CN106204662A - A kind of color of image constancy method under multiple light courcess environment - Google Patents

Abstract

The invention discloses the color of image constancy method under a kind of multiple light courcess environment, the present invention proposes multiple light courcess color constancy method based on retina physiological mechanism, simulate light dark adaptation ability and the characteristic of retinal neurons receptive field of vision, establish based on dividing bright dark areas based on ON type and the model of OFF receptive field, this model can estimate position and the color of light source well, and can adjust the scene image that parameter adaptation is different.It is few that the method for the present invention has parameter, calculates simple, and speed is fast, effective, it is possible to carry out the features such as real-time process, be very suitable for the front end being built in physical equipment pretreatment and estimate the scene light source colour in image.

Description

A kind of color of image constancy method under multiple light courcess environment

Technical field

The invention belongs to computer vision and technical field of image processing, be specifically related to the scene multiple light courcess face of coloured image Color is estimated and color of image correcting technology.

Background technology

Under natural environment, the visual system of people has the ability of light source colour change in scene that resists.Such as, to same Individual scene, under the most yellow sunlight irradiates, or under the sunlight irradiation that the time-division is the reddest at dusk, our vision system The color of the scene that system is perceived remains constant, and this ability is also referred to as the color constancy of visual system.By skill The restriction of art condition, the picture that current imaging device such as photographing unit photographs often is affected by light source colour in scene Colour cast the most in various degree, this colour cast produced by scene light source Color influences can make follow-up computer vision apply Identify accurately as target recognition, form fit etc. cannot obtain according to color and mate.Therefore, for one width input original Colour cast image, the most effectively removes comprised scene light source colour and is just particularly important.Computational color constancy causes just Power is in solving this problem, and its main purpose calculates the color of the unknown light source included in any piece image, so just After utilize calculated light source colour to original colour cast correct image, obtain the image of display under standard white light.At present The most classical Illuminant estimation method is the gray world method that Buchsbaum proposes for 1980, list of references: G.Buchsbaum,“A spatial processor model for object colour perception,Journal Of the Franklin Institute, vol.310, no.1, pp1 26, July 1980 ", although the method calculates simple, But in most cases it is difficult to entirely accurate and estimates light source colour, and algorithm is just for the field under the conditions of uniform source of light Scape.

Up to the present, the research for the multiple light courcess estimation of non-uniform Distribution is the most fewer, and comparison is the most representational is Hamid Reza Vaezi Joze et al. put forward in 2014, list of references: H.R.V.Joze and M.S.Drew, Exemplar-based color constancy and multiple illumination,Pattern Analysis and Machine Intelligence, IEEE Transactions on, Vol.36, No.5, pp.860-873,2014, but should Method calculates complexity, needs to utilize substantial amounts of sample to be trained, very flexible, is not suitable for processing in real time.

Summary of the invention

The invention aims to solve the defect that existing image scene multiple light courcess color method of estimation exists, propose A kind of color of image constancy method under multiple light courcess environment.

The technical scheme is that the color of image constancy method under a kind of multiple light courcess environment, comprise the steps:

S1. calculate the luminance graph of image: extract the color component in image, the coloured image of input is decomposed into redness, Green and blue three color components, three color components are added the luminance graph obtaining image；

S2. image is divided into bright area and dark areas: the luminance graph that S1 is obtained is divided into two classes, be specifically divided into bright area and Dark areas；

Calculating bright area and the light source of dark areas the most respectively: bright area and the dark areas different scale to artwork respectively Template be filtered, tentatively obtain the bright area of image and the light source of dark areas；

S4. synthesis obtains the light source of whole image: be added by the light source of calculated for step S3 bright area and dark areas, Obtain the light source colour of whole image；

S5. light source colour is eliminated, it is achieved color constancy: be respectively divided by correspondence by each pixel in original colour cast image The pixel of calculated light source image, the image without colour cast after being corrected.

Wherein, the purpose dividing bright dark areas in step S2 is the mechanism of simulation retina light dark adaptation, in retina Clearly, the cell that different brightness level are participated in calculating is different for cone cell and the retinal rod division of labor.

Use classical K-Mean Method to gather in the bright dark areas of actual division to be two classes, be divided into dark areas and luminance area, Original color image bright area is labeled as 1, and dark areas is labeled as 0.

Filtering Template in step S3 is the DOG template of a double gauss difference, and formula is as follows:

$R_1=\frac{1}{2{\mathrm{\π\σ}}_1^2}\exp (-\frac{x^2+y^2}{2{\mathrm{\σ}}_1^2})-\frac{k}{2\mathrm{\π}{\left({\mathrm{\λ\σ}}_1\right)}^2}\exp (-\frac{x^2+y^2}{2{\left({\mathrm{\λ\σ}}_1\right)}^2})$

$R_2=\max \left(R_1\right)+\min \left(R_1\right)-\left(\frac{1}{2{\mathrm{\π\σ}}_2^2}\exp \right(-\frac{x^2+y^2}{2{\mathrm{\σ}}_2^2})-\frac{k}{2\mathrm{\π}\left({\mathrm{\λ\σ}}_2\right)}\exp (-\frac{x^2+y^2}{2{\left({\mathrm{\λ\σ}}_2\right)}^2}\left)\right)$

Wherein, R₁Being the model of ON type receptive field, this model center just, bear, the here as filtering mould of bright area by periphery Plate, R₂Being the model of OFF type receptive field, this model center is born, outer straight, the here as Filtering Template of dark areas, its middle mold The yardstick of plate periphery is 3 times of center yardstick, and namely the value of λ is 3, and additionally Filtering Template is a unbalanced double gauss Difference, namely the value of weight k is not equal to 1, and the span of k is k ∈ (0.1,1) here, and wherein bright area filter scale is wanted Less than dark areas filter scale, namely σ₁＜ σ₂。

Here as a kind of preferred version, meets σ₁∈ [2,10], σ₂∈[3,20]。

The light source calculating bright area and dark areas respectively described in above-mentioned steps S3 specifically includes the most step by step:

S31. artwork and R₁Do convolution and obtain L₁；

S32. artwork and R₂Do convolution and obtain L₂；

S33.L₁In dark areas part be set to 0, the part being namely labeled as 0 is set to 0, L₂Middle bright area part is set to 0, the part being namely labeled as 1 is set to 0；And then obtain bright area and the light source of dark areas.

After above-mentioned S5, the light source of each pixel in image can be calculated, obtain the light source that a width is complete Image, this light source figure can be used to correct original colour cast image so that it is recovers image originally under standard light source (white light) Coloured image.

Beneficial effects of the present invention: first the cromogram of input is divided by the color constancy method of the present invention according to brightness difference For bright area and dark areas, then using positive periphery, center to bear in bright area, yardstick is different and amplitude double gauss difference type not etc. Wave filter (Difference of Gaussian, DoG), uses the outer straight DoG wave filter of central negative to enter respectively in dark areas Row Filtering Processing, the present invention has substantially drawn in human visual system retina light dark adaptation and neuron ON type and OFF The feature of type receptive field, the receptive field scope (DoG filter center and the yardstick of periphery) of suitable regulation periphery, center and The sensitivity intensity (amplitude of DoG wave filter periphery) of periphery receptive field, such that it is able to change spatial frequency modulation characteristic (such as It is changed between bandpass characteristics and low-pass characteristic), sensitivity intensity (amplitude size) energy of regulation periphery receptive field in addition Enough efficiently extracting shades of colour region in scene, color boundaries information preferably estimates the color of scene light source, thus ON The periphery sensitivity intensity (amplitude size) that type receptive field yardstick, OFF type receptive field yardstick and two receptive fields are identical is three Individual main parameter.The method of the present invention has parameter few (only three adjustable parameters, i.e. two yardsticks and an amplitude), meter Calculating simple, speed is fast, effective, it is possible to carry out the features such as real-time process, is very suitable for being built in physical equipment (such as photograph Machine) the scene light source colour in image estimated by the front end of pretreatment.

Accompanying drawing explanation

Fig. 1 is the schematic flow sheet of the scene light source colour method of estimation of the coloured image of the present invention.

Fig. 2 is the distribution of light sources figure (2a) of synthetic in embodiment, the distribution of light sources figure (2b) that bright area is estimated, dark space The distribution of light sources figure (2c) that territory is estimated, the final complete distribution of light sources figure (2d) estimated.

Fig. 3 is width original input picture (Fig. 3 a) in embodiment, adds the image (Fig. 3 b) after synthetic light source, And carry out the result (Fig. 3 c) after tint correction.

Detailed description of the invention

Inherently there is the ability of color constancy in view of human vision, it is possible to the colour cast in correction scene, retina automatically Having ability and the characteristic of retinal neurons receptive field of light dark adaptation, based on this, the present invention proposes based on retina The multiple light courcess color constancy method of physiological mechanism.This method simulates light dark adaptation ability and the retinal neurons impression of vision Wild characteristic, establish based on dividing bright dark areas based on ON type and the model of OFF type receptive field, this model can be very Estimate well position and the color of light source, and the scene image that parameter adaptation is different can be adjusted.

It is specifically described below by an embodiment.

Piece image is downloaded from the most internationally recognized image library website for estimate scene light source colour (8D5U5535.png) and the real light sources color of correspondence, image size is 340 × 511, and wherein 8D5U5535 image does not has Pretreatment (such as tint correction, gamma value corrects) through any camera itself.The schematic flow sheet of the detailed step of the present invention As it is shown in figure 1, detailed process is as follows:

S1. the luminance graph of image is calculated: first the coloured image of input is decomposed into redness, green and blue three colors Component, three color components are added the luminance graph obtaining image.

With two pixels (136,45) and (397,172) of original input image, pixel value be respectively (0.026,0.035, 0.026) with (0.868,1,0.842) as a example by, be decomposed into the value after redness, green and blue three color components respectively 0.026,0.035,0.026 and 0.868,1,0.842.RGB color component carries out being added the brightness put as each, then The brightness of two pixels is respectively 0.026+0.035+0.026=0.087,0.868+1+0.842=2.710.

S2. image is divided into bright area and dark areas: the brightness value obtained according to S1, uses classical K-Mean Method, Image is gathered for dark areas and luminance area two class.We are labeled as 1 original color image bright area, and dark areas is labeled as 0.

As a example by pixel in S1: the brightness value of pixel (136,45) is 0.087, is worth on the low side, K-average is used to gather Gathered into dark areas class after class, be labeled as 0；The brightness value of pixel (397,172) is 2.710, is worth higher, uses K-average to gather Gathered into bright area class after class, be labeled as 1.

Calculate bright area and the light source of dark areas the most respectively: use the wave filter bright area to being obtained by step S2 respectively Class carries out convolutional calculation with the pixel of dark areas class.

Bright area specifically makes in the present embodiment positive periphery, employing center bear, and yardstick is different and amplitude double gauss not etc. is poor Mode filter, uses the outer straight double gauss difference mode filter of central negative to be filtered respectively processing in dark areas.For difference Image scene, yardstick (σ can be adjusted flexibly₁And σ₂) and three parameters of amplitude (k), thus estimate the color of scene light source.

Input pixel value (0.026,0.035,0.026) is carried out with (0.868,1,0.842) by step S1～S2 Calculated it be labeled as the dark areas (0.026,0.035,0.026) of 0 and be labeled as the bright area (0.868,1,0.842) of 1 As a example by.

Here following a kind of calculation can be used:

S31. artwork and wave filter R₁Do convolution and obtain L₁, dark areas (0.026,0.035,0.026) after convolution at L₁In Value be (0.026,0.030,0.022), bright area (0.868,1,0.842) after convolution at L₁In value be (0.880, 0.998,0.880)。

S32. artwork and wave filter R₂Do convolution and obtain L₂, dark areas (0.026,0.035,0.026) after convolution at L₂In Value be (0.050,0.054,0.043), bright area (0.868,1,0.842) after convolution at L₂In value be (0.840, 0.952,0.839)。

S33.L₁In dark areas part be set to 0, be namely labeled as the part of 0 and be set to 0, dark areas (0.026, 0.035,0.026) be labeled as 0, at L₁In be set to (0,0,0).

L₂Middle bright area part is set to 0, and the part being namely labeled as 1 is set to 0, bright area (0.868,1,0.842) It is labeled as 1, at L₂In be set to (0,0,0).L₁With L₂Both are added and obtain the light source that the whole image first step is estimated.

R in this example₁Middle convolution yardstick σ₁It is 3, R₂Middle convolution yardstick σ₂It is 20, the periphery receptive field power of two convolution masks Weight k is 0.5.

S4. synthesis obtains the light source of whole image: by calculated for bright area light source L₁It is calculated light with dark areas Source L₂It is added the light source colour obtaining whole image..

As a example by two pixel values (0.026,0.035,0.026) in S1 and (0.868,1,0.842), (0.026, 0.035,0.026) light source colour that pixel is corresponding be (0.050+0,0.054+0,0.043+0)=(0.050,0.054, 0.043) light source colour that, (0.880,0.998,0.880) point is corresponding be (0.880+0,0.998+0,0.880+0)= (0.880,0.998,0.880)。

Fig. 2 illustrates the distribution of light sources figure (2a) of synthetic, bright area calculated light source L₁Scattergram (2b), secretly Region calculated light source L₂Scattergram (2c), the final complete distribution of light sources figure (2d) estimated.

S5. light source colour is eliminated, it is achieved color constancy.It is respectively divided by correspondence by each pixel in original colour cast image The pixel of calculated light source image, the image without colour cast after being corrected.

Utilize the light source colour value under step S5 each color component calculated, correct original input picture respectively The pixel value of each color component.With two pixel values (0.026,0.035,0.026) of original input picture in step S1 with As a example by (0.868,1,0.842), the result after its correction is respectively (0.026/0.050,0.035/0.054,0.026/0.043) =(0.520,0.648,0.605), (0.868/0.880,1/0.998,0.842/0.880)=(0.986,1.002,0.957), Then the value after correction is multiplied by standard white backscatter extinction logarithmic ratioRespectively obtain 0.300,0.374,0.349 with 0.569,0.579,0.553 as the pixel value of the correction chart picture of final output, and other pixel value of original input picture also does Similar calculating, finally obtains the coloured image after correction.

Above simplified example mainly illustrates with the single pixel value of image for example, is at view picture figure during Practical Calculation Carry out on all pixel values of picture.

Examples detailed above is very full on the method for the present invention and calculates scene light source colour and remove light source impact and then reality The whole process of color constancy in the case of existing multiple light courcess.

Fig. 3 c is many after the light source colour value utilizing step S5 to calculate adds artificial light sources to original image (Fig. 3 a) Light source figure (Fig. 3 b) carries out the result after tint correction.

Those of ordinary skill in the art it will be appreciated that embodiment described here be to aid in reader understanding this Bright principle, it should be understood that protection scope of the present invention is not limited to such special statement and embodiment.This area It is each that those of ordinary skill can make various other without departing from essence of the present invention according to these technology disclosed by the invention enlightenment Planting concrete deformation and combination, these deform and combine the most within the scope of the present invention.

Claims (6)

1. the color of image constancy method under multiple light courcess environment, comprises the steps:

S1. the luminance graph of image is calculated: extract the color component in image, the coloured image of input is decomposed into redness, green With blue three color components, three color components are added the luminance graph obtaining image；

S2. image is divided into bright area and dark areas: the luminance graph that step S1 is obtained is divided into two classes, be specifically divided into bright area and Dark areas；

Calculating bright area and the light source of dark areas the most respectively: bright area and the mould of dark areas different scale to artwork respectively Plate is filtered, and tentatively obtains the bright area of image and the light source of dark areas；

S4. synthesis obtains the light source of whole image: is added by the light source of calculated for step S3 bright area and dark areas, obtains The light source colour of whole image；

S5. light source colour is eliminated, it is achieved color constancy: be respectively divided by the calculating of correspondence by each pixel in original colour cast image The pixel of the light source image obtained, the image without colour cast after being corrected.

Color of image constancy method under multiple light courcess environment the most according to claim 1, it is characterised in that divide bright dark space It is two classes that territory uses classical K-Mean Method to gather, and is divided into dark areas and luminance area, original color image bright area is labeled as 1, dark areas is labeled as 0.

Color of image constancy method under multiple light courcess environment the most according to claim 1, it is characterised in that in step S3 Filtering Template is the DOG template of a double gauss difference, and formula is as follows:

$R_1=\frac{1}{2{\mathrm{\π\σ}}_1^2}\exp (-\frac{x^2+y^2}{2{\mathrm{\σ}}_1^2})-\frac{k}{2\mathrm{\π}{\left({\mathrm{\λ\σ}}_1\right)}^2}\exp (-\frac{x^2+y^2}{2{\left({\mathrm{\λ\σ}}_1\right)}^2})$

$R_2=\max \left(R_1\right)+\min \left(R_1\right)-\left(\frac{1}{2{\mathrm{\π\σ}}_2^2}\exp \right(-\frac{x^2+y^2}{2{\mathrm{\σ}}_2^2})-\frac{k}{2\mathrm{\π}\left({\mathrm{\λ\σ}}_2\right)}\exp (-\frac{x^2+y^2}{2{\left({\mathrm{\λ\σ}}_2\right)}^2}\left)\right)$

Wherein, R₁Being the model of ON type receptive field, this model center just, bear, the here as Filtering Template of bright area, R by periphery₂ Being the model of OFF type receptive field, this model center is born, outer straight, the here as Filtering Template of dark areas, wherein template periphery Yardstick be 3 times of center yardstick, namely the value of λ is 3, and additionally Filtering Template is that a unbalanced double gauss is poor, the most just It is that the value of weight k is not equal to 1, wherein, bright area filter scale σ₁It is less than dark areas filter scale σ₂。

Color of image constancy method under multiple light courcess environment the most according to claim 3, it is characterised in that institute in step S3 The light source calculating bright area and dark areas respectively stated specifically includes the most step by step:

S31. artwork and R₁Do convolution and obtain L₁；

S32. artwork and R₂Do convolution and obtain L₂；

S33.L₁In dark areas part be set to 0, the part being namely labeled as 0 is set to 0, L₂Middle bright area part is set to 0, also It is exactly to be labeled as the part of 1 to be set to 0；And then obtain bright area and the light source of dark areas.

Color of image constancy method under multiple light courcess environment the most according to claim 3, it is characterised in that described weight The span of k is k ∈ (0.1,1).

Color of image constancy method under multiple light courcess environment the most according to claim 3, it is characterised in that described σ₁∈ [2,10], σ₂∈[3,20]。