CN105096278B - The image enchancing method adjusted based on illumination and equipment - Google Patents

Abstract

The present invention proposes a kind of self adaptation Gamma Enhancement Method theoretical based on Retinex, utilize luminance component and the reflecting component of Retinex theory separate picture, luminance component is carried out the correction of self adaptation Gamma, finally utilizes reflecting component to recover details and the color of image.The present invention solves existing Retinex algorithm poor in timeliness and the undesirable problem for the treatment of effect, and the inhomogeneous illumination image of process has optimal contrast, visibility, naturality and ageing.

Description

The image enchancing method adjusted based on illumination and equipment

Technical field

The present invention relates to image processing field, particularly to the image enhaucament side adjusted based on illumination Method and equipment.

Background technology

Along with mobile device performance is constantly upgraded and popularizes, in daily life, people are increasingly Get used to anywhere or anytime things interested being taken pictures or imaging, but these images or video Generation be often under conditions of non-limiting, this results in and can there is uneven illumination in image Even situation.Process for inhomogeneous illumination image has become image processing field in the urgent need to solving Problem certainly.

In order to solve the even problem of uneven illumination, scientific research personnel has carried out substantial amounts of research.Retinex (Retina and Cortex) theory is widely used in the enhancing of the even image of uneven illumination.Should Theory hypothesis image is to be combined into by luminance component and reflecting component, and two components are permissible Separating treatment.Retinex algorithm in early days, while compensating illumination, has suppressed image Dynamic range, and there will be halation and cross-color phenomenon.In order to solve halation and color distortion Problem, use the method that multiple dimensioned Retinex tone maps to strengthen inhomogeneous illumination image, But strengthen image and seem the most natural.In conjunction with Retinex principle and double-log luminance filter, Improve visibility and the naturality of image further, but its computation complexity is the highest, the most greatly, Brightness flop is big.

Summary of the invention

The present invention solves existing algorithm poor real and the undesirable problem of reinforced effects, propose A kind of self adaptation Gamma theoretical based on Retinex strengthens algorithm, utilizes Retinex theory to divide From luminance component and the reflecting component of image, luminance component is carried out the correction of self adaptation Gamma, Reflecting component is finally utilized to recover details and the color of image.The results show, carried algorithm Having image definition higher, reinforced effects is the most natural, the ageing advantage such as more preferably.

This method is divided into several step:

Step one: acquisition observed image I (x, y)；

Step 2: the image observed is decomposed, picture breakdown is become reflecting component and Luminance component；

Retinex theory thinks that image is made up of luminance component and reflecting component, it is assumed that RGB Image I (x, y) each passage has identical brightness, it may be assumed that

I^c(x, y)=R^c(x,y) ·L(x,y),c∈{r,g,b} (1)

R in formula^c(x, y) represents the reflecting component of each passage, and (x y) represents that the brightness of image divides to L Amount, generally, using the maximum of RGB channel as eye-observation illumination V (x, y),

$V(x,y)=\underset{c\∈\{r,g,b\}}{max}{I}^c(x,y)---\left(2\right)$

Theoretical according to Retinex, with 2D Gaussian filter G, (x, y) to illumination image V, (x y) is carried out Convolution can obtain image luminance component L (x, y),

L (x, y)=G (x, y) * V (x, y) (3)

Thus, reflecting component R^c(x, y) can be separated,

R^c(x, y)=I^c(x,y)/L(x,y),c∈{r,g,b} (4)

Reflecting component mainly contains the radio-frequency component of image, including edge and details,

Step 3: after obtaining luminance component image, it is carried out the correction of self adaptation Gamma；

L_en(x, y)=L (x, y)^γ(x,y) (5)

$\mathrm{\γ}\left(l\right)=1-\underset{v=0}{\overset{l}{\Σ}}\[P_{\mathrm{\ω}}\left(v\right)/s_p\]---\left(6\right)$

$s_p=\underset{l=0}{\overset{l_{max}}{\Σ}}{P}_{\mathrm{\ω}}\left(l\right)---\left(7\right)$

L_en(x, y) represents enhanced luminance component, γ (x, y) represents Gamma correction coefficient matrix, P ω (l) is the weights distribution function corresponding to each brightness value:

$P_{\mathrm{\ω}}\left(l\right)=\frac{P\left(l\right)-p_{\min }}{p_{max}-p_{\min }}---\left(8\right)$

In formula, P (l) is the probability density function of luminance component, p_maxFor the maximum of P (l), p_minFor the minima of P (l), P (l) can be tried to achieve by following formula:

P (l)=n_l/n_p (9)

In formula, n_lPixel count contained by corresponding brightness, n_pThe pixel comprised for luminance component is total Number；

Step 4: merge L_en(x, y) and R^c(x y) i.e. can get final enhancing image I_en(x, y),

$I_{en}^c(x,y)=R^c(x,y)\·L_{en}(x,y),c\∈\{r,g,b\}---\left(10\right).$

Further, high-performance image signal processor ISP, employ the process described above.

Accompanying drawing explanation

Fig. 1 is the image enchancing method adjusted based on illumination；

Fig. 2 is the enhancing result of image Boy；

Fig. 3 is the enhancing result of image Museum；

Fig. 4 is the enhancing result of image Cockpit；

Fig. 5 is the enhancing result of image Girl；

Fig. 6 is the enhancing result of image Hall；

Fig. 7 is the enhancing result of image Building；

Fig. 8 is application image Enhancement Method in ISP chip.

Technical scheme

As it is shown in figure 1, Fig. 1 gives the algorithm for image enhancement flow chart adjusted based on illumination, Proposed algorithm for image enhancement will be passed through carry out for several different types of degraded image herein Process, and contrast from different algorithms respectively, verify the performance of carried algorithm herein with Versatility.

The algorithm for image enhancement adjusted based on illumination is divided into several step:

Step one: acquisition observed image I (x, y)；

Step 2: the image observed is decomposed, picture breakdown is become reflecting component and Luminance component；

Retinex theory thinks that image is made up of luminance component and reflecting component.Assume RGB Image I (x, y) each passage has identical brightness, it may be assumed that

I^c(x, y)=R^c(x,y)·L(x,y),c∈{r,g,b} (1)

R in formula^c(x, y) represents the reflecting component of each passage, and (x y) represents that the brightness of image divides to L Amount.Generally, using the maximum of RGB channel as eye-observation illumination V (x, y).

$V(x,y)=\underset{c\∈\{r,g,b\}}{max}{I}^c(x,y)---\left(2\right)$

Theoretical according to Retinex, with 2D Gaussian filter G, (x, y) to illumination image V, (x y) is carried out Convolution can obtain image luminance component L (x, y).

L (x, y)=G (x, y) * V (x, y) (3)

Thus, reflecting component R^c(x y) can be separated.

R^c(x, y)=I^c(x,y)/L(x,y),c∈{r,g,b} (4)

Reflecting component mainly contains the radio-frequency component of image, including edge and details.

Step 3: luminance component is carried out Gamma conversion；

After obtaining luminance component image, it is carried out the correction of self adaptation Gamma.

L_en(x, y)=L (x, y)^γ(x,y) (5)

$\mathrm{\γ}\left(l\right)=1-\underset{v=0}{\overset{l}{\Σ}}\[P_{\mathrm{\ω}}\left(v\right)/s_p\]---\left(6\right)$

$s_p=\underset{l=0}{\overset{l_{max}}{\Σ}}{P}_{\mathrm{\ω}}\left(l\right)---\left(7\right)$

L_en(x, y) represents enhanced luminance component, γ (x, y) represents Gamma correction coefficient matrix, P_ωL () is the weights distribution function corresponding to each brightness value:

$P_{\mathrm{\ω}}\left(l\right)=\frac{P\left(l\right)-p_{\min }}{p_{\max }-p_{\min }}---\left(8\right)$

In formula, P (l) is the probability density function of luminance component, p_maxFor the maximum of P (l), p_minFor the minima of P (l), P (l) can be tried to achieve by following formula:

P (l)=n_l/n_p (9)

In formula, n_lPixel count contained by corresponding brightness, n_pThe pixel comprised for luminance component is total Number.

Step 4: by the luminance component synthesis after reflecting component and conversion；

Merge L_en(x, y) and R^c(x y) i.e. can get final enhancing image I_en(x,y)。

$I_{en}^c(x,y)=R^c(x,y)\·L_{en}(x,y),c\∈\{r,g,b\}---\left(10\right)$

Step 5: form the image strengthened.

The image quality evaluation strengthened:

For evaluating the performance of algorithm herein, the Retinex inhomogeneous illumination image choosing advanced person increases Strong algorithms, carries out subjective assessment, objective evaluation and ageing comparison.

1. subjective assessment

Fig. 2～7 is the inhomogeneous illumination processing result image of different scene.Fig. 2 artwork is fine day family Outward, the face of boy is in shade, face's low visibility.Fig. 3 artwork is that indoor weak light shines Image, the scene scenery after showcase is unintelligible.Fig. 4 artwork has glass reflecting, aircraft cabin Interior things low visibility.In Fig. 5 artwork, owing to illumination is blocked, right side seems the darkest. Fig. 6 artwork is indoor inhomogeneous illumination image.Fig. 7 is the result of outdoor cloudy day building photo, In artwork red block, the result of topography is listed in the lower section strengthening image of correspondence.

Document 1 (Zhang Shangwei, Zeng Ping, Luo Xuemei, etc. there is details and compensate and color recovery Multiple dimensioned Retinex tone-mapping algorithm [J]. XI AN JIAOTONG UNIVERSITY Subject Index, 2012,46 (4): 32-37) the result reinforced effects of algorithm is obvious not, the dynamic range of image Compressed.Document 2 (Wang S, Zheng J, Hu H, et al.Naturalness preserved enhancement algorithm for non-uniform illumination images[J].IEEE Transactions on Image Processing,2013,22 (9): 3,538 3548.) result of algorithm makes image at the definition of global and local It is largely increased, but the dynamic range of image is also pressed.Inventive algorithm can not only be notable Improve image overall and the definition of regional area, and make image have good dynamic range With subjective natural sense.

2. objective evaluation

In order to process image is carried out objective appraisal, introduce EBCM (edge based herein Contrast measure), VE (visible edges), NIQE (naturalness image Quality evaluator) etc. evaluation index.Wherein EBCM is for the contrast of evaluation image, This parameter values is the biggest, then the contrast of image is the highest.VE be used for evaluating enhancing image relative to The ratio that the visibility of artwork improves, its value is the biggest, then in image, things visibility is the highest.NIQE For the naturality of evaluation image, its value is the least, and the naturality of explanatory diagram picture is the best.

Table 1～3 lists the evaluation index result of image in Fig. 2～7 respectively.As can be seen from the table, After each algorithm process, image is relative to artwork at contrast, visibility, three aspects of naturality all It is significantly improved.Methods herein obviously has more than algorithm in document [1] and document [2] Good performance: in terms of contrast, the most averagely improve 49.7% and 10.8%；At visibility Aspect, the most averagely improves 102.4% and 38.5%；In terms of naturality, it is respectively increased 31% and 16.4%.

The EBCM result of each algorithm of table 1

The VE result of each algorithm of table 2

The NIQE result of each algorithm of table 3

The most ageing comparison

In order to compare the ageing of each algorithm, table 4 shows that size is by various algorithm 2000 × 1312 images average time-consuming, these data are at hardware parameter 3.3GHz CPU, 4GB The test result of Matlab 2014 is utilized on RAM computer.Obviously, the timeliness of institute's extracting method herein Property best, 11% and document [7] that only document [6] is time-consuming time-consuming 2%.

Process 2000 × 1312 image of each algorithm of table 4 average time-consuming

It can be observed that self adaptation Gamma proposed by the invention strengthens calculation from experimental result One width observation RGB image can be separated into luminance picture and albedo image by method.Due to only pin Brightness to image processes, and does not introduce colouring information.The most isolated luminance picture Being gray level image with albedo image, wherein luminance picture has reacted surrounding well Monochrome information, has the character of space smoothing simultaneously.Albedo image then remains image self Edge and detailed information, these features have been well demonstrated that the new target that this title of the song is proposed Function, the i.e. reasonability of formula (10).And under rgb space, luminance picture and reflectance Image, in addition to having the characteristic of the result of calculation under HSV space, also show colouring information. As can be seen from the results, by Gamma rectification, illumination is adjusted in conjunction with reflectance institute The enhancing image obtained, except strengthening detail section, while promoting the brightness of dark area, Subjective vision effect also has to be improved definitely.

The algorithm that the present invention proposes can be applied at image and videos such as TV, mobile phone, video cameras On relevant common apparatus.May apply to specific occasion (medical science, military affairs, public safety etc.) Figure strengthen.This section introduces algorithm application on high-definition camera.

Generally, in order to improve graphics process performance, high-definition camera be all integrated with image letter Number processor ISP (Image Singal Processor), for imageing sensor (CCD Or CMOS) signal that exports carries out later stage process (such as: the accumulation of 3D noise reduction, frame, high light press down System), and then the quality of the exported image of enhancing, meet the demand of application-specific.Can be by this The algorithm integration of literary composition is in ISP chip, it is achieved to the image enhaucament under the conditions of inhomogeneous illumination. System forms as shown in Figure 7.

The present invention proposes a kind of self adaptation Gamma Enhancement Method theoretical based on Retinex, utilizes The luminance component of Retinex theory separate picture and reflecting component, carry out adaptive to luminance component Answer Gamma to correct, finally utilize reflecting component to recover details and the color of image.Algorithm herein Solve existing Retinex algorithm poor in timeliness and the undesirable problem for the treatment of effect, this algorithm The inhomogeneous illumination image processed has optimal contrast, visibility, naturality and ageing. Herein algorithm to hardware without particular/special requirement, may migrate to various TV, mobile phone, video camera or its He has in the electronic product of image displaying function.

Claims (2)

1. the image enchancing method adjusted based on illumination, it is characterised in that this method is divided into several Step:

Step one: acquisition observed image I (x, y)；

Step 2: the image observed is decomposed, picture breakdown is become reflecting component and Luminance component；

Retinex theory thinks that image is made up of luminance component and reflecting component, it is assumed that RGB Image I (x, y) each passage has identical brightness, it may be assumed that

I^c(x, y)=R^c(x,y)·L(x,y),c∈{r,g,b} (1)

R in formula^c(x, y) represents the reflecting component of each passage, and (x y) represents that the brightness of image divides to L Amount, generally, using the maximum of RGB channel as eye-observation illumination V (x, y),

$V(x,y)=\underset{c\∈\{r,g,b\}}{max}{I}^c(x,y)---\left(2\right)$

Theoretical according to Retinex, with 2D Gaussian filter G, (x, y) to illumination image V, (x y) is carried out Convolution can obtain image luminance component L (x, y),

L (x, y)=G (x, y) * V (x, y) (3)

Thus, reflecting component R^c(x, y) can be separated,

R^c(x, y)=I^c(x,y)/L(x,y),c∈{r,g,b} (4)

Reflecting component mainly contains the radio-frequency component of image, including edge and details,

Step 3: after obtaining luminance component image, it is carried out the correction of self adaptation Gamma；

L_en(x, y)=L (x, y)^γ(x,y) (5)

$\mathrm{\γ}\left(l\right)=1-\underset{v=0}{\overset{l}{\Σ}}\[P_{\mathrm{\ω}}\left(v\right)/s_p\]---\left(6\right)$

$s_p=\underset{l=0}{\overset{l_{max}}{\Σ}}{P}_{\mathrm{\ω}}\left(l\right)---\left(7\right)$

L_en(x, y) represents enhanced luminance component, γ (x, y) represents Gamma correction coefficient matrix, P ω (l) is the weights distribution function corresponding to each brightness value:

$P_{\mathrm{\ω}}\left(l\right)=\frac{P\left(l\right)-p_{\min }}{p_{max}-p_{\min }}---\left(8\right)$

In formula, P (l) is the probability density function of luminance component, p_maxFor the maximum of P (l), p_minFor the minima of P (l), P (l) can be tried to achieve by following formula:

P (l)=n_l/n_p (9)

In formula, n_lPixel count contained by corresponding brightness, n_pThe pixel comprised for luminance component is total Number；

Step 4: merge L_en(x, y) and R^c(x y) i.e. can get final enhancing image I_en(x, y), $I_{en}^c(x,y)=R^c(x,y)\·L_{en}(x,y),c\∈\{r,g,b\}---\left(10\right).$

2. a high-performance image signal processor ISP, it is characterised in that employ claim 1 Described method.