KR102377318B1 - High Dynamic Range technique for light exposure compensation and image quality improvement - Google Patents

Abstract

According to the present invention, an HDR technique for correcting light exposure and increasing image quality comprises the steps of: separating a brightness or luminance channel from an input image, and calculating and applying a gamma correction coefficient for correcting the entire image with the global average value (μ) and the median value (m) of the brightness or luminance value of the image, which are basic parameters extracted from an image global distribution; extracting the class ratio (P_1, P_2) and the class average value (μ_1), which are complementary parameters, from information on classes obtained by binarizing the image into bright and dark areas by applying a binarization image segmentation technique; and determining fusion weights by using the extracted complementary parameters. Therefore, the HDR technique can simply and effectively convert low-contrast and inappropriately exposed images (overexposure, underexposure, and backlight) and low-quality low dynamic range (LDR) images into high-definition HDR images.

Description

High Dynamic Range technique for light exposure compensation and image quality improvement

The present invention relates to a high dynamic range (HDR) technique, and more particularly, to an HDR technique for correcting light exposure and improving image quality.

Recently, as the high-end smartphone built-in camera, the spread of low-cost, high-performance DSLR (digital single lens reflex) cameras, and the life-logging era to record personal daily life became common, photography took up most of the recording media.

However, since ordinary people do not have expertise in photographic technology, it can be said that photographs are frequently obtained with inappropriate exposure values in environments such as backlight, sunrise and sunset. In order to solve this problem, camera manufacturers are equipped with various light metering modes and image quality improvement functions, but they cannot reflect all environments, and user errors account for a large proportion.

Therefore, in order to improve these images, various methods using recently high dynamic range (HDR) using exposure fusion, Retinex, tone mapping, gamma correction, and histogram manipulation, etc. This is being actively studied.

In particular, as for the high-definition imaging technique (HDR technique) through exposure compensation, the exposure fusion method and the tone mapping method are representative. However, since the exposure fusion method requires several photos of different exposure values, it cannot be applied to already captured images or existing systems, and the tone mapping method is difficult to implement and has poor performance because it uses a complex conversion formula.

KR 10-1717733 B

The present invention has been proposed to solve the above technical problem, and after predicting the optimal gamma compensation coefficient using only the brightness or luminance channel statistical information (the overall average value and the median value) of a single input image, improper light exposure It provides HDR techniques for photo-exposure correction and image quality improvement that improve (under-exposure and over-exposure) low dynamic range images into HDR images.

According to an embodiment of the present invention to solve the above problem, the brightness or luminance channel is separated from the input image, and then the global average value (μ) of the brightness or luminance values of the image, which is a basic parameter extracted from the global distribution of the image, and The step of calculating and applying a gamma correction coefficient that corrects the entire image with the median value (m), and the class ratio ( There is an HDR technique for light exposure correction and image quality improvement, which includes extracting P ₁ , P ₂ ) and class average values (μ ₁ , μ ₂ ), and determining a fusion weight using the extracted complementary parameters. provided - P ₁ : Class ratio of bright area, P ₂ : Class ratio of dark area, μ ₁ : Class average value of bright area, μ ₂ : Class average value of dark area -

delete

delete

delete

delete

delete

delete

delete

delete

delete

delete

delete

delete

delete

delete

- 여기서 α, β는 스케일링 상수, γ_max1는 γ₁ 의 최대값, γ_max2는 γ₂ 의 최대값, γ_min1는 γ₁ 의 최소값, γ_min2는 γ₂ 의 최소값임 - 으로 정의되는 것을 특징으로 한다.
또한, 본 발명에서 감마 보정 계수(γ₁, γ₂)로 보정된 각 영상(I(i, j))을 하기 식과 같이,

- 여기서 ω는 각 보정 영상에 대한 융합 가중치이고, 밝은 화소가 보정된 영상(I_b), 어두운 화소가 보정된 영상(I_d) 및 입력 영상(I_i)임 - 가중치 융합하여 결과 영상 R(i, j)를 획득하는 것을 특징으로 한다.
또한, 본 발명에서 클래스 평균값의 절대적 위치(class mean position)에 따른 가중치 함수(ω_cmp)는 μ₁이 0에 가깝거나 μ₂가 255에 가까울수록 큰 값을 가지도록 하기 식과 같이,

- 여기에서 σ는 클래스 평균값의 절대적 위치에 따른 가중치 결정 함수인 가우시안 함수의 폭을 조정하는 표준편차임 - 으로 처리되는 것을 특징으로 한다.
또한, 본 발명에서 어두운 클래스와 밝은 클래스의 평균값 위치(class mean position)에 따른 융합 가중치를 계산할 때, Gaussian 함수, 선형 또는 비선형함수 중 어느 하나를 사용하는 것을 특징으로 한다.
또한, 본 발명에서 클래스 비율(P₁, P₂)과 클래스 평균값(μ₁, μ₂)이 적용된 융합 가중치는,

- 여기에서 T는 밝은 영역과 어두운 영역으로 이진 분할하는 문턱치이고, ω_b 는 밝은 영역의 융합 가중치이고, ω_d 는 어두운 영역의 융합 가중치임 - 으로 정의되는 것을 특징으로 한다.
또한, 본 발명에서 기본 매개변수(primary parameter)인 영상의 전역 평균값(μ)과 중간값(m)으로 전체적인 감마 보정 계수를 계산해서 적용한 후, 이진화된 영상의 보완 매개변수인 클래스 비율(P₁, P₂)과 클래스 평균값(μ₁, μ₂)으로 융합 가중치를 결정하는 것을 특징으로 한다.In addition, the binarization image segmentation technique in the present invention is characterized in that any one of Ostu, Iterative thresholding, Kapur, and Adaptive thresholding is used.
In addition, in the present invention, from d, which is the difference between the global average value (μ) and the median value (m) of the image, a gamma correction coefficient - a gamma correction coefficient for darkening a bright area (γ ₁ ), and a gamma for brightening a dark area It is characterized by using a sigmoid function that converges on both +∞ and -∞ to calculate the correction coefficient (γ ₂ ) - .
In addition, in the present invention, when calculating the gamma correction coefficient from the difference d between the global average value (μ) and the median value (m) of the image, either a linear or non-linear function is used.
In addition, in the present invention, a gamma correction coefficient (γ ₁ ) for darkly correcting a bright area and a gamma correction coefficient (γ ₂ ) for brightly compensating a dark area are as follows:

- where α and β are scaling constants, γ _max1 is the maximum value of γ ₁ , γ _max2 is the maximum value of γ ₂ , γ _min1 is the minimum value of γ ₁ , γ _min2 is the minimum value of γ ₂ - do.
In addition, in the present invention, each image (I(i, j)) corrected with the gamma correction coefficients (γ ₁ , γ ₂ ) is expressed by the following equation,

- where ω is the fusion weight for each corrected image, the bright pixel corrected image (I _b ), the dark pixel corrected image (I _d ), and the input image (I _i ) - weighted fusion resulting image R( It is characterized in that i, j) is obtained.
In addition, in the present invention, the weight function (ω _cmp ) according to the absolute position of the class mean value has a larger value as μ ₁ is closer to 0 or μ ₂ is closer to 255, as in the following formula,

- Here, σ is the standard deviation that adjusts the width of the Gaussian function, which is a weight determining function according to the absolute position of the class average value.
In addition, when calculating the fusion weight according to the class mean position of the dark class and the light class in the present invention, any one of a Gaussian function, a linear function, or a non-linear function is used.
In addition, in the present invention, the fusion weight to which the class ratio (P ₁ , P ₂ ) and the class average value (μ ₁ , μ ₂ ) is applied is,

- where T is the threshold for binary partitioning into bright and dark regions, ω _b is the fusion weight of the bright region, and ω _d is the fusion weight of the dark region.
In addition, in the present invention, after calculating and applying the overall gamma correction coefficient using the global average (μ) and median (m) of the image, which are the primary parameters, the class ratio (P ₁ ), which is a complementary parameter of the binarized image , P ₂ ) and class average values (μ ₁ , μ ₂ ) to determine the fusion weight.

delete

delete

delete

delete

The HDR technique for light exposure correction and image quality improvement according to an embodiment of the present invention determines the coefficient of a gamma correction function that can correct an image most simply by using statistical features (overall average and median value) of the image. Compensates for dark and bright areas.

Then, in order to properly integrate each compensation image, the ratio of each class binary divided into bright and dark regions and the class average value are used to determine the integration weight to obtain the final HDR image in which all regions are improved.

Therefore, the present invention is a simple and effective method for correcting low contrast, underexposure images (overexposure, underexposure and backlight) and converting a low quality low dynamic range (LDR) image into a high quality HDR image.

1 is a view showing exemplary images with different exposure values;
2 is a diagram illustrating various global and local tone mapping operators (TMOs);
3 is a diagram showing a general gamma correction function;
4 is a block diagram (algorithm) of an HDR technique for light exposure correction and image quality improvement proposed in an embodiment of the present invention.
5 is a diagram showing an example of the Ostu threshold <(a) distribution (b) variance between classes>
6 is an example diagram of basic parameters for inappropriate light exposure (under-exposure and over-exposure) <(a) under-exposure (b) over-exposure>
7 is a view showing a part of an image for statistical analysis
Fig. 8 is a diagram showing the supplementary parameters after binary division <(a) In the case of Fig. 6(a) (b) In the case of Fig. 6(b)>
9 is a diagram showing the gamma correction range of the proposed algorithm.
10 is a diagram illustrating a proposed gamma calculation function;
11 is a diagram illustrating a weight function for a class average value position;
12 is a view showing an experimental image <(a) Image 1 (b) Image 2 (c) Image 3 (d) Image 4>
13 is a diagram showing the results for image 1 <(a) Method 1 (b) Method 2 (c) Proposal method>
14 is a diagram showing the results for image 2 <(a) Method 1 (b) Method 2 (c) Proposal method>
15 is a diagram showing the results for image 3 <(a) Method 1 (b) Method 2 (c) Proposal method>
16 is a view showing the results for image 4 <(a) Method 1 (b) Method 2 (c) Proposal method>

Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings in order to describe in detail enough that a person of ordinary skill in the art to which the present invention pertains can easily implement the technical idea of the present invention.

The present invention proposes an HDR algorithm for correcting the light exposure of a photometric failure image obtained with an improper exposure value. Existing tone mapping and gamma correction techniques produce good results when the brightness distribution is symmetric with respect to the average, because the image is improved on average by performing correction with function mapping with the same input and output ranges. Otherwise, the compensation was not sufficient for one side of the bright part.

In addition, the method using the histogram transformation function forcibly shifts the image distribution, so it works well for an image with a lot of change, but unintended noise increases in a flat area.

Therefore, in the proposed method, in order to apply sufficient correction to both dark and bright parts, the gamma coefficient for overall correction is calculated with the basic parameters extracted from the global distribution. Then, complementary parameters are used in the information segmented by the binarization image segmentation technique and reflected in the fusion weight. As a result, good correction results can be obtained not only in underexposed and overexposed images, but also in images in which the brightness distribution of the image is symmetrical or asymmetrical.

exposure fusion

In a camera that hardware exposure fusion is applied, multiple photos are taken at high speed by bracketing the EV (exposure value) value. It is used to generate and use correction images of several stages.

1 is a diagram illustrating exemplary images having different exposure values.

Most of the exposure fusion is performed in the luminance or brightness channel, and as in FIG. 1, multiple images generated by different weights are fused for each pixel, and the most representative technique is the method proposed by Tom Mertens et al. A weight map W for fusion is generated from the images of the dog as in Equation (1).

Here, (i, j) is the position of the image, and k is the weight map of the k-th image. C is contrast, S is saturation, E is well-exposedness, and ω is the reflection coefficient of each item. After normalizing the provisional decay map obtained in this way, it is multiplied by each input image to obtain a final fusion image (R _ij ) as shown in Equation (2).

은 정규화된 가중치를, I는 입력 영상을 각각 나타낸다. 그러므로 이 과정에서 가장 중요한 부분은 대조, 채도 및 노출 정도를 척도로 수치화하는 것과 각 척도에 화소별 가중치를 결정하는 것이며, 이 부분이 해결해야 할 어려운 문제라고 할 수 있다. 또한, 대부분의 노출 융합법은 식 (2)에서처럼 가중치 맵을 화소별로 다르게 적용하기 때문에 지역적 특성을 반영할 수 있다는 장점이 있지만, 가중치 맵의 데이터가 영상의 개수(N)만큼 필요하고 자료형이 실수형이므로 [4×N×영상크기] 공간이 요구된다. 그 결과, 연산량과 자료공간을 많이 소모하는 단점이 있다.here

denotes a normalized weight, and I denotes an input image. Therefore, the most important part in this process is to quantify contrast, saturation, and exposure as a scale, and to determine a weight for each pixel on each scale, which is a difficult problem to be solved. In addition, most exposure fusion methods have the advantage of reflecting regional characteristics because the weight map is applied differently for each pixel as in Equation (2), but the weight map data requires as much as the number of images (N) and the data type is [4×N×image size] space is required. As a result, there is a disadvantage of consuming a lot of computation and data space.

tone mapping

Tone mapping is a technique for extending the dynamic range of an image to a wide range by bit extension, and then reducing the brightness tone so that it can be appropriately expressed on a target display device.

2 is a diagram illustrating various global and local tone mapping operators (TMOs).

As shown in FIG. 2, tone mapping is expressed as a function that maps an output value to an input luminance or color value within the display gamut (color space/gamut) limit, and is largely divided into global and local curves, The local mapping function performs a fine transformation to the left and right of the center value of a specific region. The simplest and most representative technique is the Reinhard method, and TMO (tone mapping operator) is

It is expressed as in Equation (3), where Y _i is the input luminance and Y _o is the mapped output luminance value. However, in the case of the maximum input (Y _i = 1), the output is 1/2 and the dynamic range cannot be fully utilized, so the Extended Reinhard method using the entire dynamic range as in Equation (4) reflecting the maximum luminance value of white. is mainly used

Y _white : luminance value when white

After that, filmic techniques that mimic the properties of real films have been proposed, and Uncharted 2, also known as 'Hable Filmic' designed by John Hable, is a representative method. The filmic technique has a characteristic that the black region is saturated because the part corresponding to ‘toe’ of the conversion function is lower than the Reinhard curve. The most popular filmic technique is ACES (academy color encoding system), which rapidly converts color values through matrix transformation, and is widely used these days in the real-time game engine unreal engine 4 or computer graphics environment. The appearance of the global tone mapping curve ranges from a log function to a sigmoid function, and a local method of adding fine adjustments for a detailed area is becoming common, but the amount of computation continues to increase as local control and average and maximum brightness are taken into account.

Gamma Correction

Gamma correction can be thought of as a global tone mapping function from a large perspective, but since it is a technology for linearly compensating for non-linear physical properties of a camera or display, it is a completely different field in concept and application. The corrected output value I _C is expressed as the power of the reciprocal of the gamma coefficient γ with respect to the input IS , where _L is the maximum output value.

3 is a diagram illustrating a general gamma correction function.

The general input/output gamma characteristic of the imaging equipment is as shown in FIG. 3 , and in general, a display such as a monitor has γ=0.45, and the sensor of a camera has a reciprocal of about 2.2. Although this gamma correction is intended to compensate for the characteristics of imaging equipment, the neurobiological characteristics of human vision are also non-linear, so it is widely applied to complexly correct the characteristics of imaging devices and human perception.

- Suggested method

The technique proposed in the present invention uses gamma correction, which is a relatively simple correction technique than tone mapping, and adjusts adaptive gamma coefficients fixedly applied to the entire image according to image characteristics and fusion weights to reflect detailed characteristics. Exposure compensation method is suggested.

4 is a block diagram of an HDR technique (algorithm) for light exposure correction and image quality improvement proposed in an embodiment of the present invention.

4 shows a block diagram of the proposed technique. First, a brightness or luminance channel is separated from an input image, and then a basic parameter to be used for calculating a gamma correction coefficient is calculated from the global distribution of the image. Then, the image is binarized into a bright region and a dark region to extract complementary parameters from the information of each region class. A fusion weight is determined using this information, and an improved result image is finally obtained.

- Threshold technique

The threshold technique is the simplest technique for dividing a binary image, and there are iterative thresholding, Otsu, Kapur, and adaptive thresholding techniques. Iterative thresholding determines the median value of the averages (μ ₁ , μ ₂ ) of each bisected class as a threshold, and ends the calculation when the threshold does not change while performing iterative calculations. Because the Kapur method finds a threshold that maximizes the entropy of the histogram, a more complex formula is required for entropy calculation than other methods.

Although the Otsu technique is applied in this embodiment, various threshold determination (binarization segmentation) techniques such as iterative thresholding, Kapur, and adaptive thresholding may be applied.

Since Otsu determines the value that maximizes the between class variance as the threshold as in Equation (6), it does not require repeated calculations and the optimum value can be found with a relatively simple operation.

where T is the threshold for binary partitioning into bright and dark regions, μ is global mean - global mean (μ) -,
P ₁ and P ₂ are the interclass ratios of pixels bisected by T, P ₁ is the class ratio of the bright area, P ₂ is the class ratio of the dark area, μ ₁ is the class average value of the bright area, μ ₂ is the class ratio of the dark area is the class average of -

5 is a diagram showing an example of the Ostu threshold <(a) distribution (b) variance between classes>.

5 is an example showing the distribution of one image and the variance between classes. The variance between classes in FIG. 5(b) is always monotonically increased to the maximum value and then monotonically decreased like a Gaussian function. Since it has a shape, it is easy to find the maximum value.

- Statistical parameters

Looking at the global average value (μ) of the image, it can be seen that the image is generally bright and dark. function can be determined. For example, if the image is bright and evenly distributed throughout, there will be little difference between the average value and the median value. However, if the difference between the average value and the median value is large, it can be determined whether the exposure is insufficient or excessive depending on the position of the median value. It is used to calculate the coefficient.

6 is an exemplary diagram of basic parameters for inappropriate light exposure (underexposure and overexposure) <(a) underexposure (b) overexposure>, and FIG. 7 is a view showing a part of an image for statistical analysis.

Referring to FIG. 6 , it can be seen that the median value is higher than the average in FIG. 6( a ) because the image is underexposed, and the median value is higher than the average in FIG. 6 ( b ) because the image is overexposed. This could be confirmed from the results of analysis of 50 images used in the experiment as shown in FIG. 7, and the difference (d) between the average value (μ) and the median value (m) was

It is calculated by , and it can be seen that the larger the difference (d), the larger the correction coefficient is required. In most cases, if the correction is performed, the image quality is improved, especially when the difference is 10 or more, excessive correction is required.

Table 1 is a table showing specific examples of analyzed images. Table 1 shows examples of cases requiring large correction. In an image with an average value of around 128 but biased toward bright or dark areas, the bright and dark areas must be specially considered.

FIG. 8 is a diagram showing the supplementary parameters after binary segmentation <(a) the case of FIG. 6(a) (b) the case of FIG. 6(b)>.

FIG. 8 shows information of each class obtained from the binarization result of FIG. 6 . 8(a) shows that the difference between the threshold and the global average is small, so the ratio of each class is similar at 52% (P ₁ ) and 48% (P ₂ ), but it is biased toward a very low brightness value, so the class average value ( μ ₁ ) is very low. Therefore, in this case, it means that dark areas should be brightly corrected and reflected. In Fig. 8(b), it can be seen that the global average and median are located on the very high side, so it is a bright image. It can be seen that this excessively bright part should be corrected to the dark side. Therefore, in the proposed method, the ratio and average of each class, the class ratio in the bright region (P ₁ ), the class ratio in the dark region (P ₂ ) _, the class average value in the bright region (μ ₁ ), and the class average value in the dark region (μ ₂ ) ) is used as a complementary parameter to determine the fusion weight.

<Table 1>

- Gamma coefficient and weight calculation

As shown in FIG. 3 , a commonly used gamma correction coefficient is 2.2 and its reciprocal of 0.45 is universal, and is a basic compensation value for optical devices and human visual characteristics. However, this value cannot be said to be an appropriate correction value for an image in which the distribution of the input image is concentrated in the dark or bright side due to improper exposure. The part should be compensated largely, and vice versa, the bright part should be corrected by making it smaller than γ < 0.45.

9 is a diagram illustrating a gamma correction range of the proposed algorithm, and FIG. 10 is a diagram illustrating a proposed gamma calculation function.

Therefore, in the proposed technique, as shown in FIG. 9 , the range of the gamma correction coefficient (γ ₁ ) for darkening a bright region is (0.125, 0.45], and the gamma correction coefficient (γ ₂ ) for brightly compensating a dark region is [2.2, 8) set a wide range.

In order to calculate the gamma correction coefficients γ ₁ and γ ₂ from d, which is the difference between the global average value (μ) and the median value (m) of the image, which are the basic parameters of the proposed method, a sigmoid function that converges on both +∞ and -∞ The gamma coefficient was calculated using as in FIG. 10 . For reference, in addition to the sigmoid function, either a linear or non-linear function may be used.

Since γ ₁ is a part decreasing from 0.45 to 0.125, it is mapped to a part where d of the sigmoid function is negative, and γ ₂ is mapped to a part where d is positive to increase from 2.2 to 8. Although the maximum, minimum, and scale of each section are different, the shape is the same, so the two sections are marked together in FIG. 10 based on d=0. Since d=0 is set as the default value in each section, γ _max1 is the maximum value of γ ₁ , 0.45, and γ _min2 is the minimum value of γ ₂ , 2.2. As a result, γ ₁ and γ ₂ are

이라고 했을 때, 이 값이 각 구간의 sigmoid 함수 폭에 절반이 되도록 설정하였다. 영상 분석에 사용된 50개의 영상에서 획득한

=16.5가 나왔기 때문에 α=0.08과 β=1.45로 결정하였으며, 편의상 식 (8)과 (9) 중 하나를 사용해도 무관하다. 각 경우에 대해서 γ₁과γ₂ 을 역수로 사용한 이유는 대칭적인 보상을 해서 비대칭적 보상을 통한 급격한 평균 밝기의 변화를 막기 위함이다. 그 결과 d값에 따라 두 개의 감마 계수에 의한 보정 영상이 획득된다.is expressed as Here, α and β are scaling constants to compensate for the maximum-minimum difference in each section. The average of the absolute values of d of the image used in FIG.

, this value was set to be half the width of the sigmoid function of each section. Acquired from 50 images used for image analysis

= 16.5, so α = 0.08 and β = 1.45 were determined, and for convenience, either equation (8) or (9) can be used. The reason why γ ₁ and γ ₂ are used as reciprocals in each case is to prevent abrupt change in average brightness through asymmetric compensation by performing symmetric compensation. As a result, a corrected image using two gamma coefficients is obtained according to the d value.

Each image (I(i, j)) obtained in this way is weighted fused as follows to obtain a resulting image R(i, j).

ω is a fusion weight for each corrected image, and the subscripts indicate an image (I _b ) in which bright pixels are corrected, an image (I _d ) in which dark pixels are corrected, and an input image (I _i ), respectively, in order. In order to reflect the detailed characteristics of the image, consider the case of FIG. 8(b). Because this image is overexposed and the proportion _of bright pixels is too high (87%), the weight of the image I _d corrected with a dark image needs to be increased. It should be reflected in the weight (ω _d ) and, on the contrary, the brightly corrected I _b image should act on the fusion weight (ω _b ) of the bright region to weaken it to P ₁ =0.13, which is the ratio of class 1 (dark region). Only the two corrected images can be used, but the original image is basically reflected because the weight is too sensitive.

In addition, we consider a weighting function to consider the position of the class mean value as well as the proportion of the class. Referring to FIG. 8( a ), the ratio of each class is similar, but the average value of class 1 is very low. Therefore, since the correction may be insufficient if only the proportion of classes is considered, the fusion weight (ω _b ) of the bright region needs to be further increased so that it can be brightly corrected if it has a low average. Similarly, looking at the average of class 2 in FIG. 8(b), it is located on a very high side and is primarily weakened by the class ratio, but it is necessary to further reduce it further considering the position of the average. Therefore, the average position of each class has the effect of boosting the weight compensation, which can further improve the result.

11 is a diagram illustrating a weight function for a class average value position.

The weighting function according to the class mean position of the class mean value has a larger value as the class mean value (μ ₁ ) of the bright area is close to 0 or the class mean value (μ ₂ ) of the dark area is close to 255 ( 11) and a Gaussian function as in FIG. 11 were used. For reference, in addition to the Gaussian function, any one of a linear or non-linear function may be used.

Here, the standard deviation (σ) was set so that 6σ=128. The fusion weight considering the class ratio and the position of the class average value is the same as Equation (12), and is applied after being normalized so that the sum of the weights becomes 1. The standard deviation (σ) is the standard deviation that adjusts the width of the Gaussian function, which is a weight determining function according to the absolute position of the class average value.

To summarize the proposed technique, after calculating and applying the overall gamma correction coefficient using the global variable mean (μ) and median (m), which are the primary parameters, the class, which is the complementary parameter of the binarized image, is This is a method of fine-tuning the fusion weight with the ratios (P ₁ , P ₂ ) and class average values (μ ₁ , μ ₂ ). The class ratio in the bright area (P ₁ ) and the class ratio in the dark area (P ₂ ) are compensated by the other party's ratio, so that the class ratio in the bright area (μ ₁₎ ) and the class average value (μ ₂ ) of the dark region compensate for the excessively dark or bright image once more so that the class average value goes in the middle direction.

- where T is the threshold for binary partitioning into bright and dark regions, ω _b is the fusion weight of the bright region, and ω _d is the fusion weight of the dark region -

- Experimental results and consideration

12 is a diagram showing an experimental image <(a) Image 1 (b) Image 2 (c) Image 3 (d) Image 4>, Table 2 is a table showing experimental image parameters, Table 3 is a gamma coefficient and This is a table showing the fusion weights.

은 정규화된 가중치를 나타낸다.The images used in the experiment are shown in FIG. 12 , and images (images) 1 and 2 are underexposed images, and images (images) 3 and 4 are overexposed images. The information of each image (image) is shown in Table 2, the gray one is the weight function (ω _cmp ) when the weight operates, and the experimental coefficients calculated by the basic and supplementary parameters are specified in Table 3. . here

is the normalized weight.

Method 1, chosen to compare and evaluate the performance of the proposed method, uses fixed coefficient gamma correction and tone mapping, and uses weight map and boundary compensation, which is similar to the traditional HDR technique, T. Mertens.

Method 2 is a method using histogram adjustment and statistical deviation.

13 is a diagram showing the results for image 1 <(a) method 1 (b) method 2 (c) proposed method>, and FIG. 14 is a diagram showing the results for image 2 < (a) method 1 (b) Method 2 (c) Proposal method>, FIG. 15 is a diagram showing the results for image 3 <(a) Method 1 (b) Method 2 (c) Proposal method>, and FIG. 16 is a diagram showing the results for image 4 It is a drawing <(a) Method 1 (b) Method 2 (c) Proposal method>.

Table 4 is a table showing global parameters for the result image, FIGS. 13 to 16 show the result image, and the average and median values of each result are shown in Table 4. In the experimental results, method 1 showed overall performance improvement, but showed somewhat insufficient results in underexposure and overexposure situations. Method 2 has the characteristic of aligning the mean and median values in the center of the brightness distribution because the histogram is adjusted, so underexposure is converted into an excessively bright image and overexposure is converted into a dark image. On the other hand, the proposed method showed the best subjective image quality by correcting the underexposed and overexposed images at an appropriate ratio.

<Table 2>

<Table 3>

<Table 4>

- conclusion

In the present invention, an algorithm for correcting the light exposure of an image acquired with an inappropriate exposure value is proposed. Conventional global correction techniques, such as tone mapping and gamma correction, define a transform function between the input and output and perform correction to improve the overall image, so good results when the brightness distribution is symmetric around the global average However, in images with an asymmetric distribution skewed to one side of dark or light areas, insufficient compensation results were obtained. In addition, in the case of using the histogram transformation function, the image distribution is forcibly moved, so it works well for images with large changes. appear. Therefore, in the proposed method, in order to apply sufficient correction to both dark and bright parts, the gamma coefficient for overall correction is calculated with the basic parameters extracted from the global distribution. Then, the ratio of the two classes divided by applying the binarization image segmentation technique was reflected in the fusion weight to consider the detailed characteristics of the image. In addition, in order to improve the correction performance of underexposed and overexposed images, which are excessively focused on dark and bright areas, a technique for additionally reflecting the average value of each class in the fusion weight was proposed. As a result, good correction results were obtained not only in underexposed and overexposed images, but also in images in which the image brightness distribution was symmetrical or asymmetrical.

As such, those skilled in the art to which the present invention pertains will understand that the present invention may be embodied in other specific forms without changing the technical spirit or essential characteristics thereof. Therefore, it should be understood that the embodiments described above are illustrative in all respects and not restrictive. The scope of the present invention is indicated by the following claims rather than the above detailed description, and all changes or modifications derived from the meaning and scope of the claims and their equivalent concepts should be interpreted as being included in the scope of the present invention. do.

Claims (10)

After separating the brightness or luminance channel from the input image, a gamma correction coefficient that corrects the entire image with the global average (μ) and median (m) of the brightness or luminance values of the image, which is a basic parameter extracted from the global distribution of the image, is calculated. calculating and applying;
extracting the class ratio (P ₁ , P ₂ ) and class average value (μ ₁ , μ ₂ ), which are complementary parameters, from the information of classes in which an image is binarized into a bright area and a dark area by applying a binarization image segmentation technique; and
determining fusion weights using the extracted complementation parameters;
HDR technique for light exposure correction and image quality improvement, including.
P ₁ : Class ratio of bright areas
P ₂ : Class ratio of dark areas
μ ₁ : class average value of bright area
μ ₂ : average value of the class in the dark area According to claim 1,
The binarized image segmentation technique is
HDR technique for light exposure correction and image quality improvement, characterized in that any one of Ostu, Iterative thresholding, Kapur or Adaptive thresholding is used.
According to claim 1,
Gamma correction coefficient from d, which is the difference between the global average value (μ) and the median value (m) of the image - Gamma correction coefficient for darkening bright areas (γ ₁ ), and gamma correction coefficient for brightening dark areas (γ ₂ ) ) HDR technique for light exposure correction and image quality improvement, characterized in that it uses a sigmoid function that converges on both +∞ and -∞ to calculate -.
According to claim 1,
HDR technique for light exposure correction and image quality improvement, characterized in that either a linear or non-linear function is used when calculating the gamma correction coefficient from d, which is the difference between the global average value (μ) and the median value (m) of the image .
4. The method of claim 3,
A gamma correction coefficient (γ ₁ ) for darkly correcting a bright area and a gamma correction coefficient (γ ₂ ) for brightly compensating a dark area are as follows:

- where α, β are the scaling constants, γ _max1 is the maximum value of γ ₁ , γ _max2 is the maximum value of γ ₂ , γ _min1 is the minimum value of γ ₁ , γ _min2 is the minimum value of γ ₂ -
HDR technique for light exposure correction and image quality improvement, characterized in that it is defined as
6. The method of claim 5,
Each image (I(i, j)) corrected with the gamma correction coefficients (γ ₁ , γ ₂ ) is expressed by the following equation,

- where ω is the fusion weight for each corrected image, the bright pixel corrected image (I _b ), the dark pixel corrected image (I _d ) and the input image (I _i ) -
HDR technique for light exposure correction and image quality improvement, characterized in that the resultant image R(i, j) is obtained by weighted fusion.
7. The method of claim 6,
The weight function (ω _cmp ) according to the absolute position of the class mean value has a larger value as μ ₁ is closer to 0 or μ ₂ is closer to 255, as in the following formula,

- Here, σ is the standard deviation that adjusts the width of the Gaussian function, which is a weight determining function according to the absolute position of the class mean value -
HDR technique for light exposure correction and image quality improvement, characterized in that it is processed with
7. The method of claim 6,
An HDR technique for light exposure correction and image quality improvement, characterized in that either a Gaussian function, a linear or a non-linear function is used when calculating the fusion weight according to the class mean position of the dark class and the light class.
8. The method of claim 7,
The fusion weight with the class ratio (P ₁ , P ₂ ) and the class average value (μ ₁ , μ ₂ ) applied is,

- where T is the threshold for binary partitioning into bright and dark regions, ω _b is the fusion weight of the bright region, and ω _d is the fusion weight of the dark region -
HDR technique for light exposure correction and image quality improvement, characterized in that it is defined as
According to claim 1,
After calculating and applying the overall gamma correction coefficient using the global average (μ) and median (m) of the image, which are the primary parameters, the class ratio (P ₁ , P ₂ ) and HDR technique for light exposure correction and image quality improvement, characterized in that the fusion weight is determined by the class average value (μ ₁ , μ ₂ ).

Images