CN103177429A - FPGA (field programmable gate array)-based infrared image detail enhancing system and method - Google Patents

Abstract

The invention discloses an FPGA (field programmable gate array)-based infrared image detail enhancing system and method. The system comprises a bilateral filtering module, a gaussian filtering module, a histogram projecting module and an automatic gain control module, wherein the bilateral filtering module is connected with the gaussian filtering module which is connected with the histogram projecting module and the automatic gain control module respectively, and original input data firstly passes through the bilateral filtering module to obtain image pattern fundamental frequency information; the fundamental frequency information passes through the gaussian filtering module to be smoothened, and differencing is carried out between a result and the original input data so as to obtain image detail information; and the detail information is amplified by the automatic gain control module, the fundamental frequency information is compressed by the histogram projecting module, and the detail information and the fundamental frequency information are summed to obtain an output image. According to the invention, the contrast ratio of the image can be improved, the detail information can be enhanced, background noise can be restrained, and the common problems that the edge is fuzzy and a visual effect is poor in the image of a thermal infrared imager imaging system in the prior art can be solved.

Description

Infrared image details based on FPGA strengthens system and method thereof

Technical field

The invention belongs to the infrared thermal imaging technique field, particularly a kind of infrared image details based on FPGA strengthens system and method thereof.

Background technology

The modern high performance thermal infrared imager can the very large infrared original image of out-put dynamic range, and original detector data scope is generally in the 12-14 bit data, and this has obviously exceeded the dynamic range of display device.General typical display device, as monitor, it is merely able to receive 8 bit image signals.General human eye is merely able to differentiate 128 gray levels in addition.Therefore, after the original image that obtains high dynamic range, necessary process is exactly that original image with this high dynamic range remaps, with its dynamic range compression.This process need to reach two purposes usually: first: the dynamic range of output image can be complementary with the dynamic range that shows.Second: completing in first, keep as far as possible the details that exists in original image, make the observer can observe the image of better visual effect, and can distinguish as early as possible the weak target that is hidden in background.

contrast for infrared image strengthens problem, Chinese scholars has been done research widely, the algorithm of many uses is extensive visible (1.K.Zuiderveld in the literature also, " Contrast limited adaptive histogram equalizaiton, " in Graphics Gems IV, pp.474-485, Academic Press Professional, Inc., San Diego1994.2.S.M.Pizer, E.P.Amburm, J.D.Austin, R.Cromartie, A.Geselowitz, T.Greer, B.T.H.Romeny, and J.B.Zimmerman, " Adaptive histogram equalization and its variations, " Comput.Vis.Graph.Image Process.39 (3), 355-3681987.).Yet these contrast enhancement process major parts are for the low-dynamic range infrared image, the infrared image of 8 namely, and most of algorithm has only been considered the enhancing problem of rest image, does not consider the application in real-time system.Obviously, contrast improve strengthen this operation with details should be just more meaningful for original high dynamic range images comprised most complete information in scene because sample by AD in the original image signal that obtains, and had many faint detailed information.But, have to admit, contrast is strengthened to transfer on original image will be one and have more challenging work.Must adopt some complicated technology original signal could be mapped to the dynamic range that is fit to demonstration, and keep, even promote the observability of faint details and the overall contrast of image.

It is most popular image display technology in infrared imaging system that automatic gain is controlled balanced with square figure.Automatic gain is controlled the extremum at first reject in scene, then with the dynamic range linear mapping to 8 of integral body.The grey scale mapping function of histogram equalization image adopts the cumulative distribution function of original image, by the approximate satisfied evenly distribution of pixel distribution of image after histogram equalization.So histogram equalization is emphasized the gray level that the frequency of occurrences is larger more, so enhancing inevitably occurred through the image of histogram equalization, the homogeneous area noise amplifies, the problems such as bleaching effect.

consider that automatic gain is controlled and the deficiency of histogram equalizing method, many more complicated methods are suggested: as Retinex, and BF﹠amp, DRP method (3.WANG Yan-chen, LI Shu-jie, HUANG Lian-qing, " Enhancement of radiography based multiscale Retinex ", Optics and Precisio Engineering, Vol.14, No.1, 2005.4..F.Branchitta, M.Diani, G.Corsini, and A.Porta, " New technique for the visualization of high dynamic range infrared images, " Opt.Eng.48 (9), 0964012009.) etc., but these methods are mainly for visible images, namely has good effect for visible images, but to the infrared image poor effect, the noise problem of amplifying particularly.The universality of algorithm and real-time are not good in addition, are difficult to really be applied in real system.The effective most relative complex of detail enhancement algorithms, multiplex in computer vision or DSP embedded processing at present.Use FPGA to realize that complicated infrared image detail enhancement algorithms is quite blank, treatment effect also can't reach people's expectation.

Summary of the invention

The object of the present invention is to provide a kind of infrared image details based on FPGA to strengthen system and method thereof, improve the contrast of image, strengthen detailed information, the Background suppression noise, solved the general edge fog of image in present thermal infrared imager imaging system, contrast and detail areas subregion are low, the technical matters of poor visual effect.

The technical solution that realizes the object of the invention is: a kind of infrared image details based on FPGA strengthens system and method thereof, comprise bilateral filtering module, gaussian filtering module, projection histogram module and automatic gain control module, the bilateral filtering module is connected with the gaussian filtering module, the gaussian filtering module is connected with the automatic gain control module with the projection histogram module respectively, original input data first passes through the bilateral filtering module, obtains the image graphics fundamental frequency information; Fundamental frequency information passes through the gaussian filtering module again, with the fundamental frequency information smoothing processing, and result and original input data is done poor, obtains image detail information; Detailed information is amplified processing through the automatic gain control module, and simultaneously, fundamental frequency information compresses processing through the projection histogram module, with both output summations, namely obtains output image at last.

The present invention compared with prior art, its remarkable advantage: (1) the present invention can arrive the dynamic range that is fit to demonstration with the image with large dynamic range data compression of input, in this compression process, improve the contrast of image, strengthen detailed information, the Background suppression noise, solved the general edge fog of image in present thermal infrared imager imaging system, contrast and detail areas subregion are low, and the technical matters of poor visual effect has realized the scene objects separating capacity that existing details enhancements does not reach.(2) the invention provides a plurality of customized parameters, portable strong, and all goodish processing can be arranged for different scene informations, and can be widely used in infrared detection, chemical imaging, night vision and drive in the thermal infrared imager imaging system that auxiliary, security monitoring and target following etc. have relatively high expectations to infrared image quality.

Below in conjunction with accompanying drawing, the present invention is described in further detail.

Description of drawings

Fig. 1 is that the infrared image details that the present invention is based on FPGA strengthens system architecture diagram.

Fig. 2 is that the bilateral filtering module realizes pipelined architecture figure.

Fig. 3 is that the projection histogram module realizes schematic diagram.

Fig. 4 (a) processes by the common platform histogram effect image that obtains in the outdoor scene situation.

Fig. 4 (b) processes by histogram of the present invention the effect image that obtains in the outdoor scene situation.

Fig. 5 (a) processes by the common platform histogram effect image that obtains in large tracts of land, high temp objects scene situation.

Fig. 5 (b) processes by histogram of the present invention the effect image that obtains in large tracts of land, high temp objects scene situation.

Embodiment

in conjunction with Fig. 1, the infrared image details that the present invention is based on FPGA strengthens system and method thereof, comprise the bilateral filtering module, the gaussian filtering module, projection histogram module and automatic gain control module, the bilateral filtering module is connected with the gaussian filtering module, the gaussian filtering module is connected with the automatic gain control module with the projection histogram module respectively, the raw image data of input passes through the fundamental frequency information that bilateral filtering module and gaussian filtering module obtain image on the one hand, on the one hand the time delay of the above-mentioned processing of buffer memory and with fundamental frequency information do poor, obtain the levels of detail data of image.Image detail layer data and the fundamental frequency data sheet reason of staying alone: levels of detail is controlled to strengthen through adaptive gain and is processed; Fundamental frequency information is compressed to indication range through projection histogram.Both summations, and do suitable anti-spilled operation and namely obtain output image data.In above-mentioned processing, the most important thing is to design crucial sequential and data bit width in the complicated calculations process.To optimize the crucial sequential of FPGA inside in combined process as far as possible in design process, as set up the retention time etc., avoid occurring sequential warning and cause image abnormity.All modules are by a slice FPGA realization, and whole system is streamline, comprises 4 submodules, is respectively bilateral filtering module, gaussian filtering module, projection histogram module and automatic gain control module.The input data are first passed through the bilateral filtering module, obtain the image graphics fundamental frequency information; Pass through again the gaussian filtering module, with fundamental frequency information appropriateness smoothing processing, and result and original image are done poor, obtain image detail information; Detailed information is amplified processing through the automatic gain control module, and simultaneously, fundamental frequency information compresses processing through the projection histogram module, with both output summations, namely obtains output image at last.Modules provides a plurality of adjustable parameters, can adapt to different thermal imaging system parameter requests and different scene content.

Fig. 2 is the detail flowchart that the bilateral filtering module realizes.The bilateral filtering module receives 14 (or 16) raw image datas (being high dynamic range images) of input, in 7 * 7 processing window, 49 data wherein being carried out the associating filtering of spatial domain and intensity domain processes, the window center pixel value is finally substituted by the weighted mean of each pixel in window, and as the output image data of bilateral filtering module, after connect the gaussian filtering resume module.Want to take advantage of the weighting coefficient of trying to achieve will be as the adaptation coefficient in the automatic gain control module by spatial domain and intensity domain.Bilateral filtering module principle of work and particular hardware realize being described below:

In nonlinear filter, the most representative is bilateral filtering, and the high dynamic range images of input is carried out bilateral filtering, and to process formula as follows:

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

Wherein, k (i, j) represents normalization coefficient:

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

Mark (i ', j ') ∈ S _i,jRepresentative (i ', j ') and (i, j) are the adjacent elements in image, I _inThe input image pixels value that (i, j) denotation coordination (i, j) is located, g _sBe standardized gaussian kernel function of spatial domain

Be that all coefficient sums in filtering are 1; g _rFor adopting a standardized gaussian kernel function in intensity domain

This template S _i,jTotal weight, namely normalization coefficient k (i, j) multiplies each other by the result with two Gauss's templates of spatial domain and intensity domain to obtain.Its scope should be between 0-1.σ _sWith σ _rThe standard deviation parameter of two gaussian kernel functions of expression, it controls the expansion scope of two gaussian kernel functions.σ _sDetermined the yardstick of close region, thus must with the relation that is in proportion of image, choose the 2.5-5% of image diagonal size here.σ _rSelection more crucial because it has represented the amplitude of so-called details.If the scope of signal fluctuation is less than σ _r, this signal fluctuation will be considered to details so, namely can be level and smooth by two-sided filter, and be split in levels of detail.Otherwise, if the scope of this fluctuation is greater than σ _r, due to the nonlinear characteristic of two-sided filter, this edge will well be remained into the fundamental frequency layer so.Here select human eye can differentiate the 20-25% of gray level, namely 25 as σ _rValue.This value is considered human eye to the resolution characteristic of gray scale, and all has adaptability preferably for different scenes.

The hardware of bilateral filtering module realizes adopting the form of look-up table.As shown in Figure 2, use FIFO to make impact damper and generate 7 * 7 processing window, each data in window and window center data (being D25) are done poor taking absolute value, and move to left 7 (namely amplifying 128 times) are rear divided by Delt, Delt is the intensity domain factor sigma in formula _r, its value is 25.To make data after the business through the following formula look-up table with Delt, and find out corresponding gaussian coefficient, the set look-up table degree of depth is 512, and width is 10, by formula:

$\mathrm{Lut}\left(i\right)=512\×e^{-{\left(\frac{N_i}{128}\right)}^2}$

Obtain, wherein N _iBe the look-up table address, constant interval is 0 to 511.This coefficient is same amplifies 512 times.The spatial domain coefficient fixedly provides, σ _sValue is 25, is worth thus according to bilateral filtering spatial domain Gaussian function to calculate 49 spatial domain coefficient values in window, carries out normalized, and income value is amplified 512 times (moving to left 9).Be multiplied by respectively corresponding spatial domain coefficient with searching simultaneously 49 intensity domain coefficients that obtain, acquired results is as the weighting coefficient of each pixel in window.Multiply each other with each weighting coefficient and respective pixel value, cumulative, result obtains center pixel value after bilateral filtering is processed divided by the accumulated value of each weighting coefficient, and this value is the pixel data value that two-sided filter is finished dealing with, as an output of bilateral filtering module, after connect the gaussian filtering module.Simultaneously, the aggregate-value buffer memory of each weighting coefficient gets off, and as the control coefrficient of adaptive gain, is saving resource, and this coefficient is dwindled 256 times of storages again.

The filtering window that the gaussian filtering module is used is fixed form, and is as follows:

1	1	1
			1	2	1
1	1	1

This Gaussian filter uses FIFO to make impact damper and generates 3 * 3 windows, then the pixel in window is done weighted mean and processes.The center pixel weights are high, and the surrounding pixel weights are low.The view data of gaussian filtering output is the image fundamental frequency information, and the raw image data of these fundamental frequency data and buffer memory is done the poor image detail information that is.

In based on redundancy Gray Projection histogram module, comprise a control module and connect two ram in slice, the scope of output image is obtained by each valid gray level mean effort of input picture, no matter this gray level by how many pixels is occupied, only calculates once.And so-called available gray-scale just refers to that certain gray scale is occupied by abundant pixel.So being stored in the projection histogram H (x) of RAM1 is expressed as

$H\left(x\right)=\{\begin{array}{ccc}0& \mathrm{for}& n_x<T\\ 1& \mathrm{for}& n_x\&GreaterEqual;T\end{array}$

N wherein _xThe expression gray level is the total number of pixel of x, and threshold value T is set as 0.1% of whole image total pixel number.Thus, corresponding accumulation histogram is stored in RAM2, and distribution can be expressed as

$B\left(x\right)=\{\begin{array}{cc}0& \begin{array}{}\end{array}\mathrm{for\; x}=0\\ \underset{y=0}{\overset{x-1}{\mathrm{\&Sigma;}}}B\left(y\right)& \mathrm{other}\end{array}$

B (x) is the grey scale mapping function of image, and is last, also needs to limit the scope R of image equalization:

R＝min(n _valid,255)

n _ValidRepresent total number of effective gray level in whole image, in this way, the dynamic range of fundamental frequency image can effectively be compressed, and when the input picture dynamic range is low, has avoided traditional histogram equalization algorithm to the undue stretching of dynamic range of images.Be compared to other histograms and process, what store when carrying out statistics with histogram due to the method is histogram information after binaryzation, is compared to original histogram information, and data volume is a lot of less, can save the ram in slice space of considerable part.Fig. 4 is the description of this module.At the picture frame signal when being high, RAM1 is with the histogram information of pixel clock frequency statistics current frame image, what RAM2 stored is the complete histogram of previous frame image projection equilibrium treatment, the memory resource pressure that brings for reducing buffer memory one frame image data, current frame image adopts the histogram output of previous frame image when doing the histogram processing.Surpass in the video flowing of 25 frame/seconds at frame frequency, processing can not reduce the continuity that video shows like this.At the picture frame signal when low, the first binaryzation of the histogram information in RAM1, namely when the data of certain gray level during higher than setting threshold, this gray scale is designated as 1, otherwise is 0.The histogram of binaryzation adds up, and result store is used for next frame output in RAM2, and RAM1 zero clearing simultaneously is so that next frame begins the histogram information of adding up new.

The adaptation coefficient that automatic gain control module buffer memory after the bilateral filtering resume module is come as the gain controlling elements of levels of detail, carries out adaptive enhancing to levels of detail.Then the first normalized of this coefficient with fixing gain and a bias factor, obtains this coefficient adjustment the enhancer of respective pixel to OK range (as 1 to 2.5).Than more rich zone, corresponding enhancer is larger for details; For the flat site (as the sky background zone) of image, corresponding enhancer is less.

Suppose that the gain span of detail pictures is at G _minTo G _max, we obtain the gain expressions of detail pictures by the mode of linear mapping:

G(i,j)＝G _min+(1-f(i,j))(G _max-G _min)

.

＝G _min+(1-k(i,j))(G _max-G _min)

On the whole, in order not lose the details of original image, and avoid the amplification of noise, G _minCan be made as 1.G _maxCan select on demand, generally, G _max=2.5-3 can obtain human eye vision effect preferably.In concrete processing, normalized adaptation coefficient is amplified 512*512/256 doubly (16, wherein produce 18 amplifications by two two-sided filter coefficients of 9 of moving to left, then give up low 2 preservations of coefficient, obtain the amplification of 16).The formula of processing is as follows:

G＝1024*G(i,j)＝1024*G _min+(1024-1024*k(i,j))(G _max-G _min)

Wherein, 1024*k (i, j) is the adaptive gain control coefrficient that obtains in hardware handles, gets G _min=1, G _maxCan adjust in real time.Multiply each other with corresponding detail data with this adaptive gain value G, acquired results dwindles 1024 times of image details after being gain.

The image detail information that the fundamental frequency information that projection histogram is processed and automatic gain are controlled after strengthening has the symbol summation operation, realize the fundamental frequency information of layering processing and the fusion of detailed information, if summed result pixel value occurs for negative, illustrate to have and overflow, it is restricted to the value that is fit to demonstration.Output rusults is the final image data processing of this system.

Fig. 4 (a) processes by the common platform histogram effect image that obtains in the outdoor scene situation.Fig. 4 (b) processes by histogram of the present invention the effect image that obtains in the outdoor scene situation.Fig. 5 (a) processes by the common platform histogram effect image that obtains in large tracts of land, high temp objects scene situation.Fig. 5 (b) processes by histogram of the present invention the effect image that obtains in large tracts of land, high temp objects scene situation.Can find out that from Fig. 4 (a), Fig. 4 (b), Fig. 5 (a) with 5(b) this system can effectively adjust the integral image contrast, than histogram equalization, can not cause luminance saturation, and image detail obtained obvious enhancing, visual effect is significantly improved.

Claims (10)

1. the infrared image details based on FPGA strengthens system, it is characterized in that comprising bilateral filtering module, gaussian filtering module, projection histogram module and automatic gain control module, the bilateral filtering module is connected with the gaussian filtering module, the gaussian filtering module is connected with the automatic gain control module with the projection histogram module respectively, original input data first passes through the bilateral filtering module, obtains the image graphics fundamental frequency information; Fundamental frequency information passes through the gaussian filtering module again, with the fundamental frequency information smoothing processing, and result and original input data is done poor, obtains image detail information; Detailed information is amplified processing through the automatic gain control module, and simultaneously, fundamental frequency information compresses processing through the projection histogram module, with both output summations, namely obtains output image at last.

2. the infrared image details based on FPGA according to claim 1 strengthens system, it is characterized in that in the bilateral filtering module, the input raw image data, in 7 * 7 processing window, window data is wherein carried out the bilateral filtering processing of spatial domain and intensity domain, the window center pixel value is finally substituted by the weighted mean of each pixel in window, and as the output image data of bilateral filtering module, after connect the gaussian filtering resume module; Wanted to take advantage of the weighting coefficient of trying to achieve as the adaptation coefficient in the automatic gain control module by spatial domain and intensity domain, specific as follows: as in the bilateral filtering module, at first the high dynamic range images of input to be carried out the bilateral filtering processing:

$I_{\mathrm{bf}}(i,j)=\frac{1}{k(i,j)}\underset{(i^{\&prime;},j^{\&prime;})\&Element;S_{i,j}}{\mathrm{\&Sigma;}}{g}_s(i-i^{\&prime;},j-j^{\&prime;})g_r(I_{\mathrm{in}}(i,j)-I_{\mathrm{in}}(i^{\&prime;},j^{\&prime;}))I_{\mathrm{in}}(i^{\&prime;},j^{\&prime;})$

Wherein, k (i, j) represents normalization coefficient:

$k(i,j)=\underset{(i^{\&prime;},j^{\&prime;})\&Element;S_{i,j}}{\mathrm{\&Sigma;}}{g}_s(i-i^{\&prime;},j-j^{\&prime;})g_r(I_{\mathrm{in}}(i,j)-I_{\mathrm{in}}(i^{\&prime;},j^{\&prime;}))$

Mark (i ', j ') ∈ S _i,jRepresentative (i ', j ') and (i, j) are the adjacent elements in image, I _inThe input image pixels value that (i, j) denotation coordination (i, j) is located, g _sBe standardized gaussian kernel function of spatial domain Be that all coefficient sums in filtering are 1; g _rFor adopting a standardized gaussian kernel function in intensity domain

This template S _i,jTotal weight, namely normalization coefficient k (i, j) multiplies each other by the result with two gaussian kernel functions of spatial domain and intensity domain to obtain, its scope is between 0-1; σ _sWith σ _rThe standard deviation parameter of two gaussian kernel functions of expression, it controls the expansion scope of two gaussian kernel functions; σ _sDetermined the yardstick of close region, thus must with the relation that is in proportion of image, choose the 2.5-5% of image diagonal size, σ here _rSelect human eye to differentiate the 20-25% of gray level;

Then the hardware of bilateral filtering module realizes adopting the form of look-up table, namely use FIFO to make impact damper and generate 7 * 7 processing window, each data in window and window center data are done poor taking absolute value, and move to left 7, after namely amplifying 128 times, divided by Delt, Delt is the intensity domain factor sigma in formula _r, its value is 25; To make data after the business through the following formula look-up table with Delt, and find out corresponding gaussian coefficient, the set look-up table degree of depth is 512, and width is 10, by formula:

$\mathrm{Lut}\left(i\right)=512\&times;e^{-{\left(\frac{N_i}{128}\right)}^2}$

Obtain, wherein N _iBe the look-up table address, constant interval is 0 to 511, and this coefficient is same amplifies 512 times; The spatial domain coefficient fixedly provides, σ _sValue is 25, and the standardization Gaussian function that is worth thus according to the bilateral filtering spatial domain calculates 49 spatial domain coefficient values in window, carries out normalized, and income value is amplified 512 times, namely moves to left 9; Be multiplied by respectively corresponding spatial domain coefficient with searching simultaneously 49 intensity domain coefficients that obtain, acquired results is as the weighting coefficient of each pixel in window; Multiply each other with each weighting coefficient and respective pixel value, cumulative, result obtains center pixel value after bilateral filtering is processed divided by the accumulated value of each weighting coefficient, this value is the pixel data value that the bilateral filtering resume module is completed, as an output of bilateral filtering module, after connect the gaussian filtering module; Simultaneously, the aggregate-value buffer memory of each weighting coefficient gets off, as the coefficient of adaptive gain control module.

3. the infrared image details based on FPGA according to claim 1 strengthens system, it is characterized in that the gaussian filtering module uses filtering window to be fixed form, and is as follows:

1 1 1 1 2 1 1 1 1

This gaussian filtering module is used FIFO to make impact damper and is generated 3 * 3 windows, then the pixel in window is done weighted mean and processes; The center pixel weights are high, and the surrounding pixel weights are low; The view data of gaussian filtering output is the image fundamental frequency information, and the raw image data of these fundamental frequency data and buffer memory is done the poor image detail information that is.

4. the infrared image details based on FPGA according to claim 1 strengthens system, it is characterized in that the projection histogram module comprises a control module and connects two ram in slice, at the picture frame signal when being high, RAM1 is with the histogram information of pixel clock frequency statistics current frame image, what RAM2 stored is the complete histogram of previous frame image projection equilibrium treatment, the memory resource pressure that brings for reducing buffer memory one frame image data, current frame image adopts the histogram output of previous frame image when doing the histogram processing; At the picture frame signal when low, the first binaryzation of the histogram information in RAM1, namely when the data of certain gray level during higher than setting threshold, this gray scale is designated as 1, otherwise is 0; The histogram of binaryzation adds up, and result store is used for next frame output in RAM2, and RAM1 zero clearing simultaneously is so that next frame begins the histogram information of adding up new.

5. the infrared image details based on FPGA according to claim 4 strengthens system, it is characterized in that the projection histogram H (x) that is stored in RAM1 is expressed as

$H\left(x\right)=\{\begin{array}{ccc}0& \mathrm{for}& n_x<T\\ 1& \mathrm{for}& n_x\&GreaterEqual;T\end{array}$

N wherein _xThe expression gray level is the total number of pixel of x, and threshold value T is set as 0.1% of whole image total pixel number, and corresponding accumulation histogram is stored in RAM2, and distribution table is shown

$B\left(x\right)=\{\begin{array}{cc}0& \begin{array}{c}\begin{array}{cc}\mathrm{for}& x\end{array}\end{array}=0\\ \underset{y=0}{\overset{x-1}{\mathrm{\&Sigma;}}}B\left(y\right)& \mathrm{other}\end{array}$

B (x) is the grey scale mapping function of image, and is last, also needs to limit the scope R of image equalization:

R＝min(n _valid,255)

n _ValidThe total number that represents effective gray level in whole image.

6. the infrared image details based on FPGA according to claim 4 strengthens system, it is characterized in that the adaptation coefficient that automatic gain control module buffer memory after the bilateral filtering resume module is come, gain controlling elements as levels of detail, levels of detail is carried out adaptive enhancing, the i.e. first normalized of this coefficient, then with fixing gain and a bias factor, this coefficient adjustment is obtained the enhancer of respective pixel to the Human Perception scope;

The gain span of detail pictures is at G _minTo G _max, obtain the gain expressions of detail pictures by the mode of linear mapping:

G(i,j)＝G _min+(1-f(i,j))(G _max-G _min)

.

＝G _min+(1-k(i,j))(G _max-G _min)

G _minBe made as 1, G _max=2.5-3 obtains human eye vision effect preferably.

7. the infrared image details based on FPGA according to claim 6 strengthens system, it is characterized in that in adaptive enhancing is processed, normalized adaptation coefficient is amplified 512*512/256 doubly, 16, wherein produce 18 amplifications by two two-sided filter coefficients of 9 of moving to left, then give up low 2 preservations of coefficient, obtain the amplification of 16, the formula of processing is as follows:

G＝1024*G(i,j)＝1024*G _min+(1024-1024*k(i,j))(G _max-G _min)

Wherein, 1024*k (i, j) is the adaptive gain control coefrficient that obtains in processing, gets G _min=1, G _maxAdjust in real time, multiply each other with detail data accordingly with this adaptive gain value G, acquired results dwindles 1024 times of image details after being gain.

8. the infrared image detail enhancing method based on FPGA, is characterized in that original input data first passes through bilateral filtering, obtains the image graphics fundamental frequency information; Fundamental frequency information passes through gaussian filtering again, with the fundamental frequency information smoothing processing, and result and original input data is done poor, obtains image detail information; Detailed information is controlled to amplify through automatic gain and is processed, and simultaneously, fundamental frequency information compresses processing through projection histogram, with both output summations, namely obtains output image at last;

In bilateral filtering, the input raw image data, in 7 * 7 processing window, window data is wherein carried out the bilateral filtering processing of spatial domain and intensity domain, the window center pixel value is finally substituted by the weighted mean of each pixel in window, and as the output image data of bilateral filtering, wanted to take advantage of the adaptation coefficient of the weighting coefficient of trying to achieve in controlling as automatic gain by spatial domain and intensity domain, specific as follows:

At first the high dynamic range images of input carried out the bilateral filtering processing:

$I_{\mathrm{bf}}(i,j)=\frac{1}{k(i,j)}\underset{(i^{\&prime;},j^{\&prime;})\&Element;S_{i,j}}{\mathrm{\&Sigma;}}{g}_s(i-i^{\&prime;},j-j^{\&prime;})g_r(I_{\mathrm{in}}(i,j)-I_{\mathrm{in}}(i^{\&prime;},j^{\&prime;}))I_{\mathrm{in}}(i^{\&prime;},j^{\&prime;})$

Wherein, k (i, j) represents normalization coefficient:

$k(i,j)=\underset{(i^{\&prime;},j^{\&prime;})\&Element;S_{i,j}}{\mathrm{\&Sigma;}}{g}_s(i-i^{\&prime;},j-j^{\&prime;})g_r(I_{\mathrm{in}}(i,j)-I_{\mathrm{in}}(i^{\&prime;},j^{\&prime;}))$

Mark (i ', j ') ∈ S _i,jRepresentative (i ', j ') and (i, j) are the adjacent elements in image, I _inThe input image pixels value that (i, j) denotation coordination (i, j) is located, g _sBe standardized gaussian kernel function of spatial domain Be that all coefficient sums in filtering are 1; g _rFor adopting a standardized gaussian kernel function in intensity domain

This template S _i,jTotal weight, namely normalization coefficient k (i, j) multiplies each other by the result with two gaussian kernel functions of spatial domain and intensity domain to obtain, its scope is between 0-1; σ _sWith σ _rThe standard deviation parameter of two gaussian kernel functions of expression, it controls the expansion scope of two gaussian kernel functions; σ _sDetermined the yardstick of close region, thus must with the relation that is in proportion of image, choose the 2.5-5% of image diagonal size, σ here _rSelect human eye to differentiate the 20-25% of gray level;

The form of look-up table is adopted in the realization of bilateral filtering, namely use FIFO to make impact damper and generate 7 * 7 processing window, each data in window and window center data are done poor taking absolute value, and move to left 7, after namely amplifying 128 times, divided by Delt, Delt is the intensity domain factor sigma in formula _r, its value is 25; To make data after the business through the following formula look-up table with Delt, and find out corresponding gaussian coefficient, the set look-up table degree of depth is 512, and width is 10, by formula:

$\mathrm{Lut}\left(i\right)=512\&times;e^{-{\left(\frac{N_i}{128}\right)}^2}$

Obtain, wherein N _iBe the look-up table address, constant interval is 0 to 511, and this coefficient is same amplifies 512 times; The spatial domain coefficient fixedly provides, σ _sValue is 25, and the standardization Gaussian function that is worth thus according to the bilateral filtering spatial domain calculates 49 spatial domain coefficient values in window, carries out normalized, and income value is amplified 512 times, namely moves to left 9; Be multiplied by respectively corresponding spatial domain coefficient with searching simultaneously 49 intensity domain coefficients that obtain, acquired results is as the weighting coefficient of each pixel in window; Multiply each other with each weighting coefficient and respective pixel value, cumulative, result obtains center pixel value after bilateral filtering is processed divided by the accumulated value of each weighting coefficient, and this value is the pixel data value that bilateral filtering is finished dealing with, as an output of bilateral filtering; Simultaneously, the aggregate-value buffer memory of each weighting coefficient gets off, as the coefficient of adaptive gain control.

9. the infrared image detail enhancing method based on FPGA according to claim 8, it is characterized in that projection histogram realizes in two ram in slice that control module connects, at the picture frame signal when being high, RAM1 is with the histogram information of pixel clock frequency statistics current frame image, what RAM2 stored is the complete histogram of previous frame image projection equilibrium treatment, the memory resource pressure that brings for reducing buffer memory one frame image data, current frame image adopts the histogram output of previous frame image when doing the histogram processing; At the picture frame signal when low, the first binaryzation of the histogram information in RAM1, namely when the data of certain gray level during higher than setting threshold, this gray scale is designated as 1, otherwise is 0; The histogram of binaryzation adds up, and result store is used for next frame output in RAM2, and RAM1 zero clearing simultaneously is so that next frame begins the histogram information of adding up new;

The projection histogram H (x) of the described RAM1 of being stored in is expressed as

$H\left(x\right)=\{\begin{array}{ccc}0& \mathrm{for}& n_x<T\\ 1& \mathrm{for}& n_x\&GreaterEqual;T\end{array}$

N wherein _xThe expression gray level is the total number of pixel of x, and threshold value T is set as 0.1% of whole image total pixel number, and corresponding accumulation histogram is stored in RAM2, and distribution table is shown

$B\left(x\right)=\{\begin{array}{cc}0& \begin{array}{c}\begin{array}{cc}\mathrm{for}& x\end{array}\end{array}=0\\ \underset{y=0}{\overset{x-1}{\mathrm{\&Sigma;}}}B\left(y\right)& \mathrm{other}\end{array}$

B (x) is the grey scale mapping function of image, and is last, also needs to limit the scope R of image equalization:

R＝min(n _valid,255)

n _ValidThe total number that represents effective gray level in whole image.

10. the infrared image detail enhancing method based on FPGA according to claim 8, it is characterized in that the adaptation coefficient that automatic gain control is come from the rear buffer memory of bilateral filtering processing, gain controlling elements as levels of detail, levels of detail is carried out adaptive enhancing, the i.e. first normalized of this coefficient, then with fixing gain and a bias factor, this coefficient adjustment is obtained the enhancer of respective pixel to the Human Perception scope;

The gain span of detail pictures is at G _minTo G _max, obtain the gain expressions of detail pictures by the mode of linear mapping:

G(i,j)＝G _min+(1-f(i,j))(G _max-G _min)

.

＝G _min+(1-k(i,j))(G _max-G _min)

G _minBe made as 1, G _max=2.5-3 obtains human eye vision effect preferably.

Images