GB2530376A

GB2530376A - Method for low-loss transformation and/or compression of a data stream

Info

Publication number: GB2530376A
Application number: GB1512116.3A
Authority: GB
Inventors: Holger Vogel; Jens-Rainer Hofft
Original assignee: Airbus DS Optronics GmbH
Current assignee: Hensoldt Optronics GmbH
Priority date: 2014-09-19
Filing date: 2015-07-10
Publication date: 2016-03-23
Anticipated expiration: 2035-07-10
Also published as: GB201512116D0; DE102014113520B4; GB2530376B; DE102014113520A1

Abstract

A method for low-loss transformation and/or compression of a data stream of input signals (S12Bit) with a first bit depth into a data stream of output signals (S8Bit) with a second bit depth using a transfer function, in which a noise component of the input signals (S12Bit) is taken into account. The transformation and/or compression may be at least partly reversible and the input and output signals are preferably image signals. The method may be used for contrast enhancement of a data stream of image signals, where after low-loss transmission/compression is done, with a first and second bit depths, a contrast enhancement is performed in the data stream of intermediate image signals and after which, a number of bits, in particular the lowest bits, of the intermediate image signals are removed from the data stream to obtain an output image signal with the output bit depth. Also claimed is a signal processing unit to perform the method.

Description

Method for low-loss transformation and/or compression of a data stream The invention relates to a method for low-loss transformation and/or compression of a data stream. The invention likewise relates to a method for contrast enhancement of a data stream and to an image processing unit.

Modern available image sensors (e.g. thermal image sensors, visual image sensors or the like) have a high dynamic range or a high bit or sampling depth of e.g. 12, 14 or 16 bits (referred to below as n bits in general). However, this internally available data cannot be transmitted or processed in many systems because often standard interfaces are present, which are frequently only designed for 8 bits. Likewise, it is often necessary to transmit smaller amounts of data when transmitting image signals.

Naturally, this problem also relates to other signal types, such as audio signals, measurement signals, voltage signals or the like, from very different sensors.

Known systems convert the signals which are obtained by the sensor and are in the sensor dynamic range e.g. linearly from the internal bit depth to the externally available bit depth. Here, a significant amount of image information is lost. Although lossless methods, such as e.g. the ZIP format or the like, can reduce the internal data to the available bandwidth, these do not provide directly readable signals (e.g. audio signals or image signals) but rather a data stream which must subsequently be calculated back to form an image or sound. Moreover, such methods require a comparatively high computational power in the system, in particular of the camera, and generate additional delay of the image data or the other data.

Proceeding herefrom, the present invention is based on the object of developing a method for low-loss transformation and/or compression of a data stream, which method avoids the disadvantages of the prior art, in which, in particular, merely a small amount of useful information is lost and in which the data are directly usable.

According to the invention, this object is achieved by a method for low-loss transformation and/or compression of a data stream of input signals with a first bit depth into a data stream of output signals with a second bit depth by means of a transfer function, in which a noise component of the input signals is taken into account.

In the method according to the invention, the output signal continues to consist of directly usable data or directly observable image data. The applied transformation and/or compression can be largely undone if necessary by means of an appropriate inverse function in the receiving system. Moreover, it is possible, for example in the case of image data, to apply already introduced, present image optimization functions directly to the generated image data. It is likewise possible to use local optimizations (manually or automatically) for image optimization for an observer.

Here, the invention proceeds from the concept of dividing the available bit resources as effectively as possible amongst the multifaceted signals by means of a noise-adapted transformation. If the signal distribution is assumed to be arbitrary in the entire possible value range, the ideal solution consists of conceding the same small number of bits to the noise of each arbitrary signal value.

By way of example (in the case of negligible so-called dark current noise), an image signal S has the noise or the signal noise AS = VSfR2, where R is a noise component. The noise component can be a read-out noise or an electron noise. It can also be referred to as nRMS. The signal noise AS, which is produced when recording images, is dependent on the brightness of each picture element. The dark picture elements or pixels have significantly less signal noise AS than bright picture elements. What can be identified from Figure 1 is that the signal noise AS substantially consists of the square root of the signal (brightness of the picture element) itself. In Figure 1, the image signal S is plotted on the horizontal axis (abscissa) and the signal noise AS is plotted on the vertical axis (ordinate).

Advantageously, only the signal noise AS is affected by the transformation and/or compression. Although a bit-identical image is no longer obtained, the essential information in the image is also maintained after the transformation or compression since changes merely relate to the noise. In image signals, brighter regions have a higher signal noise and as a result experience different encoding to dark regions. The noise component R independent of the image brightness, in particular the read-out noise or the electron noise, becomes more dominant in dark pixels when considering the noise, whereas it makes no substantial contribution in the case of bright pixels. In thermal images, the signal noise AS also scales with the square root of the signal S. However, thermal images are characterized by a clear background. Figure 2 depicts signal noise AS in image signals S with an offset. What can be seen from Figure 2 is that the curve therefore does not begin at an input signal of 0. The method according to the invention can advantageously also be applied to such offset-afflicted signals.

However, the compression effect may be differently pronounced in this case, dependent on the offset.

It is very advantageous if the transfer function is configured in such a way that the noise component is at least approximately constant in the data stream of output image signals. Since it is the noise which restricts the dynamic range of a signal representation, it is advantageous to find a transformation or compression under the condition that the noise of the target representation, which e.g. has a bit depth of 8 bits, is constant or at least approximately constant, in particular within a tolerance range.

The transformation and/or the compression can be at least partly reversible.

Thereupon, if necessary, the applied transformation can be completely or at least largely undone in the receiving system by means of an appropriate inverse function.

The transfer function can comprise at least one nth root function, in particular a square root function, in which the radicand is formed by at least the input signal S, and the noise component R. s +R2-R The transfer function can emerge as m = (2 -1) , where 5m can be 1nmax +R2 -R the output signal, S, can be the input signal, n can be the first bit depth, m can be the second bit depth, 5flmax can be a maximum input signal and R can be the noise component.

The transfer function can comprise at least one logarithm function, in which a logarithm of at least the input signal Sn and the noise component R is formed.

Therefore, it is also possible to use a logarithmic transfer function as an alternative to the nth root function. (2

The transfer function can emerge as S1, = / , where Sm can be Id 1+a nm

R

the output signal, S can be the input signal, n can be the first bit depth, m can be the second bit depth, can be a maximum input signal, a can be an adjustment parameter for fitting to further system properties and R can be the noise component.

The input signal S can have a proportional statistical noise component in addition to the noise component R. The transfer function can comprise at least one area hyperbolic cosine function of the input signal, of the noise component and of the proportional statistical noise component.

If a statistical noise term proportional to the signal is present, use can be made of an area hyperbolic cosine function. The statistical noise term can be characterized by a fraction a, which is of the order of e.g. approximately 0.5 to approximately 2%.

The transfer function can emerge as 1+2a.2Sn +4(1+2a.25n)2 _1+4a.2R2 In I I 1+2aR S = (2 -1) ______________________ where Sm can be the n1+2ai2m +J(1+2aj2snmax)2 _1+4a2R2 1+2aR output signal, S can be the input signal. n can be the first bit depth, m can be the second bit depth, 5nmax can be a maximum input signal, a can be a fraction for the proportional statistical noise component and R can be the noise component.

It is advantageous if the transfer function is implemented as a lookup table.

The transformation and/or compression can be performed by means of a lookup table, as a result of which hardly any computational power is required in the system.

This also avoids an additional delay in image signals, which can be undesirable or, depending on application, even unacceptable in camera systems.

The input signals can be embodied as input image signals and the output signals can be embodied as output image signals.

Claim 12 relates to a method for increasing the contrast of a data stream of input image signals with an input bit depth, which is transferred into a data stream of output image signals with an output bit depth, wherein: a) a low-loss transformation and/or compression of the data stream of input image signals with the input bit depth as first bit depth into a data stream of intermediate image signals with the input bit depth as second bit depth is performed by means of the method according to the invention for low-loss transformation and/or compression of a data stream, after which b) a contrast enhancement in the data stream of intermediate image signals is performed, and after which c) a number of bits, in particular the lowest bits, of the intermediate image signals are removed in the data stream of intermediate image signals in such a way that a data stream of output image signals with the output bit depth is obtained.

In the case of low-contrast images, use is often made of histogram equalization functions or contrast enhancements. However, since the described methods consider the signal noise (at the recording) such corrections should not be performed before applying the method according to the invention.

In step c), the number of bits to be removed can emerge from the difference between the number of bits of the input bit depth and the number of bits of the output bit depth.

Claim 14 specifies a signal processing unit.

Advantageous embodiments and developments of the invention emerge from the dependent claims. Below, exemplary embodiments of the invention are described in terms of the principle thereof on the basis of the drawing.

In detail: Figure 1 shows a diagram of signal noise in image recordings; Figure 2 shows a diagram of signal noise in image recordings with an offset; Figure 3 shows a simplified diagram of an nth root transfer function; Figure 4 shows a simplified diagram of a 12-bit to 8-bit lookup table for an nth root transfer function; Figure 5 shows a simplified diagram of a logarithmic transfer function; Figure 6 shows a simplified diagram of a dependency between signal and signal noise in a logarithmic transfer function; Figure 7 shows a simplified diagram of a 12-bit to 8-bit lookup table for a logarithmic transfer function; Figure 8 shows a simplified diagram of a 12-bit to 8-bit lookup table for an area hyperbolic cosine function with a noise term of 0.5%; and Figure 9 shows a simplified diagram of a 12-bit to 8-bit lookup table for an area hyperbolic cosine function with a noise term of 2%.

Below, the invention is described on the basis of input image signals as input signals and output image signals as output signals. Naturally, other types of signals, such as S voltage signals, measurement signals, audio signals, etc, may come into question as input or output signals.

According to the invention, a method is proposed for low-loss transformation and/or compression of a data stream of input image signals Sl2Bit with a first bit depth of 12 bits into a data stream of output image signals S86 with a second bit depth of S bits by means of a transfer function, in which a noise component of the input image signals Sl2Bit is taken into account. Here, a first bit depth of 12 bits and a second bit depth of S bits are assumed in an exemplary manner. Naturally, any other bit depths are also conceivable here.

The method according to the invention can run on a signal processing unit or an image processing unit, which is configured to carry out the method, for example as a computer program in a storage element of the signal processing unit or image processing unit.

In general, the signal S has signal noise AS = VS + . This signal noise AS, which is produced when images are recorded, depends on the brightness of each picture element or pixel. Dark pixels have significantly less noise than bright pixels.

Essentially, the signal noise AS consists of the square root of the signal S (brightness of the respective pixel) itself. A noise component R which is independent of the image brightness and which may be e.g. the read-out noise or the electron noise of the CCD elements of the image sensors (thermal image sensors, visual image sensors or the like) becomes more dominant in dark pixels when the signal noise is considered, whereas it is less prominent in bright pixels. The noise component R can also be referred to as nRMS, where e denotes the charge electrons.

Since it is the signal noise AS which limits the dynamic range of the signal representation S or S, (with a first bit depth n), it is advantageous to find a transformation and/or compression of the data stream of input image signals S,, in which signal noise of the target representation or of the output image signals Sm (with a second bit depth m) is constant with a signal noise of ASm.

The following explanations describe a target representation S8Bt with a bit depth m of 8 bits. However, the method is accordingly also applicable to other bit depths or sampling depths m of the target representation S. The constant 255 is then, in general, replaced by 2ni-1 from which ASm or 5m emerges as a result.

The transfer function can be designed in such a way that the noise component R in the data stream of output image signals is at least approximately constant. The transformation and/or the compression can at least be partly or completely reversible.

Advantageously, the transfer function can be implemented as a lookup table, e.g. on the signal processing unit.

In a first embodiment, the transfer function can comprise at least one nth root function, in particular a square root function, in which the radicand is formed by at least the input signal S and the noise component R. Therefore, in general, the transfer function can emerge as I --R2 R Sm =(2 -1) n. Here, S, is the output image signal, Sn is the input 5nmax +R2 -R image signal, n is the first bit depth, m is the second bit depth and R is the noise component. 5nmax represents the maximum input image signal and can be equated to the maximum charge capacity of a CCD element, e.g. amounting to 20,000 electrons (referred to as e in the present case). The noise component R or the read-out noise of the CCD element may comprise e.g. 20 electrons (e). The aforementioned values constitute typical parameters for a CCC sensor. Other sensor types (e.g. CMOS sensors, etc) may have different parameters. However, the functional relationship is substantially the same.

The nth root transfer function can be established as follows: the following results if constant noise is assumed in the output image signals Sm or S S8BH (1) AS = dS8Bt AS = dSBBt Js + = const.

8Bit dS dS which can be solved by (2) =S36, =2.const.(JS+R2 R) wherein the integration constant was selected in such a way that S = 0 c=> S8BjI = 0.

The constant const. is dependent on the maximum input image signal 5nmax and on the independent noise component R and accordingly is calculated as follows: 255/2 (3) ASBBt =const. = ________ V5nmax +R2 -R According to this, the charge signal (number of charge electrons e) S transforms into an 8-bit signal SBBft according to: -VS+R2_R (4) 8Bi! -255 _________ JSnmax +R2 -R Herein, the photo charge signal S is represented by: (5) 5nmax5 Figure 3 depicts an nth root transfer function. Here, the input image signal Sl2Bjt is plotted on the horizontal axis (abscissa). The output image signal SBBjt is plotted on the vertical axis (ordinate).

The inverse function for the nth root transfer function possibly required for display purposes is derived correspondingly from the 8-bit output image signal SgBt as follows: (6) S =[ (V5nmax +R2 -R2.

Figure 4 depicts the application of the nth root transfer function in the form of a lookup table in an exemplary manner. What can be seen from Figure 4 is that the 1-bit difference signal is, independently of the signal level, only slightly greater than the signal noise. That is to say the transformation is almost maintaining the dynamic range. In Figure 4, the 12-bit input image signal S12Bd is plotted on the abscissa axis.

The left-hand ordinate axis shows the 8-bit output image signal SSBft and the right-hand ordinate axis shows the signal noise.

Alternatively, in a second embodiment, the transfer function may comprise at least one logarithm function, in which a logarithm of at least the input signal S and the noise component R is formed. Here, in general, the transfer function can emerge as (2 = / , where S, can be the output signal, S, can be the input ln1+a.°nm

R

signal, n can be the first bit depth, m can be the second bit depth, Snmax can be a maximum input signal, a can be an adjustment parameter for fitting to further system properties and R can be the noise component.

A logarithmic transfer function with an adjustment parameter a = 1 is shown in Figure 5. Here, once again, the input image signal S12Bd is plotted on the abscissa axis and the output image signal SSBiI is plotted on the ordinate axis.

Then, the 8-bit signal representation correspondingly results to be 255lnI1+a

Q R

BBit ( Id 1+a

R

As already explained above, the adjustment parameter a can be used for fitting to further system properties.

In this case, the signal noise results to be 8 AS -255 VS+R2 () BBit ( R lnH+ct. nmax R) a Figure 6 shows the dependence between the 12-bit input image signal Sl2Bjt and the signal noise AS for five different adjustment parameters a, namely 0.0025, 0.5, 1, 2 and 4. The overall noise is determined almost completely by the -constant -read-out noise in the case of a small input image signal, which is very clearly visible in the selected function, while the overall noise is dominated by the nth root of the input image signal for larger input image signals and the influence of the read-out noise becomes ever weaker. In general, the so-called dark current noise is of no significance and can be ignored in the present case. What is furthermore visible from Figure 6 is that the noise begins at relatively large bit values in the case of a small input image signal Sl2Bjt and quickly falls under the 1-bit limit in the case of moderate brightness, in order finally to achieve a fraction of a bit (approximately 15% in the case of an adjustment parameter of a = 1) in the region of great brightness. This means that, in the case of brightness, one bit of the 8-bit output image signal SSBjt represents a substantially larger difference in the original signal than the noise equivalent and hence the possible signal-to-noise ratio is perceptibly reduced. This can also be identified if the inverse function

S

(9) S=-1+a I -1 and the derivative R) s831t dS R ( S ( S = 1+a nmax I *lnfl+a rrnax 255a R) R (10) R S-( = aln1+aonmax 255 R are considered.

Figure 7 shows an exemplary implementation of a lookup table for a logarithmic transfer function with an input image signal S126 with a bit depth of 12 bits and an S output image signal S36 of 8 bits. Here, different adjustment parameters cx of 0.025, 0.5, 1, 2 and 4 are once again assumed. The 12-bit input image signal S12B, is plotted on the abscissa axis and the 8-bit input image signal SBB is plotted on the ordinate axis. The logarithmic function appears suitable for reducing the number of quantization levels since it approximately maintains contrast relationships or modulation relationships independently of the value of the background radiation.

However, the constancy of the noise representation is better ensured in the case of an nth root function as a transfer function. However, if a very small adjustment parameter, for example cc = 0.0025, is used in a logarithmic transfer function, a result similar to the nth root transfer function can be obtained.

If the input image signal S, has a proportional statistical noise component or noise term in addition to the noise component R, the transfer function may, in a third embodiment, comprise at least one area hyperbolic cosine function of the input image signal Sn, of the noise component R and of the proportional statistical noise component.

Here, in general, the transfer function results to be ln1+22 +4(i +2ai2SnT -1 + 4a2R2 1+2cxR Sm=(2m_1) In 1+2aj2Snmax +(i +2aj2Snmaj -1+ 4a2R2 1+2cxR where can be the output signal, S can be the input signal, n can be the first bit depth, m can be the second bit depth, Snmax can be a maximum input signal, a can be a fraction for the proportional statistical noise component and R can be the noise component.

The fraction ct can lie in the order of 0.5 to 2%. If the fraction ct is introduced into the formula for the noise AS, the following is obtained: (11) AS=S+R2+(ctS)2 If the procedure is now as in the case of the nth root transfer function, an 8-bit representation SB of the signal S is sought after, by means of which it is possible to satisfy the condition that the variance AS&Bft of the 8-bit output signal S36 is constant for all values of S. This is equivalent to the following condition: (12) AS88 = d555,, AS = dS55 JS + R2 +(ctS)2 = const. =: c0.

This differential equation is solved by an area hyperbolic cosine function (the inverse of the hyperbolic cosine): -c0 1+ 2a2S +V(1+2aj2S)2 -1+ t2R2 ( ) 85it n 1-f-2aR wherein the integration constant is once again selected in such a way that S = 0 <=> S8BiI = 0.

The constant c0 = ASBBt is determined by the condition SBBft (Qmax) = 5nmax = 255, to be precise 255ct (14) c0 _ASgBt = I lnl+2am +(1+2aj2Snmax)2 -1 +4a2R2 1+2ctR and hence i+2a.2S+(1+2a.2S)2 _i+4a.2R2 In I I 1+2ctR (15) SBBt(S)=255 1+2ai2Snmax +(1+2aj2Snmax)2 -1+ 4a2R2 1+2ctR The inverse function in this respect is given by -A* arccosh(B + C. S8Bt /255)-i -2 2a (16) 2 A.lnB+C.SBB/255+](B+C.SBBft/255) -i-i 2a where the following abbreviations are used: (17) A=iji_4ctj2R2 C=arccosh=lnh/A+Ji/A2_i) and B = arccoshl+2ai -C = ln(i+2aj2Snm)/A + V(i+2aj2Snmax)/A2 i) C. As an example for a statistical noise term cz of 0.5%, Figure 8 shows the implementation of a 12-bit to 8-bit lookup table with the area hyperbolic cosine law.

Analogously, Figure 9 shows such an implementation of a statistical noise term ct of 2%. In Figure 8, the lower horizontal axis shows a photocharge with a number of electrons e and the upper horizontal axis shows the 12-bit input image signal S12B.

The 8-bit output image signal S8Bt is plotted on the left-hand ordinate axis and the signal noise of the 8-bit output image signal S8BII is plotted on the right-hand ordinate axis. As mentioned above, Figure 9 plots the same variables for a noise term ct of 2%. In principle, a so-called fixed pattern noise represents an aforementioned statistical noise term. However, the latter can be eliminated subsequently, i.e. after the compression or transformation, if the inhomogeneity produced by the fixed pattern noise is known at any one time. By way of example, this can be implemented by averaging over a plurality of images with different image content.

Furthermore, a method for increasing the contrast of a data stream of input image signals with an input bit depth, which is transferred into a data stream of output image signals with an output bit depth, is proposed, wherein: a) a low-loss transformation and/or compression of the data stream of input image signals with the input bit depth as first bit depth into a data stream of intermediate image signals with the input bit depth as second bit depth is performed by means of a method according to the invention for low-loss transformation and/or compression, after which b) a contrast enhancement in the data stream of intermediate image signals is performed, and after which c) a number of bits, in particular the lowest bits, of the intermediate image signals are removed in the data stream of intermediate image signals in such a way that a data stream of output image signals with the output bit depth m is obtained.

In the case of low contrast images, use is often made of histogram equalization functions or contrast enhancements. However, since the described methods consider the signal noise (at the recording), it is advantageous not to carry out such corrections before applying the method according to the invention for low-loss transformation and/or compression, for example by means of the nth root transfer function. By way of example, additional enhancement methods can be applied to images with a low contrast after the nth root transfer function has been applied.

A contrast-enhanced signal ScE is given by the following formula: (18) SCE(SL)= L SL min flax -where L + 1 = 2 is the number of possible greyscale values of an input image signal SL, which in turn denotes the n-bit representation of a charge signal S, possibly after a transformation by means of e.g. the nth root transfer function. [mm and [max denote S the number of the lowest and highest histogram value, the occurrence of which differs from 0.

The variance ASCE of the contrast-enhanced signal S is: (19) ASCE(SL)= L ASL.

max -mm In the case where SL represents the direct linear contrast-enhanced n-bit representation, the variance ASL of SL is given by: (20) ASL = IS+R2 nmax and varies e.g. from 0.1% of L (S = 0) to 0.7% of L (S = Smax) for = 20,000 e (electrons), R = 20 e, n = 12.

In the case where SL represents the n-bit representation after applying the transformation with the nth root transfer function, the variance of SL is given by (21) ASL= L 2lSnmax -f-R2 and it has the constant value of 0.4% of L in this example, independently of S. Using the aforementioned equation (19), the following is obtained for the respective variances of the contrast-enhanced signal: (22) ASCE(SL)= L L VS+R2 max -mm nmax for the linear case and (23) ASCE(SL)= L L Lmax 1-min 2tjSnmax +R2 for the case with the nth root transformation.

To the extent that a contrast enhancement is possible, i.e. if Lmax + < L, the variance is multiplied by the same gain factor as the signal itself in both cases.

Therefore, the low bits, which are intended to be removed during compression, are in fact less significant in respect of the variance than without a contrast enhancement.

The transformation by means of an nth root transfer function with a constant variance is very well suited for low-loss compression.

The signal after the histogram equalization SHE is given by the formula (24) SHE(SL)=L, SHmax -SHmin wherein SH denotes numbers in the cumulative histogram with SHmjn SH(Lnijn) and SHmax: SH(Lmax) = SH(L) = overall number of picture elements.

The histogram is denoted by H, the following is obtained for the variance ASHE of the signal 5HE after the histogram equalization: (25) AS (S)=L ASH(SL) L H(SL) AS max -mm max -mm In the case of a constant histogram H within the boundaries of the occurrence of [max and Lmjn, the formula reduces to the formula for the direct contrast enhancement (19).

However, in the general case, H varies within these boundaries and therefore the variance multiplier of ASL will also vary about a constant value, meaning a worse behaviour in relation to the low-loss compression for a specific number of 5L. values.

Therefore, it appears to be very advantageous within the meaning of low-loss compression to apply an n-to-n bit transformation using an nth root transfer function to the full n-bit representation of the signal, perform a direct contrast enhancement thereafter and only subsequently cut off the lower (n -m) bits (input bit depth -output bit depth).

S

Claims

Patent claims: 1. Method for low-loss transformation and/or compression of a data stream of input signals (S12Bft) with a first bit depth into a data stream of output signals (S36t) with a second bit depth by means of a transfer function, in which a noise component of the input signals (Sl2Bjt) is taken into account.
2. Method according to Claim 1, wherein the transfer function is configured in such a way that the noise component is at least approximately constant in the data stream of output signals (S8Bi,).
3. Method according to Claim 1 or 2, wherein the transformation and/or the compression is at least partly reversible.
4. Method according to one of Claims 1, 2 or 3, wherein the transfer function comprises at least one nth root function, in particular a square root function, in which the radicand is formed by at least the input signal (Sl2Bjt) and the noise component.
5. Method according to one of Claims 1 to 4, wherein the transfer function results to S +R-R be Sm = (2 -1) , where Sm is the output signal, S is the input signal, *!JSnmax +R2 -R n is the first bit depth, m is the second bit depth, Snmax is a maximum input signal, and R is the noise component.
6. Method according to one of Claims 1, 2 or 3, wherein the transfer function comprises at least one logarithm function, in which a logarithm of at least the input signal (Sl2Bjt) and the noise component is formed.
7. Method according to Claim 6, wherein the transfer function results to be (2 -i)ini+a.J Sm = ,, , where Sm is the output signal, S is the input signal, n is lnl+a.mRthe first bit depth, m is the second bit depth, Snmax is a maximum input signal, a is an adjustment parameter for fitting to further system properties, and R is the noise component.
8. Method according to one of Claims 1, 2 or 3, wherein the input signal (Sl2Bjt) has a proportional statistical noise component in addition to the noise component and wherein the transfer function comprises at least one area hyperbolic cosine function of the input signal (S12Bj,), of the noise component, and of the proportional statistical noise component.
9. Method according to Claim 8, wherein the transfer function results to be ln1+22 +4(1+2aj2Sn)2 _1+4a2R2 1+2ctR m =(2 -1) , where S, is the output n1+2a25nm +J(1+2aj2snmax)2 -1+ 4a2R2 1+2aR signal, Sn is the input signal, n is the first bit depth, m is the second bit depth, Snmax is a maximum input signal, a is a fraction for the proportional statistical noise component, and R is the noise component.
10. Method according to one of Claims 1 to 9, wherein the transfer function is implemented as a lookup table.
11. Method according to one of Claims ito 10, wherein the input signals are input image signals (S12BjI) and the output signals are output image signals (SaBji).
12. Method for contrast enhancement of a data stream of input image signals (S12Bft) with an input bit depth, which is transferred into a data stream of output image signals (SaB) with an output bit depth, wherein: a) a low-loss transformation and/or compression of the data stream of input image signals (Sl2Bjt) with the input bit depth as first bit depth into a data stream of intermediate image signals with the input bit depth as second bit depth is performed S by means of a method according to Claim 11, after which b) a contrast enhancement in the data stream of intermediate image signals is performed, and after which c) a number of bits, in particular the lowest bits, of the intermediate image signals are removed in the data stream of intermediate image signals in such a way that a data stream of output image signals (Sa6t) with the output bit depth is obtained.
13. Method according to Claim 12, wherein, in step c), the number of bits to be removed results from the difference between the number of bits of the input bit depth and the number of bits of the output bit depth.
14. Signal processing unit configured to perform a method according to one of Claims ito 13.