CN104392423A - Real-time balance based infrared image detail enhancement algorithm - Google Patents

Abstract

The invention provides a real-time balance based infrared image detail enhancement algorithm aiming at characteristics that the infrared image gray level is concentrated in distribution and low in contrast ratio. The real-time balance based infrared image detail enhancement algorithm avoids details from losing during high-dynamic-range compression, overcomes the defects that a traditional histogram equalization algorithm is excessively large in noise and suddenly changed in brightness and dynamically combines a traditional gray level enhancement algorithm. Verification in FPGA hardware platform shows that the real-time balance based infrared image detail enhancement algorithm is good in contrast ratio and timeliness and meanwhile highlights the details.

Description

Based on real time equaliser infrared image detail enhancement algorithms

Technical field

The present invention relates to infrared imagery technique field, especially a kind of based on real time equaliser infrared image detail enhancement algorithms.

Background technology

Present high-performance thermal infrared imager can the very large raw image data (12 ~ 14bit) of out-put dynamic range, exceeds the dynamic range (0 ~ 8bit) of general display device.When after the original image obtaining great dynamic range, must remap, by its dynamic range compression so that display.Process means conventional at present comprise contrast strengthen, automatic growth control and histogram equalization process.

Infrared image mostly for low dynamic range echograms data, and only considers the application in real-time system to comparison Enhancement Method.Contrast improvement strengthens with details should be just more meaningful for original high dynamic range.Because sampled in the original image signal that obtains by A/D include most complete information in scene, and there is many faint detailed information.

Automatic growth control and histogram equalization are most popular image display technologies in infrared imaging system, first automatic growth control rejects the extremum in scene, then by the dynamic range linear mapping of entirety to 8, this compression roughly will inevitably cause the loss of a large amount of minor detail.The grayscale mapping function of image is adopted the cumulative distribution function of original image by histogram equalization, be similar to meet by the pixel distribution of image after histogram equalization and be uniformly distributed, so histogram equalization emphasizes the gray level that the frequency of occurrences is larger more, inevitably there is enhancing in the image through histogram equalization, homogeneous area noise amplifies the problems such as bleaching effect.

Consider the deficiency of automatic growth control and histogram equalizing method, many more complicated methods are suggested as Retinex4, but these methods are mainly for visible images.Namely for visible images, there is good effect, and to infrared image poor effect, the particularly problem of noise amplification, algorithm based on layered shaping framework can obtain reasonable effect in these class methods, its main thought is that image is divided into levels of detail and Primary layer, then processed respectively, but the ideal operand of the method effect is large and easily produce gradient flop phenomenon.

Summary of the invention

The present invention will solve the shortcoming of above-mentioned prior art, provide a kind of have good contrast and timeliness based on real time equaliser infrared image detail enhancement algorithms.

The present invention solves the technical scheme that its technical matters adopts: this based on real time equaliser infrared image detail enhancement algorithms, first linear mapping is adopted by original 14 replacement response to 10, and utilize quick bilateral filtering for image smoothing, squelch, detail extraction, then process in two steps in 10 bit data, the first step adopts linear contrast to strengthen and 10 infrared images is mapped to 8 bit image Img1; Second step adopts the plateau equalization improved that 10 bit images are mapped to 8 Img2; The weights of map image Img1 and Img2 are dynamically determined again according to the grey level range of input picture; Finally with the weights determined, Img1, Img2 are merged, obtain the image that final details strengthens.

As preferably, described quick bilateral filtering process in the following ways:

$I_{\mathrm{bf}}(i,j)=\frac{1}{k(i,j)}\underset{(i^{\′},j^{\′})\∈s_{i,j}}{\mathrm{\Σ}}{g}_s(i-i^{\′},j-j^{\′})g_r(I_{\mathrm{in}}(i,j)-I_{\mathrm{in}}(i^{\′},j^{\′}))I_{\mathrm{in}}(i^{\′},j^{\′})$

Wherein, k (i, j) is normalization coefficient, g _sfor spatial domain standard gaussian kernel function, g _rfor intensity domain standard gaussian kernel function, I _infor input picture,

$g_s=e^{-\frac{|{(i-i^{\′})}^2+{(j-j^{\′})}^2|}{2{{\mathrm{\δ}}_s}^2}},g_r=e^{-\frac{\left|{(I_{\mathrm{in}}(i,j)-I_{\mathrm{in}}(i^{\′},j^{\′}))}^2\right|}{2{{\mathrm{\δ}}_r}^2}}$

Wherein g _sfirst time Gauss's computing, according to formwork calculation space weight, available similar integer approximation, δ _sget the integer between [5,24];

G _rsecond time Gauss computing, calculate weight according to field similarity, can adopt look-up table to realize herein, look-up table scope is [0, Max-Min], δ _rget integer between [15,25].

As preferably, the linear contrast of described improvement strengthens in the following ways: after carrying out statistics with histogram to input picture, respectively from minimum and most high grade grey level, abandon several number of pixels, then grey scale mapping scope is obtained for [Hist_min, Hist_max], then by this grey scale mapping in [20,240] that are applicable to eye-observation.

As preferably, the plateau equalization of described improvement in the following ways:

In formula, R is the tonal range of output image, can according to image dynamic conditioning.

rangeHist is input picture grey level range, and the maximal value of R is 255.

Inventing useful effect is: the loss of details when this algorithm avoids high dynamic range compression, overcome the shortcoming of the excessive and jump in brightness of traditional histogram equalization algorithm noise, and the traditional gray level of dynamic bind strengthens algorithm.By verifying on FPGA hardware platform, show that this algorithm is while outstanding details, has good contrast and timeliness.

Embodiment

Embodiment:

One, algorithm content:

1) adopt linear mapping by original 14 replacement response to 10, and utilize quick bilateral filtering process, obtain image Img; 2) adopt improvement linear mapping Img to 8, obtain image Img1;

3) adopt platform improving Histogram Mapping Img to 8, obtain image Img2;

4) Img1 and Img2 is merged with formula Img_out=0.5*Img1+0.5*Img2.

Two, quick two-sided filter:

Effect: the bilateral filtering after improveing for the ease of hardware implementing.

Algorithm: utilize look-up table to reduce calculated amount.

$I_{\mathrm{bf}}(i,j)=\frac{1}{k(i,j)}\underset{(i^{\′},j^{\′})\∈s_{i,j}}{\mathrm{\Σ}}{g}_s(i-i^{\′},j-j^{\′})g_r(I_{\mathrm{in}}(i,j)-I_{\mathrm{in}}(i^{\′},j^{\′}))I_{\mathrm{in}}(i^{\′},j^{\′})$

Wherein, k (i, j) is normalization coefficient, g _sfor spatial domain standard gaussian kernel function, g _rfor intensity domain standard gaussian kernel function, I _infor input picture,

$g_s=e^{-\frac{|{(i-i^{\′})}^2+{(j-j^{\′})}^2|}{2{{\mathrm{\δ}}_s}^2}},g_r=e^{-\frac{\left|{(I_{\mathrm{in}}(i,j)-I_{\mathrm{in}}(i^{\′},j^{\′}))}^2\right|}{2{{\mathrm{\δ}}_r}^2}}$

Wherein g _sfirst time Gauss's computing, according to formwork calculation space weight, available similar integer approximation, δ _sget the integer between [5,24];

G _rsecond time Gauss computing, calculate weight according to field similarity, can adopt look-up table to realize herein, look-up table scope is [0, Max-Min], δ _rget integer between [15,25].

Three, look-up tables'implementation:

If current maximum is 1023, look-up table length is 1024,

Lut＝zeros(Max,1)；

for x＝1:Max

cc＝-(x*x/(2*sigma_r*sigma_r))^2；

Lut(x)＝round(exp(cc)*10000)；

end

First amplify 10000 times, retain the difference between data, again divided by 10000 during computing, index is the absolute value of current pixel and field pixel difference.The template of Gauss can adopt Gaussian integer template to be similar to for the first time, or similar method is searched in employing: first amplify certain multiple and ensure that it is integer and has distinction, again divided by same multiple during computing.

Advantage: can reach the bilateral similar effect with standard, speed wants fast a lot, and hardware is also convenient to be realized.

Four, the linear contrast improved strengthens:

For infrared image, sampling raw data is out 14, can obtain 16384 grades of gray scales, but the gray scale of real image only concentrates on certain interval in the middle of 0 ~ 16384.Therefore, algorithm is strengthened to linear contrast herein and done improvement: after statistics with histogram is carried out to input picture, respectively from minimum and most high grade grey level, abandon several number of pixels, then grey scale mapping scope is obtained for [Hist_min, Hist_max], then by this grey scale mapping in [20,240] that are applicable to eye-observation.The most gray levels be in like this in the middle of gray level just can demonstrate contrast preferably.

Five, the plateau equalization improved:

In order to overcome the overall brightness of image when the excessive problem of plateau equalization contrast when grey level distribution is concentrated and scene move, the problems such as comparatively macromutation occur, use platform improving histogram equalization herein, formula is as follows

$D_T\left(k\right)=\frac{(F_T\left(k\right)-F_T\left(0\right))\×R}{(F_T\left(M\right)-F_T\left(0\right))}$

In formula, R is the tonal range of output image, can according to image dynamic conditioning; l _outthe starting point that the gray level after mapping exports, can according to image dynamic conditioning.

R changes along with the grey level range of tablet pattern.When the grey level distribution of input picture is relatively concentrated, the gray level of output image just should correspondingly reduce.When the grey level range of input picture is wider, the grey level range of output image then should correspondingly increase.According to this meaning, provide the definition of R:

$R=\frac{255}{1+\frac{255}{\mathrm{RangeHist}}}$

RangeHist is input picture grey level range, and the maximal value of R is 255.

When the gray level difference of image is very little, because RangeHist is also correspondingly very little, then the grey level range R of output image is now also smaller.After equalization processing, what gray level can not be drawn opens especially, can solve the problem that when grey level distribution is relatively concentrated, contrast is excessive.And have a buffer value 255, RangeHist can be avoided to suddenly change and cause the sudden change of R.Therefore, the problem of the overall brightness sudden change of image when the method can solve that in traditional plateau equalization algorithm, image moves.

In addition to the implementation, the present invention can also have other embodiments.All employings are equal to the technical scheme of replacement or equivalent transformation formation, all drop on the protection domain of application claims.

Claims (5)

1., based on a real time equaliser infrared image detail enhancement algorithms, comprise the following steps:

1) adopt linear mapping by original 14 replacement response to 10, and utilize quick bilateral filtering process, obtain image Img;

2) adopt improvement linear contrast to strengthen and 10 infrared image Img are mapped to 8, obtain image Img1;

3) adopt platform improving histogram equalization that 10 infrared image Img are mapped to 8, obtain image Img2;

4) dynamically determine the weights of map image Img1 and Img2 according to the grey level range of input picture, finally with the weights determined, Img1 and Img2 is merged.

2. according to claim 1 based on real time equaliser infrared image detail enhancement algorithms, it is characterized in that: described quick bilateral filtering process in the following ways:

$I_{\mathrm{bf}}(i,j)=\frac{1}{k(i,j)}\underset{(i^{\′},j^{\′})\∈s_{i,j}}{\mathrm{\Σ}}{g}_s(i-i^{\′},j-j^{\′})g_r(I_{\mathrm{in}}(i,j)-I_{\mathrm{in}}(i^{\′},j^{\′}))I_{\mathrm{in}}(i^{\′},j^{\′})$

Wherein, k (i, j) is normalization coefficient, g _sfor spatial domain standard gaussian kernel function, g _rfor intensity domain standard gaussian kernel function, I _infor input picture,

$g_s=e^{-\frac{|{(i-i^{\′})}^2+{(j-j^{\′})}^2|}{2{{\mathrm{\δ}}_s}^2},g_r=e^{-\frac{\left|{(I_{\mathrm{in}}(i,j)-I_{\mathrm{in}}(i^{\′},j^{\′}))}^2\right|}{2{{\mathrm{\δ}}_r}^2}}}$

Wherein g _sfirst time Gauss's computing, according to formwork calculation space weight, available similar integer approximation, δ _sget the integer between [5,24];

G _rsecond time Gauss computing, calculate weight according to field similarity, can adopt look-up table to realize herein, look-up table scope is [0, Max-Min], δ _rget integer between [15,25].

3. according to claim 2 based on real time equaliser infrared image detail enhancement algorithms, it is characterized in that: described look-up table realizes in the following ways: set current maximum as 1023, look-up table length is 1024,

First amplify 10000 times, retain the difference between data, again divided by 10000 during computing, index is the absolute value of current pixel and field pixel difference.

4. according to claim 1 based on real time equaliser infrared image detail enhancement algorithms, it is characterized in that: the linear contrast of described improvement strengthens in the following ways: after carrying out statistics with histogram to input picture, respectively from minimum and most high grade grey level, abandon several number of pixels, then grey scale mapping scope is obtained for [Hist_min, Hist_max], then by this grey scale mapping in [20,240] that are applicable to eye-observation.

5. according to claim 1 based on real time equaliser infrared image detail enhancement algorithms, it is characterized in that: the plateau equalization of described improvement in the following ways:

In formula, R is the tonal range of output image, can according to image dynamic conditioning.

rangeHist is input picture grey level range, and the maximal value of R is 255.