CN104537615A - Local Retinex enhancement algorithm based on HSV color spaces - Google Patents

Abstract

The invention discloses a local Retinex dehazing algorithm based on HSV color spaces. The local Retinex dehazing algorithm is used for conducting image enhancement on an image with fog being unevenly distributed under the foggy day condition. The method includes the following steps that A, the multi-layered foggy day image is processed in different color spaces, so that the image contrast ratio is enhanced, and the image definition is improved; B, the size of the image is obtained, the sizes of subblock images are judged in a self-adaptation mode, and a luminance image of the foggy day image is partitioned; enhancement processing is conducted on all sub images of the luminance component, and the image definition is improved; C, the area proportion factors of all pixel points in the overlapped sub images are calculated, weighted summation is conducted on all the overlapped sub images, and image fusion is achieved; D, the saturability component of the image is processed, and the image color is restored. By means of the local Retinex dehazing algorithm, scenery details at different levels can be restored clearly and quickly, the visual effect of the image is improved, and the algorithm is suitable for aerial photo systems at different regions.

Description

A kind of local Retinex based on HSV color space strengthens algorithm

Technical field

The invention belongs to the technical field of Image Information Processing, the present invention is a kind of local Retinex mist elimination algorithm based on HSV color space, the atomization image taken under being applicable to strengthen haze weather.

Background technology

Numerous outdoor monitoring systems of computer utility all requirement can extract the feature of image exactly, and under the weather conditions such as haze, often due to the impact of fine particle scattering process in air, the image obtained is caused seriously to degrade, not only have impact on the visual effect of image, and disturb the extraction of characteristics of image, outdoor monitoring system normally cannot be run, thus bring serious potential safety hazard.Therefore, utilize effectively and fast defogging method capable to make atomization image sharpening, have important practical significance.

Existing image defogging method capable is mainly divided into two large classes: the disposal route based on image restoration and the disposal route based on image enhaucament.

Be theoretical foundation based on the algorithm of image restoration with atmospherical scattering model, atmospherical scattering model describes Misty Image causes for Degradation; Image Restoration Algorithm by inverting image degradation process, restoration scenario albedo, with obtain optimization without mist image.These class methods are from the essence of image degradation, and the result obtained is usually comparatively natural.But atmospherical scattering model is an ill-condition equation, cannot draw and need the albedo of scene to utilize the constraint conditions such as prior imformation, the parameter in atmospherical scattering model is estimated by direct solution equation, and then the albedo of reduction scene.

Algorithm based on image enhaucament does not consider image quality decrease reason, by carrying out the stretching of entirety or local to image, improves brightness of image, contrast and outstanding details, thus improves visual effect.At present, the main stream approach based on image enhaucament is Retinex algorithm, based on the color constancy of people's vision, by the luminance component of filtering estimated image, obtains reflecting component and realizes image mist elimination.But the method malleable color ratio causes color distortion, can lose local detail while stretching, and is difficult to the parameter value choosing suitable wave filter.

Paper name: A Multiscale Retinex for Bridging the Gap Between Color Images andthe Human Observation of Scenes, periodical: IEEE Transaction on image processing, time: 1997,7th phase, 965 pages to 976 pages, the people such as Daniel J.Jobson propose the multi-Scale Retinex Algorithm of band color recovery, according to color constancy principle, adopt Gaussian filter function estimated image luminance component, also original image reflecting component, linear stretch mode improves brightness and contrast.But filter parameter chooses difficulty, and multi-level Misty Image cannot be reduced.

Proprietary term: Smart Image Enhancement Process, periodical: Patent ApplicationPublication, time: in October, 2010, numbering: US 2010/0266214Al, first the people such as Daniel J.Jobson carry out judging whether image has mist to the contrast of input picture and brightness, then the enhancing algorithm that histogram equalization and MSRCR algorithm combine is carried out to Misty Image, finally iterate and strengthen until sharpness and brightness reach requirement or iterations reaches the upper limit.

Paper name: Nonlinear Transfer Function-Based Local Approach for Color ImageEnhancement, periodical: IEEE Transactions on Consumer Electronics, time: 2011, 16th phase, 858 pages to 865 pages, propose a kind of image enchancing method based on HSV color space, first image is transformed into HSV color space, then non-linear transfer function is adopted to carry out contrast strengthen to V component (i.e. luminance component) separately, and do not change S component and the H component of image, retain color, effectively can strengthen the even image of uneven illumination, but be not suitable for the process of Misty Image.

Summary of the invention

The present invention, in order to overcome the deficiency of conventional images mist elimination technology, provides a kind of local Retinex mist elimination algorithm based on HSV color space, can either strengthen the sharpness of atomization image, improve the visual effect of image, have again travelling speed faster, real-time is good, wide accommodation.

Based on a local Retinex mist elimination algorithm for HSV color space, the method comprises the steps:

A, in different color spaces, process multi-level Misty Image, realize picture superposition and sharpness promotes; Rgb color space process image easily causes color distortion; HSV color space is suitable for human visual perception, describes the feature of image saturation, tone, brightness; Saturation degree describes the impact of mist on the Scene colors depth; Brightness describes the impact of atmosphere light on image definition; Tone describes the type of Scene colors; Strengthen image at HSV color space, do not affect the ratio of RGB tri-color component, thus the skew of color can not be affected.

B, acquisition image size, the size of adaptive judgement sub-image, carries out piecemeal to the luminance picture of Misty Image; Enhancing process is carried out to each subimage of luminance component, improves the sharpness of image;

Misty Image exists without mist, mist, dense fog, thick fog four kinds of situations, and single Gaussian convolution yardstick cannot adapt to multiple situation; In order to obtain the details in more thick fog regions, choosing corresponding convolution yardstick, utilizing Gaussian convolution to estimate luminance component, reducing clear reflecting component.

According to the size of source images, determine the subimage size after the number of piecemeal and piecemeal; According to partly overlapping mode, the luminance picture (V component) of image is divided into several subimages; Calculate the scale size of Gauss around convolution function of each subimage, MSR is carried out to each subimage and strengthens process.Concrete steps are as follows:

B1: after obtaining the luminance picture of source images, the size of computed image, and the size of subimage after piecemeal; Image is divided into the indefinite subimage of number according to the image size of setting, partly overlapping mode; Picture size is large, then the piecemeal number chosen is many;

B2: Gauss is around in convolution function, and scale size determines the effect of the rear image of process: after large scale (as σ=200) process, color fidelity is good; After small scale (as σ=15) process, dynamic range compression is good; Three yardstick weighted sums are realized color fidelity and dynamic range compression by MSR algorithm simultaneously; MSR is carried out to subimage and strengthens process:

$V_o(i,j)=\underset{n=1}{\overset{N}{\mathrm{\Σ}}}{W}_n\left\{\log \right[V(i,j)]-\log [V(i,j)*F_n(i,j)\left]\right\}---\left(1\right)$

Wherein, the card coordinate that (i, j) is pixel, n is yardstick label, span: [1, N], and N is the total number of yardstick, and V (i, j) is the luminance picture of input, V _o(i, j) luminance picture for exporting, W _nfor weights, value be, for Gauss is around convolution function, expression formula is for 1/3,1/3,1/3}, F (i, j):

F(i,j)＝K exp[-(i ²+j ²)/σ ²] (2)

Wherein, the card coordinate that (i, j) is pixel, σ is yardstick, and K is constant.

C, calculate the area ratio factor of each pixel in each overlapping subimage, summation is weighted to each overlapping subimage and realizes image co-registration;

Fusion treatment is carried out to each subimage and obtains the rear luminance picture of enhancing; With neighboring pixel areas in the nearest border of current pixel point in each subimage for scale factor, with the ratio of this area and total overlapping area for weights, adopt alpha amalgamation mode to merge, eliminate blocking effect; Concrete steps are as follows:

C1, multi-Scale Retinex Algorithm are according to the value of the relativeness decision pixel of surrounding pixel; One direction (level or vertical) chooses surrounding neighbors pixel, can not reflect the impact of border surrounding pixel on current pixel point completely; But jointly determined by the relativeness of pixels all around pixel and border (horizontal and vertical direction); Therefore, choosing horizontal and vertical direction pixel, take area as the scale factor of current sub-image pixels point;

C2, calculating pixel (i, j) neighborhood territory pixel area in each subimage: with pixel (i, j) for starting point, the distance to nearest border (horizontal and vertical direction) is the area ω of the square region that limit surrounds _n; Calculate the total area ω of the overlapping region containing current pixel; With the neighborhood territory pixel area ω of each subimage _nbe weights with the ratio of overlapping region total area ω:

${\mathrm{\α}}_n=\frac{{\mathrm{\ω}}_n}{\mathrm{\ω}}---\left(3\right)$

Wherein, n is the label of subimage, span: [1, N], and N is the number of overlapping subimage.

C3, merge overlapping subimage, obtain reflecting component clearly; Adopt Alpha amalgamation mode to merge, area ratio is as weights:

$R_v(i,j)=\underset{n=1}{\overset{N}{\mathrm{\Σ}}}{\mathrm{\α}}_n{R}_n(i,j)---\left(4\right)$

Wherein, the card coordinate that (i, j) is pixel, R _v(i, j) is the brightness value after fusion, R _n(i, j) is (i, j) end value in each subimage, and n is the label of subimage, span: [1, N], and N is the number of overlapping subimage.

D, the saturation degree component of image to be processed, reduction image color;

In Misty Image, the existence of mist makes the color saturation of target scene reduce; S component in HSV color space reflects the saturation degree of each scene point in image; There is corresponding relation in saturation degree and the brightness of HSV color space, the corresponding relation according to brightness change before and after saturation degree and enhancing judges the change of pixel saturation degree; Step is as follows:

The conversion formula of D1, foundation HSV color space and rgb color space, meets following relation between saturation degree component (S component) and luminance component (V component):

$\begin{array}{c}V=\max \left(\mathrm{rgb}\right)\\ S=\frac{\max \left(\mathrm{rgb}\right)-\min \left(\mathrm{rgb}\right)}{\max \left(\mathrm{rgb}\right)}=1-\frac{\min \left(\mathrm{rgb}\right)}{\max \left(\mathrm{rgb}\right)}\\ \⇒S=1-\frac{\min \left(\mathrm{rgb}\right)}{V}\end{array}---\left(5\right)$

Wherein, S is saturation degree component, and V is luminance component, and rgb is three color components of rgb color space.

D2, supposition image are before and after enhancing, and the three-channel minimum value of RGB can not change; Obtain the changing value of luminance component before and after strengthening, calculate the S component after change:

$S=1-\frac{K}{V}$

$S^{\′}=1-\frac{K}{V^{\′}}---\left(6\right)$

$S^{\′}=S-\frac{K}{V^{\′}}+\frac{K}{V}$

Wherein, S is untreated saturation degree component, and V is not for highlighting component, and S ' is the saturation degree component after process, and V ' is for strengthening rear luminance component, and K represents the three-channel minimum value of RGB.

In D3, reality, the three-channel minimum value of RGB before and after strengthening can change; Gamma is utilized to convert the dynamic change compensating minimum value.

S _o＝(S′) ^γ(7)

Wherein, S _ofor the saturation degree component exported, S ' is the saturation degree component after first time process, and γ is Gamma coefficient.

The invention has the beneficial effects as follows: in conjunction with the advantage of Local treatment image, adopt MSR to strengthen sharpness and the color rendition of Misty Image, improve visual detail.There is blocking effect phenomenon for Local treatment image, propose area A lpha fusion method, adopt the ratio of area as the weight of subimage, remove blocking effect.Utilize the characteristic of HSV color space, reduce the execution number of times of Gaussian convolution, improve execution efficiency.Utilize the mutual conversion regime between HSV color space and rgb color space, obtain the relation that saturation degree divides change of variable and luminance component, realize color rendition, reduce image fault.

Accompanying drawing explanation

Fig. 1 is particular flow sheet of the present invention.

Fig. 2 is that V component of the present invention strengthens algorithm flow chart.

Fig. 3 is that S component of the present invention strengthens algorithm flow chart.

Fig. 4 is instance processes result figure of the present invention.Wherein, (a) is original atomization image; B () is the initial V component image of (a); C () is the initial S component image of (a); D () is V component image after the enhancing of (a); E () is the RGB triple channel minimum value of (a); F () is S component image after strengthening the first time of (a); G () is mist elimination result after strengthening the first time of (a); H () is S component image after the second time of (a) strengthens; I () is for the present invention is to the mist elimination result of (a); J () MSRCR is to the mist elimination result of (a).

Embodiment

The present invention is elaborated below in conjunction with specific embodiments and the drawings.

A, obtain atomization image (see accompanying drawing 4 (a) Suo Shi) HSV triple channel component.V component (see accompanying drawing 4 (b) Suo Shi) represents the brightness of image, and scope is [0-255], and 0 represents black, and 255 represent white; S component (see accompanying drawing 4 (c) Suo Shi) represents the saturation degree component of image, and scope is between [0-1], and be display image, each pixel is multiplied by 255.

B, utilize adaptive local Retienx strengthen algorithm strengthen image, the schematic flow sheet of method is shown in shown in accompanying drawing 2:

B1, first choose the number of image block, it is little that little image chooses piecemeal number, and large image block number is large; This image size is: 2627*1926mm, and choosing piecemeal number is 10*7, and subimage size is 262*275mm;

B2, according to the subimage size chosen, the method for the piecemeal that partly overlaps is adopted to carry out piecemeal to the V component (accompanying drawing 4 (b)) of input picture; According to formula (2), Gaussian convolution filtering is carried out to the V component subimage after piecemeal and obtains luminance component, remove luminance component acquisition reflecting component according to formula (1) and realize brightness enhancing;

C, the Alpha fusion method based on area factor is utilized to merge the V component after being enhanced to subimage:

C1, to strengthen according to step B2 and obtain each subimage, calculate the influence area of overlapping region pixel, using this pixel to border (horizontal and vertical direction) nearest distance as the length of area and wide, ask for area; Using the region of four subimage overlaps as the total area, ask for area factor α according to formula (3) _n.In concrete enforcement, overlapping region is 1/4th of subimage;

C2, according to C1 obtain area factor α _nafter, be weighted summation according to formula (4) pixel to overlapping region, obtain the image after merging.Thus, the V component after being enhanced is as shown in accompanying drawing 4 (d).

D, enhancing S component, the color of Recovery image.Wherein, the schematic flow sheet of algorithm is as shown in Figure 3:

D1, according to formula (5) (6), obtain the changing value of V component and the minimum value figure (as Suo Shi accompanying drawing 4 (e)) of input picture before and after strengthening; Then according to S component and RGB minimum value and the relation before and after strengthening between V component, the S component after strengthening is obtained as shown in accompanying drawing 4 (f).

After D2, the S component (accompanying drawing 4 (f)) obtained according to D1 strengthen, there is skew (as Suo Shi accompanying drawing 4 (g)) in the color of image; This is because counter the changing of RGB minimum value before and after strengthening; Therefore, Gamma compensation is carried out according to formula (7).In concrete enforcement, Gamma penalty coefficient γ value is 0.8.S component map after being enhanced is as shown in accompanying drawing 4 (h).

Through above-mentioned steps, Fig. 4 (i) is for the present invention is to the mist elimination result of atomization image 4 (a).Fig. 4 (j) is for MSRCR method is to the mist elimination result of Fig. 4 (a).

Claims (3)

1., based on a local Retinex mist elimination algorithm for HSV color space, it is characterized in that following steps:

A, multi-level Misty Image is transformed into HSV color space processes, realize picture superposition and sharpness promotes;

B, acquisition image size, the size of adaptively selected sub-image, carries out piecemeal according to partly overlapping mode to the luminance picture of Misty Image; MSR is carried out to each subimage of luminance picture and strengthens process, improve the sharpness of image; Concrete steps are as follows:

B1: after obtaining the luminance picture of source images, the size of computed image, and the size of subimage after piecemeal; Image is divided into the indefinite subimage of number according to the image size of setting, partly overlapping mode; Picture size is large, then the piecemeal number chosen is many;

B2: Gauss is around in convolution function, and scale size determines the effect of the rear image of process: after large scale process, color fidelity is good; After small scale process, dynamic range compression is good; Three yardstick weighted sums are realized color fidelity and dynamic range compression by MSR algorithm simultaneously; MSR is carried out to subimage and strengthens process:

$V_o(i,j)=\underset{n=1}{\overset{N}{\mathrm{\Σ}}}{W}_n\left\{\log \right[V(i,j)]-\log [V(i,j)*F_n(i,j)\left]\right\}---\left(1\right);$

Wherein, the card coordinate that (i, j) is pixel, n is yardstick label, span: [1, N], and N is the total number of yardstick, and V (i, j) is the luminance picture of input, V _o(i, j) luminance picture for exporting, W _nfor weights, value be, for Gauss is around convolution function, expression formula is for 1/3,1/3,1/3}, F (i, j):

F(i,j)＝K exp[-(i ²+j ²)/σ ²] (2)

Wherein, the card coordinate that (i, j) is pixel, σ is yardstick, and K is constant;

C, according to the alpha amalgamation mode based on area, calculate the weights of each pixel in each overlapping subimage, summation is weighted to the result of each pixel in each overlapping subimage and realizes image co-registration;

D, according to the relation between saturation degree component and luminance component, the saturation degree component of image is processed, and uses Gamma to compensate, reduce image color.

2. local according to claim 1 Retinex mist elimination algorithm, is characterized in that, in step C, carries out fusion treatment obtain the rear luminance picture of enhancing to each subimage; With neighboring pixel areas in the nearest border of current pixel point in each subimage for scale factor, with the ratio of this area and total overlapping area for weights, adopt alpha amalgamation mode to merge, eliminate blocking effect; Concrete steps are as follows:

C1, multi-Scale Retinex Algorithm are according to the value of the relativeness decision pixel of surrounding pixel; Level or the surrounding neighbors pixel selected by vertical direction, can not reflect the impact of border surrounding pixel on current pixel point completely; But jointly determined by the relativeness of all pixels around the border in pixel and horizontal and vertical direction; Therefore, choosing horizontal and vertical direction pixel, take area as the scale factor of current sub-image pixels point;

C2, calculating current pixel point neighborhood territory pixel area in each subimage: take current pixel point as starting point, the distance to nearest border, horizontal and vertical direction is the area ω of the square region that limit surrounds _n; Calculate the total area ω of the overlapping region containing current pixel; With the neighborhood territory pixel area ω of each subimage _nbe weights with the ratio of overlapping region total area ω:

${\mathrm{\α}}_n=\frac{{\mathrm{\ω}}_n}{\mathrm{\ω}}---\left(3\right)$

Wherein, n is the label of subimage, span: [1, N], and N is the number of overlapping subimage;

C3, merge overlapping subimage, obtain reflecting component clearly; Adopt Alpha amalgamation mode to merge, area ratio is as weights:

$R_v(i,j)=\underset{n=1}{\overset{N}{\mathrm{\Σ}}}{\mathrm{\α}}_n{R}_n(i,j)---\left(4\right)$

Wherein, the card coordinate that (i, j) is pixel, R _v(i, j) is the brightness value after fusion, R _n(i, j) is (i, j) end value in each subimage, and n is the label of subimage, span: [1, N], and N is the number of overlapping subimage.

3. the local Retinex mist elimination algorithm according to claim 1 and 2, is characterized in that, in step D, in Misty Image, the existence of mist makes the color saturation of target scene reduce; S component in HSV color space reflects the saturation degree of each scene point in image; There is corresponding relation in saturation degree and the brightness of HSV color space, the corresponding relation according to brightness change before and after saturation degree and enhancing judges the change of pixel saturation degree; Step is as follows:

The conversion formula of D1, foundation HSV color space and rgb color space, meets following relation between saturation degree component and luminance component:

$\begin{array}{c}V=\max \left(\mathrm{rgb}\right)\\ S=\frac{\max \left(\mathrm{rgb}\right)-\min \left(\mathrm{rgb}\right)}{\max \left(\mathrm{rgb}\right)}=1-\frac{\min \left(\mathrm{rgb}\right)}{\max \left(\mathrm{rgb}\right)}\\ \⇒S=1-\frac{\min \left(\mathrm{rgb}\right)}{V}\end{array}---\left(5\right)$

Wherein, S is saturation degree component, and V is luminance component, and rgb is three color components of rgb color space;

D2, supposition image are before and after enhancing, and the three-channel minimum value of RGB can not change; Obtain the changing value of luminance component before and after strengthening, calculate the S component after change:

$S=1-\frac{K}{V}$

$S^{\′}=1-\frac{K}{V^{\′}}---\left(6\right)$

$S^{\′}=S-\frac{K}{V^{\′}}+\frac{K}{V}$

Wherein, S is untreated saturation degree component, and V is not for highlighting component, and S ' is the saturation degree component after process, and V ' is for strengthening rear luminance component, and K represents the three-channel minimum value of RGB;

In D3, reality, the three-channel minimum value of RGB before and after strengthening can change; Gamma is utilized to convert the dynamic change compensating minimum value;

S _o＝(S′) ^γ(7)

Wherein, S _ofor the saturation degree component exported, S ' is the saturation degree component after first time process, and γ is Gamma coefficient.