CN104318529A - Method for processing low-illumination images shot in severe environment - Google Patents

Abstract

The invention discloses a method for processing low-illumination images shot in the severe environment. The method comprises the first step of processing a plurality of frames of collected continuous low-illumination images through an existing 3D noise reduction algorithm; the second step of processing a plurality of frames of continuous low-illumination images having undergone the 3D noise reduction algorithm processing through an existing non-local mean (NLM) filtering algorithm; the third step of adopting a high-dynamic range image enhancement algorithm to perform the operations of calculating the brightness value of each pixel point (x, y) in the continuous low-illumination images having undergone the NLM filtering algorithm processing through a high-dynamic range image enhancement algorithm formula (1), and outputting images through the brightness value of each pixel point (x, y) in the continuous low-illumination images. The method can obviously further improve image quality of low-illumination videos and visual effects of the low-illumination videos.

Description

The method of process captured low-light (level) image in rugged surroundings

Technical field

The present invention relates to the software execution technique program of process low-light (level) video image, particularly process the method for captured low-light (level) image in rugged surroundings.

Background technology

Usually, acquisition image or video carry out under desirable illumination condition, but actual service condition often has to obtain image or video really under low-light (level) environment.The image quality deviation of these images obtained under low-light (level) environment or video, often can not resolution image details, affects visual effect and affects follow-up further treatment effect.The image obtained under low-light (level) environment and the principal feature of video are: 1, brightness of image entirety is partially dark; 2, compared with the image obtained under desirable illumination condition, containing much noise; 3, due to the restriction of illumination condition and the impact of light source itself, in RGB, some component is on the low side relative to other components.

In current world wide, the defect of existing like product technology and existence is enumerated and is analyzed: at present, in order to clearer, image more clearly can be collected under very dark light, major part realizes by the following method: 1. when taking, add infrared light compensating lamp, this method sixty-four dollar question obviously can increase cost; 2. increase aperture when taking, this method can make camera lens size become large, significantly can increase systemic cost after aperture is greater than F1.4; 3. increasing the time shutter when taking, having obvious smear when the remarkable shortcoming of this method is subject motion, causing the impact of sharpness; 4. strengthen the optical signal intensity come from sensor, improve gain.The gain shortcoming increasing sensor is that noise also can be amplified together, so image has a lot of noise, it is not so clear that original image just becomes.In addition, noise becomes greatly, and coding can become very difficult, because be hard to tell details or the noise of image, so its can be encoded some noises as image.The emphasis of this method distinguishes noise and real image, wants the high definition that can ensure image while removing noise.Due to above-mentioned the 1. to the 3. shown in the difficulty of existing methodical open defect and technological improvement, 4. present most of method all plants method mainly for above-mentioned, namely sensor gain is improved, therefore, 4. kind method can be subdivided into two links: image enhaucament or image noise reduction.

The main method of current image enhaucament link is: a) histogram equalization-histogram equalization is one of the most frequently used, most important algorithm in image enhaucament spatial domain method, it is based on probability theory, basic thought is by the distribution of equalization processing adjustment gradation of image, reaches the object improving picture contrast.Because picture contrast is the key factor determining piece image subjective quality, therefore, histogram equalization is widely used in the enhancing process of image.After histogram equalization process, the contrast of picture is obviously strengthened, and image is also more clear, and from histogram, the image histogram distribution after process is more even.But histogram equalization enhancing also exists 2 weak points: one, and the image gray levels after process reduces to some extent, causes some details to disappear; They are two years old, some image, as histogram has peak etc., its contrast easily produces factitious undue enhancing problem after treatment, such as, some satellite image or medical image because of intensity profile concentrations, when carrying out histogram equalization process to this type of image, often there is bright or excessively dark phenomenon in its result, does not reach the object strengthening visual effect.In addition, for the limited gray level of image, quantization error also often causes information dropout, causes the edge of some sensitivities to disappear because of the merging with neighbor pixel, and this is that histogram modification strengthens unavoidable problem.B) transpositions domain-transpositions domain is, by certain conversion, image is transformed into another spatial domain, then to coefficient carry out certain process, finally to coefficient carry out inverse transformation be enhanced after image.Common transform domain method has high-pass filtering, wavelet transformation etc.First high-pass filtering method carries out Fourier transform to image, then by a Hi-pass filter, strengthen high fdrequency component (namely strengthening the details of image), suppress low frequency component simultaneously, finally again Fourier inversion and image after being enhanced are carried out to image.C) Retinex algorithm for image enhancement-Retinex (abbreviation of retina " Retina " and cerebral cortex " Cortex ") theory is that a kind of image enhaucament based on human visual system (Human Visual System) be based upon on scientific experiment and scientific analysis basis is theoretical.The basic principle model of this algorithm is a kind of theory being called as color, and a kind of image enchancing method based on theory proposed on the basis of color constancy.The substance of Retinex theory is the color of object is determined the reflection potential of long wave (red), medium wave (green) and shortwave (indigo plant) light by object, instead of to be determined by the absolute value of intensity of reflected light, the color of object does not affect by illumination is heteropic, have consistance, namely Retinex theory is based on color constancy (color constancy).This theory is thought, the brightness of object is determined jointly by incident components (i.e. luminance component) and reflecting component, i.e. expression formula I (x, y)=L (x, y) R (x, y), I (x, y) represents brightness of image, L (x, y) represent luminance component, R (x, y) represents reflecting component.By taking the logarithm, L (x can be made, y) addition is become by being multiplied, i.e. expression formula logI (x with the relation of R (x, y), y)=logL (x, y)+logR (x, y), luminance component is relevant with light source, and reflecting component is relevant with the color characteristics of object itself, namely reflecting component reflects the true colors of object.The thought of Retinex theory is the impact by removing light illumination from image, namely unitizes to illumination, obtains reflecting component, is the image of reflection object true colors.Retinex algorithm mainly contains three kinds, SSR (Single-Scale Retinex), MSR (Multi-Scale Retinex) and MSRCR (Multi-Scale Retinex with Color Restoration).The expression formula of SSR is: logR _i(x, y)=logI _i(x, y)-log (F (x, y) * I _i(x, y)), { R, G, B}, represent RGB tri-passages, * represents convolution to i ∈, and F (x, y) is low-pass filter, usually selects normalized Gaussian filter.The essence of SSR is the illumination estimation being obtained image by the weighted mean of a neighborhood, then at log-domain by subtracting each other removal, to obtain reflecting component.The shortcoming of SSR can produce halo effect in image border.For addressing this problem, MSR is then suggested, and namely the expression formula of MSR is:

$\log R_i(x,y)=\underset{n=1}{\overset{N}{\mathrm{\Σ}}}{w}_n(\log I_i(x,y)-\log (F_n(x,y)*I_i(x,y)))$

In the expression formula of MSR, N generally gets 3, w _nrepresent weighting coefficient, remove 1/3, F _n(x, y) represents the Gaussian filter that variance is different.MSR is the average of SSR in essence, it uses 3 different Gaussian filters, respectively corresponding little, in, large 3 yardsticks, average is got to the output of these 3 Gaussian filters and is MSR.SSR and MSR processes RGB tri-passages respectively, and its common shortcoming to produce cross-color.Therefore, on MSR basis, MSRCR is suggested again, and MSRCR method adds Dynamic gene on the basis of MSR, and the expression formula of MSRCR is:

$\log R_i(x,y)={\mathrm{\α}}_i(x,y)\underset{n=1}{\overset{N}{\mathrm{\Σ}}}{w}_n(\log I_i(x,y)-\log (F_n(x,y)*I_i(x,y)))$

In the expression formula of MSRCR, { R, G, B} represent RGB tri-passages, α to i ∈ _i(x, y) is the Dynamic gene based on RGB tri-components.MSRCR can solve the colour cast problem of SSR and MSR to a certain extent, but its shortcoming can not fundamentally solve colour cast problem and calculated amount is larger.

The most crucial defect of Retinex method strengthens rear result exactly and there will be colour cast phenomenon, can the visual effect of effect diagram picture and the resolution to image detail.Main method for image noise reduction is: 1. traditional airspace filter Image Denoising-traditional airspace filter Image Denoising mainly includes Wiener filtering, medium filtering, mean filter etc.For impulsive noise single in low-light (level) image or Gaussian noise, tradition airspace filter Image Denoising can play effective Noise Reduction, but when the noise in figure is the mixed form of impulsive noise and Gaussian noise, single medium filtering or mean filter are difficult to obtain gratifying noise reduction.2. mean filter is averaged the pixel in the window centered by certain pixel simply, replaces central pixel point; Mean filter is mainly used to suppress the Gaussian noise in image, and algorithm is simple, but can cause the fuzzy of whole image especially border, brings difficulty to subsequent treatment.3. medium filtering is a kind of typical nonlinear filter, and it carries out filtering by getting intermediate value to the pixel in certain window; Medium filtering is mainly used to impulse noise mitigation, when filter window is less, image detail can well be protected, but when the pixel number being subject to impulsive noise pollution in window exceedes a half of number of pixels in whole filter window, medium filtering is by complete failure; Though now increasing filter window can filtering impulsive noise, image can thicken again and significantly will increase operand.

Traditional airspace filter Image Denoising often all can only suppress for single noise, for the noise that kind in low-light (level) image is more and more complicated, the noise reduction of traditional airspace filter noise reduction technology is very limited, and its main method is: 1. frequency domain filtering Image Denoising-in frequency field, the noise reduction of image mainly adopts low-pass filtering and homomorphic filtering to realize.Low-pass filtering: the frequency spectrum that Fourier conversion obtains image is carried out to image, DC component wherein represents the average gray of image, large-area background area and slow region of variation show as low frequency component in a frequency domain, and the edge of image, details, jump part and noise all occur with high fdrequency component.Therefore, use low-pass filtering can stress release treatment to the frequency spectrum of image in a frequency domain, smoothed image, but meanwhile, also may the edge of some image of filtering and the frequency component corresponding to detail signal and image boundary is thickened.2. homomorphic filtering adjusts the tonal range of image, to improve the quality of image.Homomorphic filtering method carries out filtering to image in log-domain, while compressed image overall intensity scope, expand the interested tonal range of user.3. the algorithm of wavelet image noise reduction-wavelet transformation in image noise reduction is divided three classes substantially: linear filter Method of Noise, wavelet threshold shrink Method of Noise and wavelet coefficient model method.Linear filter Method of Noise algorithm is simple, is easy to realize.Directly the various wave filters of spatial domain are acted on wavelet field.But the image visual effect of this class methods process is general, and former figure has larger gap.It is study method the most widely at present that wavelet threshold shrinks Method of Noise, and wavelet thresholding method is divided into again hard threshold method and Adaptive Thresholding.Threshold deniosing is mainly based on the following fact, namely larger wavelet coefficient is all generally based on actual signal, smaller coefficient is then noise to a great extent, therefore can by selecting suitable wavelet threshold, the coefficient being less than threshold value is set to zero, and retain the wavelet coefficient being greater than threshold value, inverse transformation is carried out to the wavelet coefficient processed, realizes noise reduction and the reproduction of image.Hard threshold method calculated amount is less, can well retain the local features such as image border, but image there will be the phenomenons such as vision distortion.Adaptive Thresholding processing result image relative smooth is many, but calculated amount is comparatively large, and can cause the distortion phenomenons such as edge fog.Wavelet coefficient model method mainly carries out noise reduction based on following characteristics: often have very strong correlativity between the useful signal of the image wavelet coefficient on each layer relevant position, and the wavelet coefficient of noise then has weak relevant or incoherent feature.The noise reduction of the method is generally slightly good than threshold deniosing method, but shortcoming to be calculated amount comparatively large, can not scan picture be used for.4. based on the image denoising-need to carry out iterative operation based on the image de-noising method of rarefaction representation of rarefaction representation, calculated amount is comparatively large, is difficult to process in real time, seldom uses in process in real time.

Summary of the invention

The object of the present invention is to provide the method for a kind of process captured low-light (level) image in rugged surroundings, obviously can improve image quality and the visual effect thereof of low-light (level) video further.

The object of the present invention is achieved like this: the method for a kind of process captured low-light (level) image in rugged surroundings, the first step: the continuous some frame low-light (level) images gathered by the process of existing 3D noise reduction algorithm, to in chronological sequence order continuous some frame low-light (level) images gathered successively carry out contrast Screening Treatment, the noise of this continuous some frame low-light (level) image non-overlapping copies is leached automatically, makes total most similar pixel coupling wherein; Second step: by existing non-local mean (NLM) the filtering algorithm process handled continuous some frame low-light (level) images crossed of 3D noise reduction algorithm, in continuous some two field pictures gathered successively in chronological sequence order, all pixels all carry out non-mean quantization calculating; 3rd step: adopt high dynamic range images to strengthen algorithm: strengthen 1. algorithmic formula calculates wherein each pixel (x, y) brightness value to this continuous some frame low-light (level) image of non-local mean (NLM) filtering algorithm process by high dynamic range images and with the brightness value of each pixel (x, y) in this continuous some frame low-light (level) image output image, its formula is:

① $L_{\mathrm{out}}^{\tanh }(x,y)={\left\{\frac{N(x,y)L_{\mathrm{in}}(x,y)}{L_{\mathrm{in}}^{\mathrm{avg}}(x,y)+\mathrm{\ϵ}}\tanh \left(\frac{L_{\mathrm{in}}^{\mathrm{avg}}(x,y)+\mathrm{\ϵ}}{m(x,y)}\right)\right\}}_0^1$

High dynamic range images strengthen algorithmic formula 1. in, normalized factor, for the maximum brightness value in local, pixel (x, a y) corresponding place rectangular window in input picture, m (x, y) as linear function with control hyperbolic tangent function curvature and by input picture local rectangular window statistical property and calculating, for the average brightness value in local, pixel (x, a y) corresponding place rectangular window in input picture, m _maxmaximum quantization value then for calculating based on input picture local rectangular window statistical property, m _minbe then the minimum quantization value calculated based on input picture local rectangular window statistical property, s as scale factor to control the curvature of hyperbolic tangent function, L _in(x, y) is the intrinsic brilliance value of a pixel (x, y) in input picture, and ε corresponds to the average brightness value in the rectangular window of local, the corresponding place of pixel (x, y) in input picture residual correction value.

The present invention creatively property ground proposes colored low-light (level) and strengthens algorithm frame, 3D noise reduction algorithm, non-local mean noise reduction algorithm and high dynamic range images is strengthened algorithm and carries out organically combining solving the problem that low-light (level) image is difficult to strengthen.On image processing effect, be not only better than domestic current product, and the effect of process in real time can be reached in Video processing.

The present invention is according to the correlativity between the feature of the image obtained under actual low-light (level) environment and video and frame of video, propose a brand-new algorithm frame strengthened for low-light (level), this framework includes 3D noise reduction, non-local mean noise reduction and wide dynamic range image enhaucament etc.The present invention is directed to the feature that low-light (level) image contains much noise, carry out noise reduction process, in order to avoid amplify existing noise further in the middle of subsequent treatment.The present invention is in noise reduction process, the inter-frame information in video is utilized in its first step, perform the process of preliminary 3D noise reduction algorithm, in its second step, non-local mean noise reduction is carried out to each two field picture after the process of 3D noise reduction algorithm, to reduce noise further, just subsequent treatment, for the image after noise reduction in it is the 3rd, again for the characteristic of its low-light (level) (i.e. low-dynamic range), the ultimate handling procedure of core of the present invention-high dynamic range images is performed to each two field picture and strengthens algorithm, what can reach final optimization pass is enhanced low-light (level) (video) image.

The present invention passes through through practice, and it obviously can improve image quality and the visual effect thereof of low-light (level) video further.

Accompanying drawing explanation

Below in conjunction with accompanying drawing, the invention will be further described;

Fig. 1 is the example schematic of existing non-local mean (NLM) filtering algorithm of the present invention.

Embodiment

The method of a kind of process captured low-light (level) image in rugged surroundings, the first step: the continuous some frame low-light (level) images gathered by the process of existing 3D noise reduction algorithm, to in chronological sequence order continuous some frame low-light (level) images gathered successively carry out contrast Screening Treatment, the noise of this continuous some frame low-light (level) image non-overlapping copies is leached automatically, makes total most similar pixel coupling wherein; Second step: by existing non-local mean (NLM) the filtering algorithm process handled continuous some frame low-light (level) images crossed of 3D noise reduction algorithm, in continuous some two field pictures gathered successively in chronological sequence order, all pixels all carry out non-mean quantization calculating; 3rd step: adopt high dynamic range images to strengthen algorithm: strengthen 1. algorithmic formula calculates wherein each pixel (x, y) brightness value to this continuous some frame low-light (level) image of non-local mean (NLM) filtering algorithm process by high dynamic range images and with the brightness value of each pixel (x, y) in this continuous some frame low-light (level) image output image, its formula is:

① $L_{\mathrm{out}}^{\tanh }(x,y)={\left\{\frac{N(x,y)L_{\mathrm{in}}(x,y)}{L_{\mathrm{in}}^{\mathrm{avg}}(x,y)+\mathrm{\ϵ}}\tanh \left(\frac{L_{\mathrm{in}}^{\mathrm{avg}}(x,y)+\mathrm{\ϵ}}{m(x,y)}\right)\right\}}_0^1$

High dynamic range images strengthen algorithmic formula 1. in, normalized factor, for the maximum brightness value in local, pixel (x, a y) corresponding place rectangular window in input picture, m (x, y) as linear function with control hyperbolic tangent function curvature and by input picture local rectangular window statistical property and calculating, for the average brightness value in local, pixel (x, a y) corresponding place rectangular window in input picture, m _maxmaximum quantization value then for calculating based on input picture local rectangular window statistical property, m _minbe then the minimum quantization value calculated based on input picture local rectangular window statistical property, s as scale factor to control the curvature of hyperbolic tangent function, L _in(x, y) is the intrinsic brilliance value of a pixel (x, y) in input picture, and ε corresponds to the average brightness value in the rectangular window of local, the corresponding place of pixel (x, y) in input picture residual correction value.

After completing the first step or second or the 3rd, the image of existing linear color mapping algorithm to the corresponding first step or second or the 3rd process is adopted to process.

Algorithm detailed step of the present invention is introduced: 1. 3D noise reduction algorithm: in video image, modal noise model is white Gaussian noise, and the appearance of its noise is random, and therefore, the position of the noise that each two field picture occurs is not identical with intensity.3D noise reduction is by carrying out contrast Screening Treatment to the image of a few frame in front and back, nonoverlapping noise automatically being leached, and goes back the original real information in original image.(note: existing noise reduction algorithm is while reduction noise, easily there is the problem such as blocking effect and edge fog, reduce the subjective quality of image on the contrary, and this algorithm to utilize between multiframe the method for most similar pixel coupling, effectively stop the generation of the phenomenons such as fuzzy and smear, while reducing noise, farthest ensure that the quality of image.Adopt the video camera of this 3D noise reduction algorithm, image noise can obviously reduce, and image can be more thorough, thus demonstrate the picture of purer exquisiteness.) 2. non-local mean noise reduction: non-local mean (NLM) filtering algorithm is a kind of image noise reduction algorithm put forward in 2005 by people such as Buades, its cardinal principle is: as shown by way of example in fig. 1, first 1 s in image is got, centered by s, 3 obtain a similar window Δ s for diameter, and then 7 obtain a search window Ws for diameter centered by s.From the most left top point t in the most upper left corner of Ws, centered by t, 3 obtain similar window Δ t for radius, according to the second order norm [1] that Δ s and Δ t subtracts each other || Δ s – Δ t||2 can obtain the weight wt (second order norm value less weight larger) of t point relative to s point by tabling look-up, then similar window Δ t moves through pixel along direction shown in the reciprocal Z-shaped broken lined arrows shown in Fig. 1 so that zigzag trajectory is continuous successively line by line downwards to the right, in like manner, profit uses the same method, a window often mobile pixel just can calculate the weight of this window center point t relative to s point, until the rightest lowest point that t point moves to the most upper right corner of search window terminates, the weight finally points all in search window being multiplied by self is added the s point value that can obtain after noise reduction again.To the institute in original image a little all according to non-local mean (NLM) filtering algorithm process one time, the image after noise reduction can be obtained.The core concept of this filtering algorithm is that other pixels that (or even in full images) searches as much as possible and filtered pixel similar in the large pixel coverage of search window participate among filtering, it not only with reference to the gray scale of partial points, and compare the geometric configuration of whole neighborhood, to realize better filter effect, while noise reduction, farthest remain the details of image.3. high dynamic range images strengthens a kind of low dynamic range echograms with local contrast protection of algorithm-this algorithm use and strengthens disposal route; this algorithm not only can control the saturation degree of color; cross-color is minimized; on the quantizating index of image procossing and visual effect relatively go up; good image output effect can be arrived; and the counting yield of whole process in video source modeling process can be improved, can process in real time vision signal.For this algorithm, the image after enhancing meets local contrast protection, that is:

Formula one: $L_{\mathrm{out}}(x,y)=\frac{L_{\mathrm{out}}^{\mathrm{avg}}(x,y)}{L_{\mathrm{in}}^{\mathrm{avg}}(x,y)}{L}_{\mathrm{in}}(x,y)$

In formula one, L _in(x, y), represent the brightness value of pixel (x, y) in input picture and corresponding local average luminance value respectively.L _out(x, y), represent the brightness value of pixel (x, y) in output image and corresponding local average luminance value respectively.Can be obtained by above formula:

Formula two: $L_{\mathrm{out}}(x,y)={\left\{r(x,y)L_{\mathrm{in}}(x,y)\right\}}_0^1,$ Wherein $r(x,y)=\frac{T\left[L_{\mathrm{in}}^{\mathrm{avg}}(x,y)\right]}{L_{\mathrm{in}}^{\mathrm{avg}}(x,y)}$

In formula two, the span of operator representation L is { between a, b}, r (x, y) represents the mapping coefficient put at (x, y), and T is a continuously differentiable mapping function.By carrying out rational comparison and selection, in the method, T is the self-adaptation hyperbolic tangent function that feature is determined by image local statistical property.It has following major advantage: 1) it can provide adaptive dynamic range of images to adjust; 2) for arbitrary positive input value, the span of output valve is 0 to 1, and this just can ensure that output area is positioned within the scope of expectation all the time; 3) it can strengthen the pixel portion of low-light (level), and is retained for the normal pixel portion of illumination, that is:

Formula three: $L_T^{\tanh }(x,y)=N(x,y)\tanh \left(L_{\mathrm{in}}(x,y){\left[m(x,y)\right]}^{-1}\right)$

In formula three, $N(x,y)={\left[\tanh (L_{\mathrm{in}}^{\max }/m(x,y))\right]}^{-1}$ Be normalized factor, be used for ensureing $L_T^{\tanh }(x,y)=1$ Time, parameter m (x, y) is used for controlling the curvature of hyperbolic tangent function and it is calculated by the partial statistics characteristic of image.The simplest image local statistical property is the local mean value using local window, therefore, and the linear function that definition m (x, y) is image local average, that is:

Formula four: $m(x,y)=L_{\mathrm{in}}^{\mathrm{avg}}(x,y)s+m_{\min }$

In formula four, for scale factor, according to parameter m _minand m _maxbe set, m _maxbe then the maximum quantization value then for calculating based on input picture partial statistics characteristic, m _minminimum quantization value then for calculating based on input picture partial statistics characteristic, can control the curvature of hyperbolic tangent function, and is used for determining the dynamic range compression ability of transport function T.

The final expression formula that high dynamic range images strengthens algorithm can be derived by above four formula:

Formula five: $L_{\mathrm{out}}^{\tanh }(x,y)={\left\{\frac{N(x,y)L_{\mathrm{in}}(x,y)}{L_{\mathrm{in}}^{\mathrm{avg}}(x,y)+\mathrm{\ϵ}}\tanh \left(\frac{L_{\mathrm{in}}^{\mathrm{avg}}(x,y)+\mathrm{\ϵ}}{m(x,y)}\right)\right\}}_0^1$

In formula five, ε is for corresponding to the corresponding local average luminance value of pixel (x, y) in input picture residual correction value.In the middle of enhancing process, in order to avoid colour cast phenomenon, this algorithm can combine with linear color mapping algorithm, is intended to the color information retaining former figure while low-light (level) strengthens.

Claims (2)

1. the method for a process captured low-light (level) image in rugged surroundings, it is characterized in that following treatment step: the first step: the continuous some frame low-light (level) images gathered by the process of existing 3D noise reduction algorithm, to in chronological sequence order continuous some frame low-light (level) images gathered successively carry out contrast screen, the noise of this continuous some frame low-light (level) image non-overlapping copies is leached automatically, makes total most similar pixel coupling wherein; Second step: by existing non-local mean (NLM) the filtering algorithm process handled continuous some frame low-light (level) images crossed of 3D noise reduction algorithm, in continuous some two field pictures gathered successively in chronological sequence order, all pixels all carry out non-mean quantization calculating; 3rd step: adopt high dynamic range images to strengthen algorithm: strengthen 1. algorithmic formula calculates wherein each pixel (x, y) brightness value to this continuous some frame low-light (level) image of non-local mean (NLM) filtering algorithm process by high dynamic range images and with the brightness value of each pixel (x, y) in this continuous some frame low-light (level) image output image, its formula is:

① $L_{\mathrm{out}}^{\tanh }(x,y)={\left\{\frac{N(x,y)L_{\mathrm{in}}(x,y)}{L_{\mathrm{in}}^{\mathrm{avg}}(x,y)+\mathrm{\ϵ}}\tanh \left(\frac{L_{\mathrm{in}}^{\mathrm{avg}}(x,y)+\mathrm{\ϵ}}{m(x,y)}\right)\right\}}_0^1$

High dynamic range images strengthen algorithmic formula 1. in, normalized factor, for the maximum brightness value in local, pixel (x, a y) corresponding place rectangular window in input picture, m (x, y) as linear function with control hyperbolic tangent function curvature and by input picture local rectangular window statistical property and calculating, for the average brightness value in local, pixel (x, a y) corresponding place rectangular window in input picture, m _maxmaximum quantization value then for calculating based on input picture local rectangular window statistical property, m _minbe then the minimum quantization value calculated based on input picture local rectangular window statistical property, s as scale factor to control the curvature of hyperbolic tangent function, L _in(x, y) is the intrinsic brilliance value of a pixel (x, y) in input picture, and ε corresponds to the average brightness value in the rectangular window of local, the corresponding place of pixel (x, y) in input picture residual correction value.

2. the method for process according to claim 1 captured low-light (level) image in rugged surroundings, it is characterized in that: after completing the first step or second or the 3rd, adopt the image of existing linear color mapping algorithm to the corresponding first step or second or the 3rd process to process.