KR102686707B1 - Ghost Image Acquisition Method and Electronic Device Using The Same - Google Patents

Abstract

The image acquisition method according to the present invention uses a ghost image acquisition method to generate a plurality of correlation data pairs consisting of the intensity of light sequentially measured by light coming from a light source and reacting (passing or reflecting) with an object and the beam pattern of the light source. 1 Obtaining a ghost image, removing noise from the acquired first ghost image using a Markov random field-based Bayesian noise removal model, and sequentially measuring the light intensity of the light coming out of the light source and reacting with the object and acquiring a second ghost image using a ghost image acquisition method on a plurality of correlation data pairs composed of the beam pattern of the light source, and Markov random field-based Bayesian noise using the first ghost image from which the noise has been removed as prior information. Removing noise from the second ghost image using a removal model, wherein the beam pattern of the light source may be a beam pattern at the time of measuring the intensity of light through the optical sensor.

Description

Ghost Image Acquisition Method and Electronic Device Using The Same }

The present invention relates to a method of acquiring a ghost image using a Bayesian noise removal model and an electronic device using the same.

Images (or videos) are generally obtained through the process of collecting light that reacts with an object with a two-dimensional optical sensor and converting it into an electrical signal. In contrast, a method that can acquire images even with light that does not directly react with the object was introduced relatively recently. In terms of being able to photograph non-existent objects, this is called a ghost image, and the system and processing process for acquiring it are called ghost imaging.

Traditional ghost imaging requires processing large amounts of data to obtain high-quality images. Accordingly, much research has been conducted to obtain high-quality ghost images even with a small amount of data. For example, if the size of the object is extremely small compared to the background, differential ghost imaging (DGI) or normalized ghost imaging (NGI) can be used as a ghost image acquisition formula to improve the SNR (signal to noise ratio) of the ghost image, or Studies have been conducted using uniformly weighted ghost imaging (UWGI) and uniformly weighted differential ghost imaging (UWDGI) as ghost image acquisition formulas to correct noise caused by beam patterns with uneven light intensity.

In addition, deep learning-based ghost imaging is being researched, but due to the nature of deep learning, there is a limitation that it is difficult to acquire images of untrained objects.

Ghost imaging using two correlated beams has the advantage of being able to obtain higher SNR images than conventional imaging under adverse conditions. However, there is a disadvantage in that acquiring a high SNR ghost image requires a large amount of data and requires a lot of time to accumulate it.

The image acquisition method according to the present invention uses a ghost image acquisition method to generate a plurality of correlation data pairs consisting of the intensity of light sequentially measured by light coming from a light source and reacting (passing or reflecting) with an object and the beam pattern of the light source. 1 Obtaining a ghost image, removing noise from the obtained first ghost image using a Markov random field-based Bayesian noise removal model, and sequentially measuring the light intensity of the light coming out of the light source and reacting with the object and acquiring a second ghost image using a ghost image acquisition method on a plurality of correlation data pairs composed of the beam pattern of the light source, and Markov random field-based Bayesian noise using the first ghost image from which the noise has been removed as prior information. Removing noise from the second ghost image using a removal model, wherein the beam pattern of the light source may be a beam pattern at the time of measuring the intensity of light through the optical sensor.

The image acquisition device according to the present invention includes an optical sensor that measures the intensity of light, and a processor, and the processor sequentially measures light that comes from a light source and reacts (passes or reflects) with an object through the optical sensor. A first ghost image is acquired using a ghost image acquisition method on a plurality of correlated data pairs consisting of the light intensity and the beam pattern of the light source, and noise is removed from the obtained first ghost image using a Markov random field-based Bayesian noise removal model. Remove the light coming out of the light source and reacting with the object, sequentially measuring the light intensity through the optical sensor and using a ghost image acquisition method to obtain a second ghost using a plurality of correlation data pairs consisting of the beam pattern of the light source. An image is acquired, and noise is removed from the second ghost image using a Markov random field-based Bayesian noise removal model using the noise-removed first ghost image as prior information, and the beam pattern of the light source is set to the optical sensor. It may be a beam pattern at the time of measuring the intensity of light.

According to the present invention, an image with high SNR can be obtained even with a relatively small amount of data.

According to the present invention, images with high SNR can be acquired in a short time.

The present invention provides not only classical ghost imaging with the characteristics of a quasi-thermal light source and a Hanbury-Brown and Twiss interferometer structure, but also computational ghost imaging using a digital micromirror device or spatial light module. ) can also be applied. In addition, it can be applied to existing imaging technology fields that acquire images using accumulated data for each pixel over time, such as non-destructive testing using radiation.

The present invention can be used in various fields such as military LiDAR, self-driving cars, bio-imaging, and space telescopes that require high-quality image acquisition in adverse conditions.

Figure 1 shows an example of a ghost imaging system 100.
FIG. 2A shows an example of a flow chart performed by the image processing device 180, and FIG. 2B shows an example of an image derived from each step of FIG. 2A.
Figure 3 shows a hidden Markov random field (Hidden MRF, HMRF) among various structures of MRF as an example.
Figure 4 shows the simulation process of acquiring a ghost image based on a random beam pattern actually measured using a pseudo-thermal light source.
5 to 7 show binary ghost images obtained through TGI, DGI, UWDGI, CGI, and the method proposed by the present invention (BDGI) according to the accumulated data number.
Figure 8 shows the peak signal-to-noise ratio (PSNR) and structural similarity index measure (SSIM) for each object.
Figures 9 and 10 show black-and-white scale ghost images obtained through TGI, DGI, UWDGI, CGI, and the method proposed by the present invention (BDGI) according to the accumulated data number.
Figure 11 shows quantitative analysis results for the results shown in Figures 9 and 10.

The terms used herein are for descriptive purposes only and should not be construed as limiting. Singular expressions include plural expressions unless the context clearly dictates otherwise. In this specification, terms such as “comprise” or “have” are intended to designate the presence of features, numbers, steps, operations, components, parts, or combinations thereof described in the specification, but are not intended to indicate the presence of one or more other features. It should be understood that this does not exclude in advance the possibility of the existence or addition of elements, numbers, steps, operations, components, parts, or combinations thereof.

Unless otherwise defined, all terms used herein, including technical or scientific terms, have the same meaning as generally understood by a person of ordinary skill in the technical field to which the embodiments belong. Terms defined in commonly used dictionaries should be interpreted as having a meaning consistent with the meaning in the context of the related technology, and should not be interpreted in an ideal or excessively formal sense unless explicitly defined in the present application. No.

Hereinafter, the present invention will be described in more detail. However, the following examples only illustrate the present invention, and the content of the present invention is not limited to the following examples.

Figure 1 shows an example of a ghost imaging system 100.

Referring to FIG. 1 , the ghost imaging system 100 may include two optical sensors 130 and 140 and an image processing device 180. Light emitted from the light source unit 110 may pass through the beam splitter 120, have its intensity divided by a predetermined ratio (e.g., 50:50), and enter the optical sensors 130 and 140 located in each beam path. The beam path can be divided into a path 160 where the object 150 is placed and a path 170 where the object does not exist. The path 160 along which the object 150 is placed may be referred to as an objective arm, and the optical sensor 140 located on the objective arm 160 is a sensor that can measure the intensity of light reacting with the object, for example. For example, it may be at least one of a bucket detector, a photodiode-based optical sensor, and a pixelized camera. The path 170 where no object exists may be referred to as a reference arm, and the optical sensor 130 located on the reference arm 170 is a sensor capable of measuring the beam pattern of received light, for example, a CCD It may be at least one of a (charge couple device) camera, a complementary metal-oxide semiconductor (CMOS) camera, and a pixel-based camera. The image processing device 180 includes a processor and can process light received through the objective arm 160 and the reference arm 170. The image processing device 180 can process operations described below.

According to one embodiment, in a computing ghost imaging system using a digital micro-reflective indicator or a spatial light modulation device, information about the light source can be known in advance, so the beam pattern through the reference arm 170 may not be measured. Accordingly, the beam splitter 120 and/or the optical sensor 130 may be omitted.

Hereinafter, it is assumed that the light intensity is divided 50:50 by the beam splitter 120.

The beam pattern ( ) is divided in intensity 50:50 in the beam splitter 120, and when the 2D camera 130 and the object 150 are located at the same distance (Z ₁ = Z ₂ ), each beam pattern is the same. The beam pattern refers to the light intensity signal after reacting with an object, and can be expressed as [Equation 1] below.

[Equation 1]

= d

here, and indicates the positions of the optical sensor 130 and the object 150, respectively. represents the response function (transmission or reflection) of the beam pattern and the object, and integration represents the process of collecting responses from all two-dimensional planar light sources with the single pixel bucket detector 140.

The beam pattern may be received by the bucket detector 140 through the objective arm 160, and the same beam pattern may be received by the optical sensor 130 through the reference arm 170. The signal received by the bucket detector 140 may be a scalar signal (B), and the signal received by the optical sensor 130 may be a beam pattern ( ) can be. The ghost image is divided into two signals B, obtained at the same time point t, divided by the intensity of 50:50 in the beam splitter 120, It can be obtained using the spatial correlation of . The ghost image is a scalar value (B) and a beam pattern (B) obtained from the two optical sensors (130, 140) as shown in [Equation 2] It can be obtained by the covariance of , and the beam pattern of the reference arm (170) ( ) is an accumulated image weighted by the light intensity signal (B) of the objective arm (160).

[Equation 2]

=

where , N is the number of measurements

As described above, ghost imaging is a new imaging method that can obtain an image of an object from a light pattern that does not directly react with the object using the correlation of light intensity. A typical cause that hinders the acquisition of images with a high signal-to-noise ratio is image blur caused by the scattering medium located in the beam path. Image blur is a phenomenon in which the boundary of the target object in the image becomes unclear due to the optical sensor's inability to accurately obtain a response function with the object due to impurities other than the object in the beam path. Another cause of SNR deterioration is a low-illuminance light source in which the amount of light flux reacting with the object is small. This literally means that the amount of light reacting with the object is small, so the object is not revealed in the image. Ghost imaging has the characteristic of being more robust than existing imaging methods to image quality degradation caused by scattering media and low-illuminance light sources.

The above describes the case of measuring the beam pattern through the reference arm 170, but it can also be applied to cases where the beam pattern of the light source can be known even without directly measuring it, such as in a computing ghost imaging system.

FIG. 2A shows an example of a flow chart performed by the image processing device 180, and FIG. 2B shows an example of an image derived from each step of FIG. 2A.

Referring to FIG. 2A, the image processing device 180 stores a ghost image in a plurality of correlation data pairs consisting of the intensity of light sequentially measured from the light that comes out of the light source and reacts (passes or reflects) with the object and the beam pattern of the light source. The first ghost image (Y ⁱ ) can be acquired using an acquisition method (S201). The image processing device 180 may directly measure the beam pattern of the light source, but may not measure it if the beam pattern of the light source is known. Here, the path through which the light from the light source reacts with the object and is received by the light sensor is the objective arm (160), and the path through which the light from the light source does not react with the object and is received by the light sensor can be the reference arm (170). there is. Specifically, the image processing device 180 may sequentially receive data through each of the objective arm 160 and the reference arm 170. Alternatively, the image processing device 180 can sequentially receive data only through the objective arm 170 if the beam pattern of the light source is known. The image processing device 180 manages data received simultaneously or at a point in time determined to have been received simultaneously as a correlation data pair (data received through the objective arm 160, data received through the reference arm 170). or processing). Data received through the objective arm 160 may be light intensity, and data received through the reference arm 170 may be a beam pattern. The image processing device 180 may acquire a first ghost image using a ghost image acquisition method for a plurality of correlated data pairs. The first ghost image can be acquired using a ghost image acquisition method that accumulates the beam pattern received through the reference arm 170 or already known, weighted by the beam intensity received through the objective arm 160. For example, the image processing device 180 may acquire one ghost image using a ghost image acquisition method for 100 received correlation data pairs.

The image processing device 180 may remove noise from the acquired first ghost image using a Markov random field-based Bayesian noise removal model (S203). Specifically, the image processing device 180 may initialize the acquired first ghost image Y ⁱ into a binarized image. The image processing device 180 may obtain an image (X ⁱ⁺¹ ) from which noise has been removed from the binarized image using a Markov random field-based Bayesian noise removal model. The noise-removed image (X ⁱ⁺¹ ) is the prior information ( Prior).

The image processing device 180 creates a second ghost image (Y ^{i +1} ) can be obtained (S205). The image processing device 180 may acquire the second ghost image (Y ⁱ⁺¹ ) in the same manner as S201.

The image processing device 180 may remove noise from the second ghost image using a Markov random field-based Bayesian noise removal model using the first ghost image from which the noise has been removed as prior information (S207). The Markov random field may be a hidden Markov random field. The image processing device 180 may remove noise from the second ghost image by using the first ghost image from which the noise has been removed as prior information for modeling the joint probability distribution of the Markov random field-based Bayesian noise removal model.

Specifically, the noise removal method using the Markov random field-based Bayesian noise removal model converts the second ghost image into a binary image or a black-and-white scale image, performs Gibbs sampling using the first ghost image, and calculates the posterior probability using Gibbs sampling. Thus, a mask M can be generated, and noise can be removed from the converted second ghost image using the generated mask. Figure 2b shows images generated at each stage. A more specific Markov random field-based Bayesian noise removal model is described below.

The image processing device 180 uses a ghost image acquisition method to generate a third image using a ghost image acquisition method on a plurality of correlation data pairs consisting of the light intensity and the beam pattern of the light source, which are sequentially measured for the light that comes out of the light source and reacts (passes or reflects) with the object. A ghost image can be acquired, and noise can be further removed from the third ghost image using a Markov random field-based Bayesian noise removal model using the noise-removed second ghost image as prior information. The image processing device 180 repeatedly acquires more ghost images and removes noise from the additionally acquired ghost images using a Markov random field-based Bayesian noise removal model that uses the ghost image from which noise was just removed as prior information. .

Below, the above-described Markov random field-based Bayesian noise removal model will be described in detail.

Bayesian inference is a method of statistical inference based on the Bayesian theorem expressed in [Equation 3], where the likelihood of measurement data Y for parameter ))) and the posterior probability (P(X|Y)) through prior information (P(X)) about the parameter X to be inferred.

[Equation 3]

When applying Bayesian inference to modeling for image noise removal, X can be assumed to be an image without noise (original image), and Y can be assumed to be an image containing noise (acquired image). An ideal image of a natural object has similar pixel values between adjacent pixels, but noise caused by various causes can cause differences in values between adjacent pixels, deteriorating image quality. Here, the pixel value means the intensity of light, and the intensity of light can be expressed in various ways and units. A representative probability model that can represent these image characteristics is the Markov random field (MRF). MRF is a type of graphical model that approximates a joint probability distribution (e.g., P(X={x _{1 …} It is a probabilistic model that can be done. This conditional independence structure based on Markov properties can allow the distribution of a multidimensional space to be expressed as a decomposed dimensional space.

Figure 3 shows a hidden Markov random field (Hidden MRF, HMRF) among various structures of MRF as an example.

In a hidden Markov random field, each pixel in the image is a random variable called a node. According to the Markov property, one pixel (x _ij ) (i.e., node) in a noise-free image (X) has four adjacent nodes (x' _ij ∈ N(x _ij ,y _ij )) and ∈ corresponding observations. (y _ij ∈ N(x _ij ,y _ij )) Depends only on nodes. Here, the set of neighboring nodes (x _ij ^mb ) may be called a Markov blanket. For example, the set of neighboring nodes (x ₂₂ ^mb ) of one pixel (x ₂₂ ) (301) in a noise-free image (X) is {y ₂₂ (311), x ₁₂ (303), x ₂₁ (305) ), x ₂₃ (307), x ₃₂ (309)}. On the other hand, it has a conditional independent relationship with the remaining nodes (x _ij ⊥R(x _ij )). Accordingly, the posterior probability of [Equation 3] can be decomposed and expressed as [Equation 4].

[Equation 4]

here, , h is the number representing the height of the image in pixels, w is the number representing the width of the image in pixels, α and β are hyperparameters, and C is the clique.

In [Equation 4], means a set of all nodes belonging to one maximal clique, and each different node is necessarily dependent on each other. Here, the maximum clique refers to a graph in which no more nodes can be added and two different nodes must be connected by one edge. There are two terms ( , ) can be re-expressed as Here the first term represents the relationship between a pixel value in an image (y) containing noise and a corresponding pixel in an image (x) that does not contain noise, and may be called a data cost function or unary term. second term can be called a smooth function or a binary term that represents the similarity between a pixel (x _ij ) and its neighboring pixel (x _ij ') in an image (x) that does not contain noise. The two terms correspond to likelihood (P(Y|

Meanwhile, the wider the range of pixel values, the more operations are required to calculate the posterior probability inference. Accordingly, the Ising model, which was proposed as a model that simply describes the magnetism of materials, can be applied to the present invention. The Easing model makes the marginal probability (Z(y)) relatively simple, making it easier to calculate posterior probabilities.

The present invention assumes that each pixel constituting the image is a binary image with a value of -1 or 1. Considering the characteristics of the easing model, the data cost function is (x _ij ,y _ij )=x _ij Хy _ij and the smooth function is When setting (x _ij ,x' _ij )=x _ij Хx' _ij , a posterior probability model such as [Equation 5] can be created.

[Equation 5]

At this time, the probability that one pixel (x _ij ) in the noise-free image has a value of 1 may be equal to [Equation 6] below.

[Equation 6]

[Equation 6] is Gibbs sampling, a type of Markov chain Monte Carlo method, and the Markov chain Monte Carlo (MCMC) method is a representative method used to calculate posterior probabilities in Bayesian inference problems. It is a sampling method. Gibbs sampling can effectively estimate parameters by sampling each random variable rather than the entire sample when conditional probabilities between random variables are given. In the present invention, it is possible to calculate a distribution P(x _ij =1 | y _ij ) representing the probability that the value of one pixel in a noise-free image estimated through Gibbs sampling is 1.

The posterior probability obtained in this way can be used as a mask to remove noise from noisy images. At this time, if the image to be acquired is a binary image, a binary denoising mask M can be obtained by converting each pixel value to 1 if the value is above a certain threshold for the posterior probability, and -1 otherwise. there is. On the other hand, if the image to be acquired is a gray-scale image with values between 0 and 255, the posterior probability image itself, which consists of probability values between 0 and 1, is used as a gray-scale mask M. It can be used as. By applying a Gaussian kernel to the mask obtained in this way, a modified mask, G(M), can be generated so that the values between neighboring pixels are similar. Finally, the estimated image (X) can be obtained by projecting this noise-removed image (denoised image) onto the existing input noise image (G(M)ХY).

The estimated original image

Hereinafter, simulation results for verifying the ghost imaging method based on the Bayesian noise removal structure presented in the present invention will be described.

Figure 4 shows the simulation process of acquiring a ghost image based on a random beam pattern actually measured using a pseudo-thermal light source.

When the beam pattern emitted from a quasi-thermal light source passes through the beam splitter, the same beam pattern appears in both paths: the reference arm and the objective arm. is hired. Signal received by bucket detector can be considered the sum of the pixel values of the pattern after reacting with the virtual object. A correlation-based ghost image was obtained using TGI, DGI, NGI, UWGI, and UWDGI for the two correlated signals, and a compressed sensing ghost image was also obtained according to the method proposed in the present invention.

In order to verify the method proposed in the present invention, images acquired using the most basic ghost image acquisition method were prepared in advance. A total of 10 ghost images Y={y ⁰ ,y ¹ ... in units of 100 data units. y ⁹ } was initialized as a binary image that was 1 if the pixel value was greater than the average pixel value of each image, and -1 otherwise. Hyperparameters α and β were set to 0.5 and 0.5, respectively. The initial value was input into the MRF-based Bayesian noise removal model, and the process of Gibbs sampling, posterior probability calculation, and binary image acquisition was performed. At this time, the threshold for the binary image acquisition process was set to 0.95. The estimated original image (x) for the corresponding step was obtained by masking the binary image to which the Gaussian filter was applied to the input image. The above steps were repeatedly applied to a total of 10 ghost images by inputting the image from which the noise was removed in the previous step (k) as prior information for the noise removal model in the next step (k+1).

5 to 7 show binary ghost images obtained through TGI, DGI, UWDGI, CGI, and the method proposed by the present invention (BDGI) according to the accumulated data number.

In the case of a relatively simple object, a double slit, the object can be identified rather easily even from a ghost image using existing methods. However, the more complex objects 'GI' and 'KR' could be visually confirmed when there were more than about 800 pieces of data, but their shapes were still not clear due to noise. On the other hand, the ghost image obtained through the method proposed in the present invention was able to identify objects with relatively low noise with only about 500 pieces of data.

Below, a quantitative analysis was conducted on the ghost images obtained through each method.

Figure 8 shows the peak signal-to-noise ratio (PSNR) and structural similarity index measure (SSIM) for each object.

PSNR is an image quality index based on the numerical difference between two image pixels, and SSIM is known as an index that can comprehensively evaluate image quality in three aspects: brightness, contrast, and structure, which are difficult to explain with PSNR. While correlation-based ghost imaging techniques (TGI, DGI, NGI, UWGI, UWDGI) showed poor performance in terms of PSNR for all cases (double-slit, GI, KR), the method proposed by the present invention (BDGI) showed poor performance in terms of PSNR. It showed higher PSNR in all cases (double-slit, GI, KR) than the CGI method, which previously showed excellent performance.

In addition, for SSIM, the method proposed by the present invention (BDGI) showed higher performance than the existing acquisition method. For all cases (double-slit, GI, KR), the method proposed by the present invention (BDGI) improves up to 1.89 dB, 3.44 dB, and 3.54 dB in PSNR, respectively, and up to 37%, 100%, and 109% in SSIM, respectively. showed.

Figures 9 and 10 show black-and-white scale ghost images obtained through TGI, DGI, UWDGI, CGI, and the method proposed by the present invention (BDGI) according to the accumulated data number.

Unlike when acquiring a binary image, when generating the mask (M), a binary mask was not created through a threshold, but the posterior probability itself was used as a black-and-white scale mask. All hyperparameters were set to 0.5. Figure 9 shows a case where artificial white Gaussian noise is not included, and Figure 10 shows a case where it is included. In both cases, the method proposed by the present invention (BDGI) displayed the “cameraman” target more clearly. In particular, while ghost images obtained through existing methods in a noisy environment were vulnerable to noise to the extent that the object could not be identified, the method proposed by the present invention (BDGI) acquired the object relatively well even in a noisy environment. You can check that.

Figure 11 shows quantitative analysis results for the results shown in Figures 9 and 10.

Among existing ghost imaging methods, the CGI method showed relatively good performance for objects expressed in binary form, but was very weak for objects in black and white scale. When using the same number of data (20,000), the method proposed by the present invention (BDGI) is up to 2.0 dB and 2.5 dB in terms of PSNR, respectively, and up to 14% and 65% in terms of SSIM, respectively, when there is no noise compared to existing methods (20,000). showed an improvement of %.

As above, specific parts of the present invention have been described in detail, and those skilled in the art will understand that these specific techniques are merely preferred embodiments and do not limit the scope of the present invention. It will be obvious. Accordingly, the actual scope of the present invention will be defined by the appended claims and their equivalents.

110: Light source unit
120: Beam splitter
130: optical sensor
140: optical sensor
150: object
160: objective arm
170: reference arm
180: image processing device

Claims (16)

Each performed by an image processing device,
Obtaining a first ghost image using a ghost image acquisition method to obtain a ghost image using the spatial correlation of the two signals consisting of the light intensity that sequentially measures the light coming out of the light source and reacting with the object and the beam pattern of the light source step;
Removing noise from the acquired first ghost image using a Markov random field-based Bayesian noise removal model;
A second ghost image is created using a ghost image acquisition method that obtains a ghost image using the spatial correlation of the two signals consisting of the light intensity that sequentially measures the light coming out of the light source and reacting with the object and the beam pattern of the light source. acquiring; and
Removing noise from the second ghost image using a Markov random field-based Bayesian noise removal model using the first ghost image from which the noise has been removed as prior information,
An image acquisition method, wherein the beam pattern of the light source is a beam pattern at the time of measuring the intensity of light through an optical sensor. According to paragraph 1,
An image acquisition method, wherein the optical sensor is at least one of a bucket detector, a photodiode-based optical sensor, and a pixel-based camera. According to paragraph 1,
Each performed by an image processing device,
A ghost image acquisition method is used to obtain a ghost image using the spatial correlation of the two signals consisting of the intensity of the light that comes out of the light source and reacts with the object sequentially through the optical sensor and the beam pattern of the light source. Obtaining a third ghost image; and
An image acquisition method further comprising removing noise from the third ghost image using a Markov random field-based Bayesian noise removal model using the second ghost image from which the noise has been removed as prior information. According to paragraph 1,
An image acquisition method, wherein the Markov random field is a hidden Markov random field. According to paragraph 1,
The ghost image acquisition method is a method of acquiring a ghost image by accumulating the beam pattern weighted by the light intensity. The method of claim 1, wherein the step of removing noise from the second ghost image using a Markov random field-based Bayesian noise removal model using the first ghost image from which the noise has been removed as prior information is performed by an image processing device. ,
Removing noise from the second ghost image by using the first ghost image from which the noise has been removed, performed by an image processing device, as prior information for modeling a joint probability distribution of the Markov random field-based Bayesian noise removal model. In,method of image acquisition. The method of claim 1, wherein the step of removing noise from the second ghost image using a Markov random field-based Bayesian noise removal model using the first ghost image from which the noise has been removed as prior information is performed by an image processing device. ,
Each performed by an image processing device,
Converting the second ghost image into a binary image or a black-and-white scale image;
Performing Gibbs sampling using the first ghost image;
generating a mask by calculating a posterior probability using the Gibbs sampling;
An image acquisition method comprising removing noise from the second ghost image converted to a binary image or a black-and-white scale image using the generated mask. According to paragraph 1,
An image acquisition method wherein the beam pattern of the light source is measured by another optical sensor or is predetermined. An optical sensor that measures the intensity of light; and
Includes a processor,
The processor,
A first ghost image acquisition method is used to obtain a ghost image using the spatial correlation of two signals consisting of the intensity of light sequentially measured from the light coming from the light source and passing through the object through the optical sensor and the beam pattern of the light source. Acquire a ghost image,
Noise is removed from the obtained first ghost image using a Markov random field-based Bayesian noise removal model,
A ghost image acquisition method is used to obtain a ghost image using the spatial correlation of the two signals consisting of the intensity of the light that comes out of the light source and reacts with the object sequentially through the optical sensor and the beam pattern of the light source. Acquire a second ghost image, and
Removing noise from the second ghost image using a Markov random field-based Bayesian noise removal model using the first ghost image from which the noise was removed as prior information,
The beam pattern of the light source is a beam pattern at the time of measuring the intensity of light through the optical sensor. According to clause 9,
The optical sensor is at least one of a bucket detector, a photodiode-based optical sensor, and a pixel-based camera. The method of claim 9, wherein the processor:
A ghost image acquisition method is used to obtain a ghost image using the spatial correlation of the two signals consisting of the intensity of the light that comes out of the light source and reacts with the object sequentially through the optical sensor and the beam pattern of the light source. Acquire a third ghost image,
An image acquisition device that removes noise from the third ghost image using a Markov random field-based Bayesian noise removal model using the second ghost image from which the noise has been removed as prior information. According to clause 9,
The Markov random field is a hidden Markov random field. According to clause 9,
The ghost image acquisition method is a method of acquiring a ghost image by accumulating the beam pattern weighted by the light intensity. The method of claim 9, wherein the processor:
An image acquisition device that removes noise from the second ghost image by using the first ghost image from which the noise has been removed as prior information for modeling a joint probability distribution of the Markov random field-based Bayesian noise removal model. The method of claim 9, wherein the processor:
Converting the second ghost image into a binary image or a black-and-white scale image,
Perform Gibbs sampling using the first ghost image,
A mask is generated by calculating the posterior probability using the Gibbs sampling,
An image acquisition device that further removes noise from the second ghost image converted to a binary image or a black-and-white scale image using the generated mask. According to clause 9,
An image acquisition device wherein the beam pattern of the light source is measured by another optical sensor or is predetermined.