CN113947535B

CN113947535B - Low-illumination image enhancement method based on illumination component optimization

Info

Publication number: CN113947535B
Application number: CN202010694542.9A
Authority: CN
Inventors: 王正勇; 龙庆延; 何小海; 卿粼波; 吴小强; 滕奇志; 吴晓红
Original assignee: Sichuan University
Current assignee: Sichuan University
Priority date: 2020-07-17
Filing date: 2020-07-17
Publication date: 2023-10-13
Anticipated expiration: 2040-07-17
Also published as: CN113947535A

Abstract

The invention provides a low-illumination image enhancement method based on illumination component optimization. Firstly, estimating an initial illumination component by using a max-RGB model on a low-illumination original image, and optimizing the initial illumination component by using singular value decomposition and guided filtering to obtain a final illumination image. According to Retinex theory, dividing the low-illumination original image and the final illumination image point by point, and performing guided filtering on the obtained image by utilizing a G channel in three channels of the low-illumination original image RGB to remove noise. Due to the combination of singular value decomposition and guided filtering, the illumination component can be estimated more accurately, and the subjective visual effect and objective indexes of the enhanced picture are improved. The validity of the method was verified in several experiments of image enhancement.

Description

Low-illumination image enhancement method based on illumination component optimization

Technical Field

The invention relates to a low-illumination image enhancement problem in the field of image processing, in particular to a low-illumination image enhancement method based on illumination component optimization.

Background

Low-intensity image enhancement has long been a research hotspot in the field of image processing. Under the condition of lacking good illumination conditions, the image acquisition process is inevitably influenced by a plurality of factors such as parameters of equipment, illumination change and the like, so that the image shooting effect is poor, the problems of uneven illumination, color distortion and the like occur, and the follow-up processing such as target detection, target identification and the like can be influenced to a certain extent. Therefore, how to obtain high quality images under low illumination conditions is a critical issue.

Classical low image enhancement methods can be divided into three types: histogram transformation based, retinex model based and defogging model based. The histogram transformation achieves an enhancement effect by extending the grey level of the artwork and increasing the dynamic range of the image. The Retinex model is based on the principle of color constancy, and the traditional method uses gaussian filtering to estimate the illumination in the original image, and uses logarithmic transformation to remove the illumination to obtain an enhanced image. In the defogging model, the image is reversed to obtain an effect similar to that of a foggy day image, and the image is reversed again after being processed by a defogging algorithm to obtain an enhanced image. All three methods have advanced to some extent, but the visual effect and objective index are still not ideal.

Disclosure of Invention

Aiming at the problem that the low-illumination image enhancement result is not ideal, the invention provides a low-illumination image enhancement method based on illumination component optimization. And optimizing the initial illumination by utilizing singular value decomposition and guided filtering, and denoising the enhanced image by using the guided filtering again after the enhanced image is obtained by simplifying the Retinex model. The present invention achieves the above object by the following procedure:

(1) Obtaining an initial illumination component diagram by using a max-RGB model on the low-illumination original diagram;

(2) Processing the initial illumination component diagram obtained in the step (1) by using singular value decomposition, and normalizing;

(3) Obtaining an optimized illumination component diagram by using three guide filtering on the normalized result in the step (2);

(4) According to the simplified Retinex model, dividing the RGB three channels of the low-illumination original image with the illumination component image obtained in the step (3) point by point to obtain an enhanced image;

(5) And (3) denoising the enhanced image obtained in the step (4) by using a G channel in the RGB three channels of the low-illumination original image as a guide image to obtain a final required enhanced image.

Drawings

FIG. 1 is a block diagram of a low-intensity image enhancement method based on illumination component optimization;

Detailed Description

The invention is further described below with reference to the accompanying drawings:

the illumination map calculation and optimization method specifically comprises the following steps:

(1) Acquisition of initial illumination map

The initial illumination map is extracted by adopting a max-RGB model, and the calculation formula is as follows:

L(x,y)＝max(R(x,y),G(x,y),B(x,y)) (1)

wherein R (x, y), G (x, y) and B (x, y) are RGB three-channel images of the low-illumination original image respectively.

(2) Optimization of illumination patterns

Singular value decomposition is a common matrix diagonalization decomposition scheme, and matrix a can be decomposed into the following forms by using singular value decomposition:

A＝UΣV ^T (2)

wherein U and V are orthogonal matrices, i.e. UU for matrix U, V ^T =i and VV ^T =i, Σ is a non-negative diagonal matrix, the orthogonal matrices U and V contain structural information of the original matrix, and the singular matrix Σ contains energy information of the original matrix. The singular matrix Σ may be expressed as:

Σ＝diag(s ₁ ,s ₂ ,…,s _N )(s ₁ ＞s ₂ ＞…＞s _N ) (3)

wherein s is ₁ ,s ₂ ,…,s _N Is the singular value of the original matrix a.

After singular value decomposition is carried out on a 3 multiplied by 3 area in an initial illumination image, S in a singular matrix is selected ₁ I.e. the largest singular value as the illumination intensity of that point. After the whole image is processed, normalization is carried out.

The guide filter is a common edge-preserving filter, and can well preserve the edges of the image while removing texture detail information. To further optimize the illumination map, the illumination map after singular value decomposition and normalization is reprocessed using three-time guided filtering. If the initial illumination pattern is L _ini (x, y), singular value decomposition and normalization to obtain an illumination map L _s (x,y)，GF _R (A, B) represents the guided filtering with a window radius R, and the following formula exists:

wherein the window size is r ₁ ＝15、r ₂ ＝7、r ₃ ＝3，L ₃ (x, y) is the illumination map for final use after optimization.

The enhanced image and denoising method are obtained as follows:

(1) According to Retinex theory, a low-light image S (x, y) may be expressed as the product of the light component L (x, y) and the reflection component R (x, y), i.e.:

S(x,y)＝L(x,y)×R(x,y) (5)

the above-mentioned deformable is:

where τ is a constant that prevents the denominator from becoming zero, taking 0.01.

(2) An enhanced image can be obtained according to equation (6). The low-illumination image amplifies noise originally hidden in the dark while enhancing, and a denoising process is required after the enhancement. Currently used CCD cameras and CMOS cameras mostly use bayer array sensors to acquire color images, and bayer array is a 4×4 array consisting of 8 green, 4 blue and 4 red pixels. Therefore, the noise level of the green channel in the original image is usually lower, the RGB three channels of the enhancement result can be respectively denoised by using the guide filtering as the guide image, and the final result is obtained after three channels are fused.

In order to verify the effectiveness of the low-illumination image enhancement method based on illumination component optimization, experiments are carried out on 20 images, and comparison experiments are carried out on a classical low-illumination image enhancement method MSRCR, dong, SRIE and the method provided by the invention.

Table 1 average value of each index of 20 images

Tab.1Mean value of indicators for 20images

The table shows that the invention has better effect on experimental pictures, has a certain practical value, is lower than the method A in the index of color entropy and exceeds the comparison method in the indexes of LOE, NIQE, BRISQUE.

Claims

1. The low-illumination image enhancement method based on illumination component optimization is characterized by comprising the following steps of:

(2) Processing the initial illumination map obtained in the step (1) by using singular value decomposition, and normalizing;

(3) For the initial illumination map obtained in (1)L _ini Normalized result graph L obtained in (x, y) and (2) _s (x, y) conducting guided filtering with a window radius R, and the formula is as follows:wherein GF is _R (A, B) means that the guided filtering is performed with the window radius R, L ₃ (x, y) is the illumination map for final use after optimization;

(4) According to the simplified Retinex model:c epsilon { R, G, B }, taking 0.01 for preventing denominator from being zero, and dividing the RGB three channels of the low-illumination original image with the illumination component image obtained in the step (3) point by point to obtain an enhanced image;

2. The method according to claim 1, characterized in that in step (1) the initial illumination map is obtained using a max-RGB model, i.e. using the maximum of the RGB three channels of the original image at point (x, y) as illumination for that point.

3. The method of claim 1, wherein the initial illumination map is processed in step (2) using singular value decomposition, the method comprising:

singular value decomposition is performed for a 3×3 region of the image centered around point (x, y), namely:

A＝UΣV ^T (1)

wherein A represents a 3×3 region in the initial illumination map, U and V are both orthogonal matrices, Σ is a non-negative diagonal matrix composed of singular values of A, and there are:

Σ＝diag＝(s ₁ ,s ₂ ,…,s _N )(s ₁ ＞s ₂ ＞…＞s _N ) (2)

with maximum singular value S ₁ Instead of the point illumination, the whole image is processed and then fedAnd (5) normalizing the rows.

4. The method of claim 1, wherein in step (5), the enhanced image is guided, filtered and denoised by using a green channel component with a lower noise level in the original image, so as to improve the visual effect and various indexes of the image.