulated samples x _k are distinguished for even and odd values of k and divided into two data branches, q Is the base value of k / 2, with a confidence input according to FIG. 13, but two instances of a digital filter with different filter impulse responses (see FIGS. 18 and 19), with even and odd exponents k, respectively It is employed to process the samples x _k. Each data branch in this example has its own peak noise detector that produces a confidence value c. The sample rate on each branch is half the input sample rate, and the packet length considered on each branch is half the packet length of the input samples x _k , for example, where the original packet length is L = 16 and the packet length for each digital filter with trust input is L ′ = L / 2 = 8. The outputs of both filters are summed to produce the final result.

  The digital filter described above can be formed by hardware (eg, a programmable logic device) or software (eg, in a microcontroller, processor, or digital signal processor).

(Bibliography)
B. Boashhash, M .; M.M. (2004). Signal Enhancement by Time-Frequency Peak Filtering. IEEE Transactions on Signals Processing, 929-937.
Blahut, R.A. E. (1983). Theory and Practice of Error Control Codes. Reading, MA: Addison-Wesley.
Bossert, M.M. (1999). Channel Coding for Telecommunication. New York, NY, USA: John Wiley & Sons, Inc.
Commission, I.D. E. (Nd). Electrical Fast Transient / Burst Immunity Test (IEC 61000-4-4).
Grubbs, F.M. E. (1950, Mar). Sample criteria for testing out Observations. Ann. Math. Statist. , Pp. 27-58.
H. Hwang, R.A. H. (1995). Adaptive media filters: New algorithms and results. IEEE Transactions on Image Processing, 499-502.
Hampel, F.M. R. (1974). The influenza curve and its roles in robust establishment. Journal of the American Statistical Association, 69, 383-393.
J. et al. Horentrup, M .; S. (2014). Confidence-aware guided image filter. IEEE International Conference on Image Processing, (pp. 3243-3247).
K. Aiswarya, V.M. J. et al. (2010). A new and effective algorithm for the removal of high density salt and pepper noise in images and videos. Second International Conference on Computer Modeling and Simulation, (pp. 409-413).
Kalman, R.A. E. (1960). A new approach to linear filtering and prediction programs. Transactions of the ASME-Journal of Basic Engineering, 35-45.
Lee, J .; -S. (1983). Digital image smoothing and the sigma filter. Computer Vision, Graphics, and Image Processing, 255-269.
Microchip Touch and Input Sensing Solutions. (Nd). Retrieved from http: // www. microchip. com / touchandinputsensing
S. -J. Ko, Y .; H. (1991). Center weighted median filters and therair applications to image enhancement. IEEE Transactions on Circuits and Systems, 984-993.
T. T. et al. Benazir, B.M. I. (2013). Removal of high and low density impulse noise digital images using non linear filter. International Conference on Signal Processing, Image Processing and Pattern Recognition, (pp. 229-233).

Claims (38)

A digital filter comprising an assigned filter function, the digital filter having assigned filter coefficients, an input for receiving input samples, another input for receiving a confidence value, and an output;
Each input sample value is associated with an input confidence value, and each input sample is weighted with its associated confidence value,
The filter output depends on both the input sample and the input confidence value;
The filter comprises an accumulator, the accumulator further comprising a predetermined number of the confidence weighted input samples, the associated confidence value, the confidence value weighted with an assigned filter coefficient, and the assigned filter coefficient. A filter configured to accumulate the weighted confidence weighted input samples. A first branch having a first accumulator, wherein the first accumulator receives the input confidence value weighted with a coefficient from a coefficient set and generates a first cumulative value; A branch,
A second branch having a second accumulator, wherein the second accumulator receives the input confidence value and generates a second cumulative value;
A third branch having a third accumulator, wherein the third accumulator receives an input sample value weighted by a coefficient from the coefficient set and the input confidence value, and obtains a third accumulated value; A third branch to generate,
A fourth branch having a fourth accumulator, the fourth accumulator comprising: a fourth branch receiving the confidence weighted input value and generating a fourth cumulative value; Item 2. The filter according to Item 1.   The first cumulative value is subtracted from a constant value, and the result of the subtraction is divided by the second cumulative value, multiplied by the fourth cumulative value, and added to the third cumulative value, The filter of claim 2, wherein the first, second, third, and fourth accumulators are subsequently cleared.   A plurality of instances of the filter are operated in parallel, each instance being operated with a dedicated coefficient on a subset of the confidence values associated with the input samples. The filter according to item 1.   The filter of claim 4, wherein input samples are alternately assigned to one of the two instances of the filter.   The filter according to claim 1, wherein the confidence value is represented by a digital logic value.   The filter according to claim 3, wherein the constant value is a sum of all coefficients.   The filter according to claim 1, wherein the assigned filter function is a low-pass filter function.   9. The filter of claim 8, wherein the low pass is obtained from converting a high pass or band pass to an equivalent low pass region.   10. A filter according to one of claims 1-9, wherein the assigned filter function has only positive value coefficients or only negative value coefficients.   11. A filter according to one of the preceding claims, wherein the assigned filter function has at least one non-zero valued coefficient having a magnitude different from another non-zero coefficient.   The filter according to claim 1, wherein the DC gain of the digital filter is constant or substantially constant.   The filter according to claim 1, wherein the filter is formed by software. A digital filter having an assigned filter function, the digital filter having first and second filter coefficient sets, an input for receiving input samples, another input for receiving a confidence value, and an output; And
Each input sample value is associated with an input confidence value,
The filter output depends on both the input sample and the input confidence value;
The digital filter is
A first branch having a first accumulator, wherein the first accumulator receives the input confidence value weighted with a coefficient from the first coefficient set and generates a first cumulative value; The first branch;
A second branch having a second accumulator, wherein the second accumulator receives the input confidence value weighted with a coefficient from the second coefficient set and generates a second cumulative value; The second branch,
A third branch having a third accumulator, wherein the third accumulator receives an input sample value weighted with a coefficient from the first coefficient set and the input confidence value; A third branch generating a cumulative value;
A fourth branch having a fourth accumulator, wherein the fourth accumulator receives the input value weighted with a coefficient from the second coefficient set and the input confidence value; A filter comprising: a fourth branch for generating a cumulative value.   The first cumulative value is subtracted from a constant value, and the result of the subtraction is divided by the second cumulative value, multiplied by the fourth cumulative value, and added to the third cumulative value, 15. A filter according to claim 14, wherein the first, second, third and fourth accumulators are subsequently cleared.   16. The filter of claim 15, wherein multiple instances of the filter are operated in parallel, each instance being operated with a dedicated coefficient for a subset of confidence values associated with input samples.   The filter according to one of claims 14 to 16, wherein a confidence value is represented by a digital logic value.   18. A filter according to claim 15-17, wherein the constant value is the sum of all coefficients.   19. A filter according to one of claims 14-18, wherein the assigned filter function is a low pass filter function.   20. The filter of claim 19, wherein the low pass is obtained from converting a high pass or band pass to an equivalent low pass region.   21. A filter as claimed in one of claims 14 to 20, wherein the assigned filter function has only positive value coefficients or only negative value coefficients.   The filter according to one of claims 14-21, wherein the assigned filter function has at least one non-zero value coefficient having a magnitude different from another non-zero coefficient.   The filter according to one of claims 14-21, wherein the DC gain of the digital filter is constant or substantially constant. A filter system, the filter system comprising:
First and second digital filters, each comprising an assigned filter function, having assigned filter coefficients, an input for receiving input samples, another input for receiving a confidence value, and an output , Each input sample value is associated with an input confidence value, each input sample is weighted with its associated confidence value, and the filter output depends on both the input sample and the input confidence value; The filter comprises an accumulator, wherein the accumulator is further weighted with the confidence weighted input sample, the associated confidence value, the confidence value weighted with an assigned filter coefficient, and the assigned filter coefficient. First and second digital filters configured to accumulate the confidence weighted input samples ,
A demultiplexer that receives an input signal and generates input samples for the first and second digital filters. A first outlier detector that receives the input samples for the first digital filter and generates an associated confidence value;
25. The filter system of claim 24, further comprising: a second outlier detector that receives the input samples for the second digital filter and generates an associated confidence value.   26. The filter system according to claim 24 or 25, wherein the input samples for the first digital filter are high samples and the input samples for the second digital filter are low samples. A filter system, the filter system comprising:
First and second digital filters, each of which
An assigned filter function, having first and second filter coefficient sets, an input for receiving input samples, another input for receiving confidence values, and an output;
Each input sample value is associated with an input confidence value,
The filter output depends on both the input sample and the input confidence value;
Each digital filter
A first branch having a first accumulator, wherein the first accumulator receives the input confidence value weighted with a coefficient from the first coefficient set and generates a first cumulative value; The first branch;
A second branch having a second accumulator, wherein the second accumulator receives the input confidence value weighted with a coefficient from the second coefficient set and generates a second cumulative value; The second branch,
A third branch having a third accumulator, wherein the third accumulator receives an input sample value weighted with a coefficient from the first coefficient set and the input confidence value; A third branch generating a cumulative value;
A fourth branch having a fourth accumulator, wherein the fourth accumulator receives the input value weighted with a coefficient from the second coefficient set and the input confidence value; First and second digital filters further comprising: a fourth branch that generates a cumulative value;
A demultiplexer that receives an input signal and generates input samples for the first and second digital filters. A first outlier detector that receives the input samples for the first digital filter and generates an associated confidence value;
28. The filter system of claim 27, further comprising: a second outlier detector that receives the input samples for the second digital filter and generates an associated confidence value.   29. A filter system according to claim 27 or 28, wherein the input samples for the first digital filter are high samples and the input samples for the second digital filter are low samples. A digital filter comprising an assigned filter function, the digital filter having assigned filter coefficients, an input for receiving input samples, another input for receiving a confidence value, and an output;
Each input sample value is associated with an input confidence value,
The filter output depends on the input sample, the input confidence value, and the filter coefficient;
The filter includes a plurality of accumulators,
An output sample is generated after a predetermined number of sample values and associated confidence values are input to the filter. A method of filtering a digital input sample, the method comprising:
Receiving a digital input sample value and an associated input confidence value;
Accumulating the input confidence values weighted with coefficients from a coefficient set to generate a first accumulated value;
Accumulating the input confidence value to generate a second accumulated value;
Accumulating the input sample values weighted with a coefficient from the coefficient set and the input confidence value to generate a third accumulated value;
Accumulating the confidence weighted input value and generating a fourth accumulated value.   Subtracting the first cumulative value from a constant value, wherein the result of the subtraction is divided by the second cumulative value, multiplied by the fourth cumulative value, and the third cumulative value is 32. The method of claim 31, further comprising adding and subsequently clearing the first, second, third, and fourth accumulators.   The method of claim 32, wherein the constant value is the sum of all coefficients.   34. A method as claimed in any one of claims 31 to 33, wherein the input confidence value is binary. A method of filtering a digital input sample, the method comprising:
Receiving a digital input sample value and an associated input confidence value;
Accumulating the input confidence value weighted with a coefficient from a first coefficient set to generate a first accumulated value;
Accumulating the input confidence value weighted with a coefficient from a second coefficient set to generate a second accumulated value;
Accumulating input sample values weighted with coefficients from the first coefficient set and the input confidence values to generate a third accumulated value;
Accumulating the input value weighted with a coefficient from a second coefficient set and the input confidence value to generate a fourth accumulated value.   Subtracting the first cumulative value from a constant value, wherein the result of the subtraction is divided by the second cumulative value, multiplied by the fourth cumulative value, and the third cumulative value is 36. The method of claim 35, further comprising adding and subsequently clearing the first, second, third, and fourth accumulators.   37. The method of claim 36, wherein the constant value is the sum of all coefficients.   38. A method as claimed in any one of claims 35 to 37, wherein the input confidence value is binary.