KR102176093B1 - Method for acquiring illumination invariant image using color space compression technique and device therefor - Google Patents

Abstract

을 추정하는 과정, 및 상기 추정된

을 기반으로 1D 조명 불변 이미지로 변환하는 과정을 포함한다.Disclosed is a method for obtaining an illumination invariant image using a color space compression technique and an apparatus therefor.
According to an aspect of the present embodiment, a method of obtaining an illumination invariant image using a color space compression technique is a process of acquiring an RGB image, a process of applying color space compression using a color offset to the acquired RGB image, and the color space compression. The process of converting the applied image to a 2D color ratio space image, and a PCA-based chromaticity projection line using the converted image

The process of estimating, and the estimated

It includes the process of converting to a 1D illumination invariant image based on.

Description

A method for acquiring an illumination invariant image using a color space compression technique, and an apparatus therefor {METHOD FOR ACQUIRING ILLUMINATION INVARIANT IMAGE USING COLOR SPACE COMPRESSION TECHNIQUE AND DEVICE THEREFOR}

The present invention relates to a method and apparatus for obtaining an illumination invariant image using a color space compression technique.

The content described in this section merely provides background information on the present invention and does not constitute prior art.

For the stable operation of the robot (or mobile robot), precise location recognition and environmental awareness of the robot are very important issues. If the robot runs without location awareness and environment awareness, the robot frequently collides with obstacles. Therefore, an environmental map is required for the stable operation of the robot. Environmental maps are made using an expensive GPS (Global Positioning System) to increase accuracy, but since even expensive GPS cannot be used indoors, they are made through actual measurement. Simultaneous Localization And Mapping (SLAM) has been studied as a technology to solve problems such as the inability to use GPS or accuracy in environmental map production. SLAM is a technology that enables the robot to simultaneously recognize the location and create a map based on the measured sensor data. In the case of using SLAM, the robot can drive through relative location recognition without an environment map created in advance, and obtain an environment map based on the route traveled. Recently, SLAM is being used in various fields such as indoor robots, autonomous vehicles, unmanned drones, and Augmented Reality (AR), and its importance is rising.

SLAM can be largely divided into Filter-based SLAM and Graph-based SLAM according to the type of algorithm to be applied. Filter-based SLAM has been widely used in SLAM in the early days, but it is rarely used at present because of the problem of error accumulation and the problem that the amount of computation increases exponentially as information is acquired. Graph-based SLAM was rarely used in the early days due to the relatively larger initial computational load than Filter-based SLAM. However, graph-based SLAM has been widely used in recent years because of the advantages of simplicity of implementation, stable computational management, and continuous map optimization.

SLAM can be divided into LiDAR SLAM, where the distance sensor is the main sensor, and visual SLAM (vSLAM), where the image sensor is the main sensor, depending on the type of sensor used. In the early study of SLAM, a distance sensor that is more convenient to use was mainly used due to the problem of computation. However, in recent years, the development of computers has solved the problem of computational volume, and since the image sensor can acquire more information than the distance sensor, research on SLAM using an image sensor is actively progressing. Furthermore, since the strengths and weaknesses of each sensor are different, studies on SLAM in a method of using two sensors by fusion are also actively progressing.

One of the most important technologies in SLAM is Loop Closure Detection (LCD). The LCD detects that the traveling robot revisits to the previously visited location. The location information of the robot can be more accurately estimated through the LCD, and if the estimated location information is used, it is possible to optimize the environment map. In other words, it is impossible to optimize the environment map without LCD.

The method of LCD differs depending on the sensor used. In general, LCDs using a distance sensor use ICP (Iterative Closest Point) to compare the similarity between keyframes to detect whether the robot has visited the location. However, since this method has low accuracy and a high probability of acquiring other information even at the same location, it is not well used unless the SLAM uses a distance sensor alone. LCDs using image sensors are mainly used in Bag of Words (BoWs). BoWs is to detect the presence or absence of the object in the acquired image by learning the image features detected in the object. Compared to feature matching, this method is widely used because it has a lower computational amount and enables probabilistic object detection. Initially, BoWs began to be detected based on local features, but later developed based on binary features capable of faster computation, but are now being developed to detect in real time without prior learning.

However, in SLAM using an image sensor, LCD has a fatal disadvantage that distance sensors do not have. That is, the information obtained from the image sensor is light source dependent. Even if the image is acquired from the same location, feature matching may not be possible if the light source, that is, the intensity, location, and type of illumination is different. Therefore, even if an image sensor is used in an indoor environment in which the type and direction of lighting can be controlled, it is not a problem, but in an outdoor environment where the location and intensity of the light source changes over time, this becomes a very fatal problem.

The main object of the present embodiment is to provide a method and apparatus for acquiring an image irrespective of the location, intensity, and type of a light source.

을 추정하는 과정, 및 상기 추정된

을 기반으로 1D 조명 불변 이미지로 변환하는 과정을 포함한다. According to an aspect of the present embodiment, a method of obtaining an illumination invariant image using a color space compression technique is a process of acquiring an RGB image, a process of applying color space compression using a color offset to the acquired RGB image, and the color space compression. The process of converting the applied image to a 2D color ratio space image, and a PCA-based chromaticity projection line using the converted image

The process of estimating, and the estimated

It includes the process of converting to a 1D illumination invariant image based on.

을 추정하고, 상기 추정된

을 기반으로 1D 조명 불변 이미지로 변환하는 이미지 처리부를 포함한다.According to another aspect of the present embodiment, an apparatus for obtaining an illumination invariant image using a color space compression technique includes an image acquisition unit for obtaining an RGB image, and a color space compression using a color offset to the obtained RGB image, and the Converts an image to which color space compression is applied to a 2D color ratio space image, and uses the converted image to use the PCA-based chromaticity projection line.

And the estimated

It includes an image processing unit that converts the 1D illumination constant image based on.

As described above, according to the present embodiment, an image irrespective of the location, intensity, and type of the light source can be obtained. These images allow the use of image sensors in SLAM.

1 is a diagram showing a flow chart of a method of obtaining an illumination invariant image using a color space compression technique according to the present disclosure;
2 is a diagram illustrating a configuration of an apparatus for obtaining an illumination invariant image using a color space compression technique according to the present disclosure.

Hereinafter, some embodiments of the present invention will be described in detail through exemplary drawings. In adding reference numerals to elements of each drawing, it should be noted that the same elements are assigned the same numerals as possible even if they are indicated on different drawings. In addition, in describing the present invention, if it is determined that a detailed description of a related known configuration or function may obscure the subject matter of the present invention, a detailed description thereof will be omitted.

In addition, in describing the constituent elements of the present invention, terms such as first, second, A, B, (a), (b) may be used. These terms are only used to distinguish the component from other components, and the nature, order, or order of the component is not limited by the term. Throughout the specification, when a part'includes' or'includes' a certain element, it means that other elements may be further included rather than excluding other elements unless otherwise stated. . In addition, the'... Terms such as'sub' and'module' mean a unit that processes at least one function or operation, which may be implemented by hardware or software or a combination of hardware and software.

The data acquired from the image sensor is dependent on the light source (or illumination). This is because the data acquired in the environment in which the light source is present is greatly affected by the shadow generated from the intensity or direction of the light source. Therefore, in the case of using an image sensor, although it is possible to obtain a wealth of information, image sensors have been mainly used in environments where light sources can be controlled, such as indoors. In other words, since the image acquired from the image sensor can make an extreme difference in contrast depending on the intensity of the light source, if the intensity of the light source is strong, information on the shadow area is not acquired or information on the shadow area, that is, the general area, disappears. There may be a problem of throwing away. Furthermore, due to the non-contiguous brightness of the shadow area and the general area, it may be judged as an edge even though it is actually a non-edge area. In order to solve this problem, research has been conducted to acquire an illumination invariant image. In general, research for acquiring an illumination invariant image can be largely divided into three categories.

The first is a study on how to detect and remove shadows from images. Research on this method was first started and is still ongoing. This method can be divided into a shadow detection process that separates a shadow area and a general area of an image, and a shadow removal process that restores the detected part. The shadow detection process may be a process of detecting an edge portion of a shadow or a shadow region itself. The shadow detection and removal method is to detect the edge map of the shadow within a single image and use the shadow removal algorithm, or simply classify the shadow edge without finding the edge of the shadow, noting that all shadows in the image do not have the same brightness difference. There is a method of finding shadow areas with the same lighting intensity value through (classification) and then removing the shadows with the weights of each area. In addition, after clustering shadow areas using a 2-D chromaticity space, shadow removal is performed based on the detected shadow area, or the illumination intensity level of the entire image area is calculated, and the calculated intensity level is used. There is also a way to restore the image. However, these methods have a large amount of computation as the process is divided into shadow detection and shadow removal. Therefore, it is not suitable when real-time calculation is required, and because it focuses on removing shadows, the robustness to lighting is also poor.

The second is a study on how to apply deep learning to the process of acquiring lighting invariant images as research on deep learning has been actively conducted recently. In general, the best result is a method using these algorithms to acquire shadow-free images through learning. However, it is very difficult to construct a dataset for learning shadowless images. In addition, unlike clustering or segmentation in which a person can designate a ground-tooth (Ground-Truth), there is a difficulty in learning due to the nature of a deep learning model that can obtain good results only by securing a sufficient amount of learning. In addition, in order to use the learning model, a high-end GPU (Graphics Processing Unit) is basically required.

The third is a study on the IIT (Illumination Invariant Transform) technique, which is commonly used in vision algorithms. The IIT technique is a technique in which only the reflectance value is left using a specific model in the acquired image, and lighting information is removed from it and converted into an original image (Intrinsic Image, II). In addition, various methods have been studied in order to acquire the original image. For example, a method of minimizing entropy or removing lighting information using a 2D wiener filter has been studied. In addition, an algorithm for obtaining an optimal original image using the maximum spectral response value of each RGB sensor or PCA technique has been proposed.

In addition, research has been conducted to independently acquire an illumination constant image without acquiring the original image. For example, using the YCbCr color space rather than the RGB color space to acquire an illumination invariant image. An image that is robust to lighting is acquired by using an Offset RGB (ORGB) image that can efficiently remove ambient light from an image. As a result, the IIT technique was applied to road detection and classification to obtain better results than the original image. This IIT technique is advantageous for real-time operation because the amount of computation is small. In addition, compared to other methods, the IIT technique has the advantage of obtaining a constant image regardless of the type and intensity of illumination. On the other hand, since the IIT technique changes the color space itself, a heterogeneous image different from the actual image can be obtained, but the characteristics of the image are maintained, so it is suitable as a preprocessing process for the vision algorithm.

Hereinafter, a method for obtaining an illumination invariant image from a general RGB image using a color space compression technique according to the present disclosure will be described in detail. The method according to the present disclosure has two main features.

The first feature is to apply a color space compression technique to the IIT technique.

In general, in order to acquire an intrinsic image using the IIT technique, a process of converting into a color ratio space using sensor values of each channel in an RGB color space is required. At this time, in the case of a dark area (eg, a shadow area), since the absolute value of the sensor value is small, it is significantly affected by noise compared to other areas. Therefore, if a general noise removal algorithm (eg, noise reduction, etc.) is applied, the amount of information other than noise in the dark region is also reduced. In addition, when the color space is compressed without weights to remove the dark area, a correct original image cannot be obtained because a color ratio is shifted. In the present disclosure, in order to solve this problem, a color compression technique using a color offset that makes the sensor response according to the intensity of illumination linear is applied to the IIT technique.

In order to linearize the sensor response, we first grasp the relationship of the sensor response according to the lighting. In general, an RGB sensor used in a color image sensor (or a color image sensor) samples the amount of light of a long wavelength (red), a medium wavelength (green), and a short wavelength (blue), respectively. The RGB sensor may be configured as one, or may be a configuration in which R, G, and B sensors are combined. Accordingly, the sensor response of one pixel in the RGB sensor can be expressed as three values. The sensor response of the RGB sensor measured at the pixel at the x position can be expressed as R ^x = [R ^x,1 , R ^x,2 , R ^x,3 ]. The sensor response R ^x,n measured by the RGB sensor may be modeled as in [Equation 1].

Here, n means any one of the RGB sensors (i.e., any one of R, G, B), λ is the wavelength of light, F ⁿ (λ) is a spectral sensitivity function of any one n of the RGB sensors, E ^x (λ) represents the spectral power distribution of the light source, and S ^x (λ) represents the reflectance of the object surface. L ^x (λ) represents the intensity of the incident light. The direct light L ^x,d (λ) that is reflected directly from the light from the light source and the light from the light source is reflected back to the target object. It can be expressed by dividing the ambient light L ^x,e (λ). If L ^x (λ) is expressed as L ^x,d (λ) and L ^x,e (λ), it can be expressed as [Equation 2].

와

는 광원과 반사면 간의 반사각을 나타내는 단위 벡터이고, μ는 객체 가려짐(object occlusion)과 빛 감쇠(light attenuation)에 의해 조정되는 변수이다. 따라서, 광원으로부터 직접광이 없는 영역이라면 μ는 0 (ex. 건물 그림자, 양산 안), 광원으로부터 직접광을 받는 영역이라면 μ는 1값을 가진다. 다시 말해, 빛의 강도를 나타내는 [수학식 2]는 직접광으로부터 야기되는 센서반응과 간접광으로부터 야기되는 센서반응의 합으로 나타낼 수 있다.

Wow

Is a unit vector representing the angle of reflection between the light source and the reflective surface, and μ is a variable that is adjusted by object occlusion and light attenuation. Therefore, if there is no direct light from the light source, μ has a value of 0 (ex. building shadow, in mass production), and if the area receives direct light from the light source, μ has a value of 1. In other words, [Equation 2] representing the intensity of light can be expressed as the sum of the sensor response caused by direct light and the sensor response caused by indirect light.

[Equation 3] shows the sensor response caused by direct light, and [Equation 4] shows the sensor response caused by indirect light.

The sensor response at pixel x R ^x represents using the equation 3 and equation 4 may be expressed by [Equation 5].

Again, if [Equation 5] is substituted into [Equation 1], the same result as [Equation 6] can be derived.

= [

,

,

] 라고 할 때, 수렴점을 통해 보정된 센서반응

은 [수학식 7]과 같이 나타낼 수 있다.In a general environment, an image acquired through an image sensor contains both an ambient light component and a direct light component. Acquiring an image containing only direct light or ambient light is impossible under normal circumstances. However, since an image containing ambient light does not guarantee color consistency, in order to restore a correct original image using the IIT technique, it is necessary to remove the ambient light component from the acquired image and leave only the direct light component. In order to acquire an image in which only the direct light component is physically present, it is necessary to subtract only the ambient light component from the acquired image, but this is practically impossible. However, since the sensor response in an image containing only the direct light component converges to one point, the same effect as removing the ambient light component can be obtained by finding the point where the sensor response converges and compressing the color space of the image. The convergence point of the RGB sensor

= [

,

,

], the sensor response corrected through the convergence point

Can be expressed as [Equation 7].

를 찾기 위해 선형화 방법이 이용될 수 있으며, 수렴점

는 컬러 오프셋이 된다.At this time, the convergence point

The linearization method can be used to find the convergence point

Becomes the color offset.

The second feature is to apply Principal Compression Analysis (PCA) to the IIT technique.

The existing IIT technique restored the image acquired using the maximum spectral response of each RGB sensor to the original image. However, this method is possible only if you know the specifications of the sensor in advance, and the acquired image has a lot of noise, which is a problem. In order to solve this problem, the present disclosure proposes a method of obtaining an optimal result by using a PCA with a small amount of computation in the IIT technique.

을 찾아서 1D gray-scale 조명 불변 이미지(즉, 원본 이미지)로 변환하는 것이다.The IIT technique can be roughly divided into two parts. The first is to convert the RGB color space to a 2D color ratio space, and the second is to convert the optimal chromaticity projection line in the converted color ratio space.

Is to find and convert it to a 1D gray-scale illumination invariant image (i.e. the original image).

First, PNL-assumption is applied to an image to be converted, that is, an image to which color space compression is applied. PNL-assumption assumes that the image is reflected on Lambertian surfaces under Planckian illumination illumination and acquired with narrow-band sensors. When PNL-assumption is used, it is possible to convert into a 2D color ratio space from which lighting information has been removed through the log ratio of each RGB sensor. The conversion equation can be expressed as [Equation 8].

However, the image sensor used in a typical camera is not a narrow band sensor. Therefore, an error occurs according to PNL-assumption, and [Equation 8] cannot be applied. A conversion equation applicable to a general broad-band sensor other than a narrow-band sensor can be expressed as [Equation 9].

의 각도를 추정한다. 이러한 방법은 기존의 엔트로피 최소화(entropy minimization)로

을 찾아내는 것보다 노이즈에 강인하며 동작속도도 빠르다. 또한, 센서특성을 사용하여 원본 이미지를 취득하는 것에 비해 높은 정보량의 이미지를 취득할 수 있다.Acquire a 2D color ratio space image using the color ratio, and then use PCA to optimize the chromaticity projection line.

Estimate the angle of This method can be accomplished by conventional entropy minimization.

It is more robust to noise than to find and operates faster. In addition, it is possible to acquire an image of a higher information amount than that of acquiring an original image by using the sensor characteristics.

의 각도를 추정하기 위해서 PCA를 사용해 2D color ratio space 이미지의 값이 최대 분산을 이루는 고유벡터

과

를 추정한다. 먼저 이미지의 픽셀을 랜덤으로 샘플링하여 2 x n 행렬(X)를 구성한다. 구성된 행렬(X)를 사용하여 공분산 행렬 C = XX ^T 를 구성하고, 여기에 특이값 분해(singular value decomposition)를 적용하여 고유벡터

과

를 추정한다. 상기 공분산 행렬은 [수학식 10]을 이용해 구할 수 있다.Optimized

Eigenvectors where the values of the 2D color ratio space image have the maximum variance using PCA to estimate the angle of

and

Estimate First, the pixels of the image are randomly sampled to form a 2 xn matrix ( X ). Construct a covariance matrix C = XX ^T using the constructed matrix ( X ), and apply a singular value decomposition to the eigenvector

and

Estimate The covariance matrix can be obtained using [Equation 10].

과

는 행렬 X의 고유벡터이고,

과

는 각각의 고유값이다. 둘 중 큰 분산 값을 가지는

과 고유벡터

는 투영선

로 선택된다. 원본 이미지로 변환하기 위해 2D color ratio space에서 투영선

과 직교하는 선분 위에 있는 픽셀들을 같은 반사율을 가진 표면으로 가정하고 1D gray scale space로 투영시킨다. 최종결과로는 조명 정보가 제거되어 반사율 정보만이 남은 원본 이미지를 획득할 수 있다.

and

Is the eigenvector of matrix X ,

and

Is each eigenvalue. Whichever has the greater variance

And eigenvectors

Is the projection line

Is selected as Projection line in 2D color ratio space to convert to original image

The pixels on the line segment perpendicular to and are projected in 1D gray scale space, assuming that the surface has the same reflectivity. As a final result, the lighting information is removed, and the original image with only the reflectance information remaining can be obtained.

[Table 1] below shows the overall process.

Input : RI _t , O
Output : II _t
begin
for every input 3D Color Image RI _t
AutoBalanceWhite( RI _t )
for every channels n = R, G, B
for every pixels x
CI _t ^{x , n} = ( RI _t ^{x , n} - O ⁿ )/(1- O ⁿ )
end
end
for every pixels x
CRI _t ^{x , r} = log( CI _t ^{x , R} /( CI _t ^{x , R} × CI _t ^{x , G} × CI _t ^{x , B} ) ^1/3 )
CRI _t ^{x , b} = log( CI _t ^{x , B} /( CI _t ^{x , R} × CI _t ^{x , G} × CI _t ^{x , B} ) ^1/3 )
end
for sample size ss
Random sampling CRI _t ^x
if mean of sampled pixel value 100 < CI _t ^x <200
add CRI _t ^x to sampled pixel SI _t ^y
end
end
get e ₁ , e ₂ using SI _t
α = | arctan ( e ₁ / e ₂ ) |
for every pixels x
II _t ^x = (cos α × CRI _t ^r ) + (sin α × CRI _t ^bb )
end
normalization( II _t )
return II _t
end
end

RI _t is a 3D RGB image acquired at time t, O is a color offset acquired in advance, and II _t is a 1D illumination invariant image acquired as a final result. First, if RI _t is input, white balance is corrected in the input image. This is because an image with an accurate color temperature is required to apply the IIT technique. Next, a compressed image CI _t is obtained by applying a color space compression (CSC) technique to RI _t whose color temperature is corrected. CI _t can be obtained by applying [Equation 7] to each RGB channel using O obtained in advance. The obtained CI _t is converted into a 2D color ratio image CRI _t . CRI _t can be derived using [Equation 9]. Next, randomly select ss samples from CRI _t . ss is calculated as column * row / m by reflecting the size of the image. In this disclosure, m = 1000 is applied. However, when selecting a sample, there may be constraints. When the dark and bright areas of the image information are noisy and applied to the sample pixel, a CRIt can be added to the sample pixel to prevent a problem that the PCA does not converge. Using the sampled image SI _t , PCA is applied to obtain the eigenvectors e ₁ and e ₂ , and α is obtained using their ratio. Finally, convert CRI _t ^x to II _t using α and apply normalization to II _t . However, when normalizing, the normalization range is set to a multiple of 0.8 instead of 1. Almost no useful information exists in 0~0.1 and 0.9~1, and by doing this, the contrast is increased, and an image with a higher amount of information can be obtained.

1 is a flowchart illustrating a method of obtaining an illumination invariant image using a color space compression technique according to the present disclosure.

First, an RGB image is acquired (S110). For example, the RGB image is an image acquired at time t using an RGB sensor. In order to obtain a more accurate color temperature image, the white balance may be corrected on the acquired RGB image.

Color space compression is applied to the acquired RGB image using a color offset (S120). The color offset can be obtained in advance as an offset that makes the sensor response according to the intensity of illumination linear.

Thereafter, the image to which the color space compression is applied is converted into a 2D color ratio space image (S130). The 2D color ratio space image can be derived using [Equation 8] or [Equation 9]. [Equation 8] is used if the sensor for acquiring an image is a narrow band sensor, and [Equation 9] is used for a wide band sensor.

을 추정한다(S140). 상기 변환된 2D color ratio space 이미지에서 임의로 샘플을 선정하고, 선정된 샘플에 PCA 기반으로 색도 투영선

을 추정한다. 상기 샘플은 이미지의 크기를 반영할 수 있다. 또한, 필요에 따라서는 샘플 선정시 제약조건이 붙을 수 있다. 상기 선정된 샘플로 행렬을 구성하고, 공분산 행렬을 구해 고유벡터를 추정하고, 상기 추정된 고유벡터로부터 투영선

을 추정한다.PCA-based chromaticity projection line using the converted image

Is estimated (S140). Randomly select a sample from the converted 2D color ratio space image, and a chromaticity projection line based on PCA on the selected sample

Estimate The sample may reflect the size of the image. In addition, if necessary, constraints may be imposed upon sample selection. Construct a matrix from the selected samples, obtain a covariance matrix to estimate an eigenvector, and a projection line from the estimated eigenvector

Estimate

을 기반으로 1D 조명 불변 이미지로 변환한다(S150). 이후, 상기 변환된 1D 조명 불변 이미지에 1이 아닌 0.8로 맞춰 정규화를 적용할 수 있다. Above estimated

Converts to a 1D illumination constant image based on (S150). Thereafter, normalization may be applied to the converted 1D illumination invariant image by setting it to 0.8 instead of 1.

In FIG. 1, steps S110 to S150 are described as sequentially executing, but this is merely illustrative of the technical idea of an embodiment of the present invention. In other words, if one of ordinary skill in the technical field to which an embodiment of the present invention belongs, changes the order shown in FIG. 1 and executes one of the processes S110 to S150 without departing from the essential characteristics of the embodiment of the present invention. Since the above processes are executed in parallel, various modifications and variations may be applied, and thus FIG. 1 is not limited to a time series order.

Meanwhile, the processes shown in FIG. 1 can be implemented as computer-readable codes on a computer-readable recording medium. The computer-readable recording medium includes all types of recording devices that store data that can be read by a computer system. That is, the computer-readable recording medium is a magnetic storage medium (e.g., ROM, floppy disk, hard disk, etc.), optical reading medium (e.g., CD-ROM, DVD, etc.), and carrier wave (e.g., Internet And storage media such as transmission through In addition, the computer-readable recording medium can be distributed over a computer system connected through a network to store and execute computer-readable codes in a distributed manner.

2 is a block diagram of an apparatus for obtaining an illumination invariant image using a color space compression technique according to the present disclosure.

In FIG. 2, a description is made by dividing into a plurality of configurations, but multiple configurations may be integrated into one configuration, or one configuration may be divided into multiple configurations.

Referring to FIG. 2, an apparatus for obtaining an illumination invariant image using a color space compression technique according to the present disclosure may include an image acquisition unit 210 and an image processing unit 220.

The image acquisition unit 210 acquires an RGB image using an RGB sensor. At this time, the acquired RGB image may be a 3D RGB image.

The image processing unit 220 applies color space compression to the image acquired by the image acquisition unit 210 by using a color offset, and performs 2D color ratio space conversion. The color offset can be obtained in advance as an offset that makes the sensor response according to the intensity of illumination linear. The converted 2D color ratio space image can be obtained using [Equation 8] or [Equation 9]. [Equation 8] is used if the sensor for acquiring an image is a narrow band sensor, and [Equation 9] is used for a wide band sensor.

을 추정하고, 상기 추정된

을 기반으로 1D 조명 불변 이미지로 변환한다. 상기 색도 투영선

은 상기 변환된 2D color ratio space 이미지에서 임의로 샘플을 선정하고, 선정된 샘플에 PCA 기반으로 추정할 수 있다. 상기 샘플은 이미지의 크기를 반영할 수 있다. 또한, 필요에 따라서는 샘플 선정시 제약조건이 붙을 수 있다. 상기 선정된 샘플로 행렬을 구성하고, 공분산 행렬을 구해 고유벡터를 추정하고, 상기 추정된 고유벡터로부터 색도 투영선

을 추정한다.Then, the image processing unit 220 is a PCA-based chromaticity projection line

And the estimated

It converts to a 1D illumination invariant image based on. The chromaticity projection line

May randomly select a sample from the converted 2D color ratio space image, and estimate the selected sample based on PCA. The sample may reflect the size of the image. In addition, if necessary, constraints may be imposed upon sample selection. Construct a matrix from the selected samples, obtain a covariance matrix to estimate an eigenvector, and a chromaticity projection line from the estimated eigenvector

Estimate

The image processing unit 220 may correct a white balance on the acquired RGB image in order to acquire an image having a more accurate color temperature. Thereafter, normalization may be applied to the converted 1D illumination invariant image by setting it to 0.8 instead of 1.

The above description is merely illustrative of the technical idea of the present embodiment, and those of ordinary skill in the technical field to which the present embodiment belongs will be able to make various modifications and variations without departing from the essential characteristics of the present embodiment. Accordingly, the present exemplary embodiments are not intended to limit the technical idea of the present exemplary embodiment, but are illustrative, and the scope of the technical idea of the present exemplary embodiment is not limited by these exemplary embodiments. The scope of protection of this embodiment should be interpreted by the following claims, and all technical ideas within the scope equivalent thereto should be interpreted as being included in the scope of the present embodiment.

Claims (10)

In a method of obtaining an illumination invariant image using a color space compression technique,
The process of acquiring an RGB image;
Generating an RGB image corrected by using a color offset on the acquired RGB image;
Applying color space compression to the corrected RGB image;
Converting the image to which the color space compression is applied into a 2D color ratio space image;
PCA-based chromaticity projection line using the converted image

The process of estimating; And
Above estimated

The process of converting to a 1D lighting invariant image based on
Including,
The color offset is an offset that causes the sensor response according to the intensity of illumination to become linear, and the R, G, and B color components of each pixel of the corrected RGB image are determined by the following equation. How to acquire an image.

Where x represents the associated pixel, n represents the associated color component,

Represents the n color component of the x- th pixel of the corrected RGB image,

Is the n color component of the x- th pixel of the acquired RGB image acquired at time t,

Is the color offset corresponding to each color component. The method of claim 1,
Before applying the color space compression,
A method of obtaining an illumination invariant image further comprising a step of correcting a white balance on the acquired RGB image. The method of claim 1,
remind

The process of estimating is,
An illumination invariant image acquisition method, characterized in that a process of using a sample randomly selected from the 2D color ratio space image. The method of claim 1,
A method of obtaining an illumination invariant image comprising the process of further applying normalization to the converted 1D illumination invariant image. delete In an apparatus for obtaining an illumination invariant image using a color space compression technique,
An image acquisition unit that acquires an RGB image; And
A corrected RGB image is generated using a color offset to the acquired RGB image, color space compression is applied to the corrected RGB image, the color space compression applied image is converted into a 2D color ratio space image, and the converted PCA-based chromaticity projection line using images

And the estimated

Including an image processing unit for converting the 1D illumination constant image based on,
The color offset is an offset that causes the sensor response according to the intensity of illumination to become linear, and the R, G, and B color components of each pixel of the corrected RGB image are determined by the following equation. Image acquisition device.

Where x represents the associated pixel, n represents the associated color component,

Represents the n color component of the x- th pixel of the corrected RGB image,

Is the n color component of the x- th pixel of the acquired RGB image acquired at time t,

Is the color offset corresponding to each color component. The method of claim 6,
The image processing unit,
Before applying the color space compression, a white balance is further corrected on the acquired RGB image. The method of claim 6,
The image processing unit,
Using a randomly selected sample from the 2D color ratio space image, the

Illumination invariant image acquisition device, characterized in that to estimate. The method of claim 6,
The image processing unit,
An illumination invariant image acquisition device, characterized in that normalization is further applied to the converted 1D illumination invariant image. delete

Images