Noise reduction device and noise reduction method

Abstract

To provide a technique and a noise reduction device that can improve noise elimination performance while suppressing signal degradation.SOLUTION: A noise reduction device 10 includes: an input unit 11 that acquires an input signal; a first processing unit 12 that generates, by using a first filter that attenuates a high frequency region and the input signal, a first signal; a second processing unit 13 that applies a noise reduction filter to a first signal to remove noise and generates a second signal; a third processing unit 14 that generates, by using a second filter that emphasizes a high frequency region and the second signal, a third signal; and an output unit 15 that outputs the third signal.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] The present disclosure relates to a noise reduction device and a noise reduction method. [Background technology]

[0002] There are known techniques for removing noise from signals such as images. For example, when removing noise from an image, the structure of the image is smoothed, which may cause degradation of the image. Specifically, in the noise removal process, edges of structural parts other than noise in the image tend to become rounded, grooves tend to be filled, and sharpness tends to be lost. In other words, the noise removal process removes not only noise but also structural parts in the image.

[0003] Patent Document 1 describes a technique for removing noise while suppressing a decrease in image resolution. The technique described in Patent Document 1 controls the mixing ratio of a smoothed image and a sharpened image to output a mixed image. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Patent No. 4755087 [Non-patent literature]

[0005] [Non-Patent Document 1] Zhang K, Zuo W, Chen Y, Meng D, Zhang L. Beyond a Gaussian Denoiser:Residual Learning of Deep CNN for Image Denoising. IEEE Transactions on ImageProcessing 2017; 26:3142-3155. [Non-Patent Document 2] Wang Z, Bovik AC, Sheikh HR, Simoncelli EP. Image qualityassessment: from error visibility to structural similarity. IEEE Transactionson Image Processing 2004; 13:600-612. [Non-Patent Document 3] Dabov K, Foi A, Katkovnik V, et al. Image denoising by sparse 3-D transform-domain collaborative filtering. IEEE Trans Image Process 2007; 16:2080-2095. [Non-Patent Document 4] Gu S, Xie Q, Meng D, et al. Weighted nuclear norm minimization and its applications to low level vision. Int J Comput Vis 2017; 121:183-208. Summary of the Invention [Problem to be solved by the invention]

[0006] There is a demand for suppressing signal degradation and further improving noise removal performance. An object of the present disclosure is to provide a technique capable of improving noise removal performance while suppressing signal degradation. [Means for solving the problem]

[0007] A noise removal device according to one aspect of the present disclosure includes an input unit that acquires an input signal, a first processing unit that generates a first signal using a first filter that attenuates high frequency ranges and the input signal, a second processing unit that applies a noise removal filter that removes noise to the first signal to generate a second signal, a second filter that emphasizes high frequency ranges and a third processing unit that generates a third signal using the second signal, and an output unit that outputs the third signal.

[0008] A noise removal method according to one aspect of the present disclosure includes the steps of acquiring an input signal, generating a first signal using a first filter that attenuates high frequency ranges and the input signal, applying a noise removal filter that removes noise to the first signal to generate a second signal, generating a third signal using a second filter that emphasizes high frequency ranges and the second signal, and outputting the third signal.

[0009] In the noise removal device and the noise removal method of the present disclosure, a first filter is applied to an acquired input signal to generate a first signal. In the first signal, the noise in the signal is reduced by reducing the amount of noise in the high frequency range. A noise removal filter is applied to the first signal to generate a second signal. In noise removal, the greater the amount of noise, the greater the effect of noise removal and the greater the degree of degradation of the signal. In the present disclosure, the noise in the signal is reduced by reducing the amount of noise in the high frequency range in the first signal, so the effect of noise removal by the noise removal filter is reduced. This suppresses signal degradation caused by the noise removal process. In addition, since the noise removal process is applied to the first signal whose edge sharpness has already been reduced, the degree of reduction in edge sharpness caused by the noise removal process is reduced. Then, a second filter is applied to the second signal to generate a third signal, which is output. In the third signal, the edges of the signal are restored. As a result, it is possible to improve the noise removal performance while suppressing signal degradation.

[0010] The second filter may have a frequency characteristic that is the reciprocal of the amplitude ratio of the output signal to the input signal of the first filter. In this case, the first filter and the second filter have inverse frequency characteristics. That is, the high frequency range attenuated by the first filter when the first signal is generated is accurately restored by the second filter.

[0011] The first processing unit may generate the first signal by performing a Fourier transform on the input signal, applying a first filter in Fourier transform space, and performing an inverse Fourier transform. The third processing unit may generate the third signal by performing a Fourier transform on the second signal, applying a second filter in Fourier transform space, and performing an inverse Fourier transform. In this case, the first filter and the second filter are applied in the Fourier transform space. This can reduce the calculation cost associated with applying the first filter and the second filter.

[0012] The noise removal filter may be a deep learning model generated by deep learning using a first data set, which is a plurality of signal data that does not include noise, and a second data set, which is a plurality of signal data that corresponds to the first data set and has a non-uniform amount of noise. In this case, a noise removal filter trained with various amounts of noise is used. This makes it possible to remove noise according to the amount of noise without specializing in a specific amount of noise. As a result, it is possible to further improve the noise removal performance while suppressing image degradation. Effect of the Invention

[0013] According to the present disclosure, it is possible to provide a technique capable of improving noise removal performance while suppressing signal degradation. [Brief description of the drawings]

[0014] [Figure 1] FIG. 1 is a block diagram illustrating a configuration of a noise reduction system. [Diagram 2] FIG. 2 is a schematic diagram showing an example of the operation of the noise elimination device. [Diagram 3] FIGS. 3(a) to 3(c) are graphs showing the relationship between the filter coefficient value and the evaluation index for each amount of noise. [Figure 4] FIGS. 4(a) to 4(c) are graphs showing the PSNR of the present method and the PSNR of the conventional method when optimal filter coefficient values ​​are used. [Diagram 5]FIG. 5(a) to (g) are diagrams for explaining an example of the result of noise removal. [Figure 6] FIG. 6(a) to (h) are diagrams for explaining another example of the result of noise removal. [Figure 7] 7(a) and (b) are schematic diagrams for illustrating a comparison between the conventional method and the present method in terms of structural preservation. [Figure 8] FIG. 8 is a diagram illustrating an example of a hardware configuration related to the noise reduction system. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0015] Hereinafter, embodiments of the present disclosure will be described in detail with reference to the accompanying drawings. In the description of the drawings, the same elements are given the same reference numerals and duplicated explanations will be omitted.

[0016] The noise removal device of the present disclosure removes noise contained in an input signal. Examples of the input signal include, but are not limited to, an image, which is a two-dimensional or higher signal, and a voice or an electrocardiogram, which is a one-dimensional signal. In the present disclosure, an image signal is used as an example of the input signal for explanation. The type of image and noise is not limited. The noise removal device of the present disclosure performs three stages of processing on an image. As a first process, the noise removal device performs a pre-processing to attenuate the high frequency range in the image. The pre-processing reduces the apparent amount of noise in the image. As a second process, the noise removal device performs a noise removal process on the image with the reduced amount of noise. As a third process, the noise removal device performs a post-processing to restore the high frequency range in the image. The post-processing restores the sharpness of degraded parts in the image.

[0017] 1 is a block diagram illustrating a configuration of a noise reduction system 1 according to an embodiment. The noise reduction system 1 includes a noise reduction device 10, an image capture device 20, a storage device 30, and a terminal 40. The noise reduction device 10 is connected to the image capture device 20, the storage device 30, and the terminal 40 so as to be able to communicate with each other.

[0018] The image capturing device 20 is, for example, but is not limited to, a camera. The image capturing device 20 outputs the captured image to the noise elimination device 10. The image capturing device 20 may be a component of the noise elimination device 10.

[0019] The storage device 30 is a non-transitory storage medium or storage device that stores various data used by the noise reduction device 10. The storage device 30 may store, for example, images containing noise and images without noise. The storage device 30 may be constructed as a single database or a collection of multiple databases. The storage device 30 may be a component of the noise reduction device 10.

[0020] The terminal 40 is a computer used by a user of the noise reduction system 1. The type and configuration of the terminal 40 are not limited. The terminal 40 may be, for example, a personal computer, a tablet terminal, a smartphone, or the like. The terminal 40 may be a component of the noise reduction device 10.

[0021] The noise elimination device 10 is a device that eliminates noise from an input signal. There are no limitations on the type and configuration of the noise elimination device 10. The noise elimination device 10 includes an input unit 11, a first processing unit 12, a second processing unit 13, a third processing unit 14, and an output unit 15 as functional elements.

[0022] The input unit 11 acquires an input signal. The input unit 11 acquires an image including noise from, for example, the image capturing device 20, the storage device 30, or the terminal 40.

[0023] The first processing unit 12 generates a first signal using a first filter that attenuates a high frequency range and an input signal. The first filter is, for example, a low-pass filter that passes a low frequency range and attenuates a high frequency range. For example, the first processing unit 12 may generate the first signal by applying the first filter to the input signal. The first processing unit 12 may generate the first signal by Fourier transforming the input signal, applying the first filter in Fourier transform space (frequency space), and performing an inverse Fourier transform.

[0024] The second processing unit 13 applies a noise removal filter that removes noise to the first signal to generate a second signal. The noise removal filter is, for example, DnCNN (Denoising Convolutional Neural Network). DnCNN is a method of noise removal using deep learning (see Non-Patent Document 1). The noise removal filter may be a blind DnCNN (hereinafter referred to as "BDnCNN") that is trained to remove noise according to the amount of noise. For example, BDnCNN is a deep learning model generated by deep learning using a first data set that is a plurality of signal data that does not contain noise, and a second data set that corresponds to the first data set and is a plurality of signal data with a non-uniform amount of noise. The noise removal filter may be a non-learning type method.

[0025] The third processing unit 14 generates a third signal using a second filter that emphasizes a high frequency range and the second signal. The second filter is a high frequency emphasis filter that emphasizes a high frequency range. For example, the third processing unit 14 may generate the third signal by applying the second filter to the second signal. The third processing unit 14 may generate the third signal by Fourier transforming the second signal, applying a high frequency emphasis filter in the Fourier transform space, and performing an inverse Fourier transform.

[0026] The output unit 15 outputs a third signal. The third signal can be called an output signal. The output unit 15 may display the third image on a display device of the noise removal device 10, for example. The output unit 15 may transmit the third image to the terminal 40, for example.

[0027] [Noise reduction system operation] An example of an operation method (noise removal method) by the noise removal device 10 will be described with reference to Fig. 2. Fig. 2 is a schematic diagram showing an example of the operation of the noise removal device 10.

[0028] In step S1, the input unit 11 acquires a noisy image I as an input signal. The input unit 11 acquires the noisy image I from, for example, the imaging device 20, the storage device 30, or the terminal 40. For example, the noisy image I is an image in which noise having properties similar to white Gaussian noise is added to an MRI (Magnetic Resonance Imaging) image of the brain.

[0029] In step S2, the first processing unit 12 performs a Fourier transform on the noise image I. If the Fourier transform is denoted as FT, the process of step S2 can be expressed as FT[I].

[0030] In step S3, the first processing unit 12 applies a low-pass filter L as a first filter in the Fourier transform space. That is, the first processing unit 12 applies the low-pass filter L to the Fourier transform signal FT[I]. The process of step S3 can be expressed as L FT[I].

[0031] In step S4, the first processing unit 12 performs an inverse Fourier transform on the signal to which the low-pass filter L has been applied. If the inverse Fourier transform is IFT, the process of step S4 can be expressed as IFT[L FT[I]]. By the inverse Fourier transform, the first image I L is generated.

[0032] In this way, the first processing unit 12 uses the low-pass filter L and the noise image I to generate the first image I L The first image I L is an image in which the amount of noise appears to be suppressed more than in the noisy image I. The processes in steps S2 to S4 can be rephrased as pre-processing. The pre-processing can be expressed by the following equation (1).

[0033]

number

[0034] In step S5, the second processing unit 13 applies a noise reduction filter to the first image I LApply to the second image I LD The noise reduction filter may be, for example, BDnCNN. The first image I L Since the noise removal filter is applied to the noisy image I, the effect of noise removal is suppressed compared to the case where the noise removal filter is applied to the noisy image I.

[0035] In step S6, the third processing unit 14 processes the second image I LD The process in step S6 is to Fourier transform FT[I LD ] can be expressed as follows.

[0036] In step S7, the third processing unit 14 applies a high-frequency emphasis filter Q as a second filter in the Fourier transform space. That is, the third processing unit 14 applies a high-frequency emphasis filter Q as a second filter in the Fourier transform space. LD ] is applied with a high-frequency emphasis filter Q. The process in step S7 is LD ] can be expressed as follows.

[0037] The second filter has a frequency characteristic that is the reciprocal of the amplitude ratio of the output signal to the input signal of the first filter. For example, the low-pass filter L used by the first processing unit 12 and the high-frequency emphasis filter Q used by the third processing unit 14 have a relationship of Q=1 / L.

[0038] In step S8, the third processing unit 14 performs an inverse Fourier transform on the signal to which the high-frequency emphasis filter Q has been applied. LD By inverse Fourier transform, the noise-removed image I D is generated.

[0039] In this way, the third processing unit 14 processes the high-frequency emphasis filter Q and the second image I LD Using this, the noise-removed image I D Generate the denoised image I D is the second image I LDThis is an image in which signals in the high frequency range are restored, rather than the image in the normal range. The processes in steps S6 to S8 can be rephrased as post-processing. The post-processing can be expressed by the following equation (2).

[0040]

number

[0041] In step S9, the output unit 15 outputs the noise-removed image I D The output unit 15 outputs the noise-removed image I on, for example, a display device of the noise removal device 10. D The output unit 15 may display the noise-removed image I on, for example, a terminal 40. D may be transmitted.

[0042] [Example] Hereinafter, an embodiment using BDnCNN as a noise reduction filter will be described. In the embodiment, the frequency characteristic of the low-pass filter L is expressed by the following formula (3), and the frequency characteristic of the high-frequency emphasis filter Q is expressed by the following formula (4).

[0043]

number

[0044]

number

[0045] Here, Nx, Ny are the two-dimensional number of pixels of the image. i, j are data indices in the Fourier transform space of the image (i≦Nx, j≦Ny). c is a coefficient that adjusts the strength of the filter characteristics (hereinafter referred to as the "filter coefficient"). When the filter coefficient c=0, L(i,j)=1 and Q(i,j)=1, so the low-pass filter L and the high-frequency emphasis filter Q are not actually applied.

[0046] Figures 3(a) to (c) are graphs showing the relationship between the value of the filter coefficient c and the evaluation index for each noise level. Figure 3(a) is a graph showing the relationship between the value of the filter coefficient c and the evaluation index when the noise level is 2.5%. Figure 3(b) is a graph showing the relationship between the value of the filter coefficient c and the evaluation index when the noise level is 5.0%. Figure 3(c) is a graph showing the relationship between the value of the filter coefficient c and the evaluation index when the noise level is 7.5%.

[0047] In the embodiment, the peak signal-to-noise ratio (PSNR) and the structural similarity index measure (SSIM) were used as the evaluation indexes. The PSNR is the ratio of the error standard deviation σ to the maximum luminance value Imax of the image, expressed as 20log 10 (Imax / σ). SSIM is an index for evaluating the structural preservation of an image (see Non-Patent Document 2).

[0048] In (a) to (c) of Fig. 3, the horizontal axis indicates the value of the filter coefficient c (0≦c≦2), the vertical axis on the left indicates PSNR [dB], and the vertical axis on the right indicates SSIM. In (a) to (c) of Fig. 3, the PSNR values ​​are plotted with circles, and the SSIM values ​​are plotted with squares.

[0049] As shown in (a) and (b) of FIG. 3, improvements in PSNR and SSIM were observed depending on the value of the filter coefficient c. That is, by applying a low-pass filter L and a high-frequency emphasis filter Q before and after the noise reduction filter BDnCNN, PSNR and SSIM are improved. In particular, when the amount of noise is small, 2.5% (see (a) of FIG. 3), the degree of improvement in PSNR and SSIM is large. Based on the results shown in (a) to (c) of FIG. 3, the optimal value of the filter coefficient c in the embodiment is set to c=0.75, which will be described below.

[0050] FIG. 4 (a) to (c) are graphs showing the PSNR of the present method (Example) and the PSNR of the conventional method (Comparative Example) when the optimal filter coefficient c value (c=0.75) is used. In the conventional method, a conventional DnCNN trained with a specific noise amount is applied to a noisy image. FIG. 4 (a) is a graph showing the PSNR of the present method and the PSNR of the conventional method when the noise amount is 2.5%. FIG. 4 (b) is a graph showing the PSNR of the present method and the PSNR of the conventional method when the noise amount is 5.0%. FIG. 4 (c) is a graph showing a comparison of the PSNR of the present method and the PSNR of the conventional method when the noise amount is 7.5%.

[0051] In (a) to (c) of Fig. 4, the vertical axis indicates the PSNR value. As shown in (a) to (c) of Fig. 4, an improvement in PSNR was observed in the proposed method compared to the conventional method.

[0052] Fig. 5(a) to (g) are diagrams for explaining an example of the result of noise removal. Fig. 5(a) is an example of an MRI image. Fig. 5(b) is a partial enlarged view of range R1 in Fig. 5(a). Fig. 5(a) and (b) are images (target images) to which no noise has been added. Fig. 5(c) is a noisy image in which 5.0% noise has been added to the image in Fig. 5(b).

[0053] Fig. 5(d) is a noise-removed image obtained by applying the conventional DnCNN to Fig. 5(c). Fig. 5(e) is a noise-removed image obtained by applying the BDnCNN of the embodiment when the filter coefficient c=0.50. Fig. 5(f) is a noise-removed image obtained by applying the BDnCNN of the embodiment when the filter coefficient c=0.75. Fig. 5(g) is a noise-removed image obtained by applying the BDnCNN of the embodiment when the filter coefficient c=1.0.

[0054] The noise-removed image by the conventional method in (d) of Figure 5 is smoothed to a greater extent and the brain structure is not as clear as the target image in (b) of Figure 5. The noise-removed image by the present method (c=0.75) in (f) of Figure 5 is a sharp image that retains a structure closer to the target image in (b) of Figure 5. In addition, the noise-removed image by the present method (c=0.5) in (e) of Figure 5 and the noise-removed image by the present method (c=1.0) in (g) of Figure 5 can also be said to be sharp images close to the target image in (b) of Figure 5.

[0055] FIG. 6(a) to FIG. 6(h) are diagrams for explaining another example of the result of noise removal. FIG. 6(a) is another example of an MRI image. FIG. 6(b) is a noisy image in which 5.0% noise is added to the image of FIG. 6(a). FIG. 6(c) is a noise-removed image in which the conventional DnCNN is applied. FIG. 6(d) is a noise-removed image in which the BDnCNN of the embodiment is applied. FIG. 6(e) is a partial enlarged view of range R2 in FIG. 6(a). FIG. 6(f) is a partial enlarged view of range R3 in FIG. 6(b). FIG. 6(g) is a partial enlarged view of range R4 in FIG. 6(c). FIG. 6(h) is a partial enlarged view of range R5 in FIG. 6(d).

[0056] The noise-removed image using the conventional method (g) in Fig. 6 is smoothed to a greater extent and the brain structure is not as clear as the target image (e) in Fig. 6. The noise-removed image using the present method (h) in Fig. 6 is a sharper image that retains a structure closer to that of the target image (e) in Fig. 6.

[0057] 7(a) and (b) are schematic diagrams for explaining a comparison between the conventional method and the present method in terms of structural preservation. FIG. 7(a) is a schematic diagram showing the noise removal process of the conventional method. FIG. 7(b) is a schematic diagram showing the noise removal process of the present method. FIG. 7(a) and (b) show signal values ​​with continuous lines and noise with discontinuous lines.

[0058] FIG. 7(a) shows the generation of a noise-removed image P10 according to a comparative example by applying a noise removal filter to an input signal P containing noise. In noise removal processing, generally, the greater the amount of noise, the greater the noise removal effect. In noise removal processing, structural parts other than noise in the signal are also degraded. In the noise-removed image P10, the degree of smoothing increases according to the amount of noise, so that, for example, the sharpness of the edges of the image is significantly degraded.

[0059] FIG. 7(b) shows a process in which a first filter is applied to an input signal P containing noise to generate a first signal P1, a second signal P2 is generated using the first signal P1 and a noise reduction filter, and a third signal P3 is generated by applying a second filter to the second signal P2. In the first signal P1, the amount of noise in the high frequency range is reduced compared to the input signal P, thereby suppressing the noise in the signal. At this time, for example, the sharpness of the edge of the image is also deteriorated. Since the noise in the signal is suppressed by reducing the amount of noise in the high frequency range, the noise removal effect of the noise removal filter that generates the second signal P2 can be suppressed. In addition, since the sharpness of the edge has already deteriorated in the first signal P1, the degree of deterioration of the sharpness of the edge caused by the noise removal process is reduced. In the third signal P3, for example, the sharpness of the edge of the image is restored. At this time, the noise is also amplified at the same time, but when the degree of noise amplification and the degree of image restoration are compared, the effect of the degree of image restoration is high, so the PSNR and SSIM are high in the overall evaluation.

[0060] [Hardware configuration] Fig. 8 is a diagram showing an example of a hardware configuration related to the noise reduction system 1. Fig. 8 shows a computer 100 functioning as the noise reduction device 10. The computer 100 has a processor 101, a main memory unit 102, an auxiliary memory unit 103, a communication control unit 104, an input device 105, and an output device 106. The noise reduction device 10 is composed of one or more computers 100 composed of the above hardware and software such as a program.

[0061] When the noise elimination device 10 is composed of multiple computers 100, these computers 100 may be connected locally or via a communication network such as the Internet or an intranet. This connection logically constitutes one noise elimination device 10.

[0062] The processor 101 is a central processing unit (CPU) that executes an operating system, application programs, etc. The main memory 102 is composed of a read-only memory (ROM) and a random access memory (RAM). The auxiliary memory 103 is a storage medium composed of a hard disk, a flash memory, etc. The auxiliary memory 103 generally stores a larger amount of data than the main memory 102. The communication control unit 104 is composed of a network card or a wireless communication module. At least a part of the communication function with other devices in the noise elimination device 10 may be realized by the communication control unit 104. The input device 105 is composed of a keyboard, a mouse, a touch panel, a microphone for voice input, etc. The output device 106 is composed of a display, a printer, etc.

[0063] The auxiliary storage unit 103 stores in advance a program 110 (noise removal program) and data necessary for processing. The program 110 causes the computer 100 to execute each functional element of the noise removal device 10. The program 110 causes, for example, processing related to the above-mentioned noise removal method to be executed in the computer 100. For example, the program 110 is read by the processor 101 or the main storage unit 102, and causes at least one of the processor 101, the main storage unit 102, the auxiliary storage unit 103, the communication control unit 104, the input device 105, and the output device 106 to operate. For example, the program 110 reads and writes data in the main storage unit 102 and the auxiliary storage unit 103.

[0064] The program 110 may be provided in a form recorded on a tangible storage medium such as a CD-ROM, a DVD-ROM, a semiconductor memory, etc. The program 110 may be provided as a data signal via a communication network.

[0065] [effect] As described above, a noise removal device 10 according to one aspect of the present disclosure includes an input unit 11 that acquires an input signal, a first processing unit 12 that generates a first signal using a first filter that attenuates high frequency ranges and the input signal, a second processing unit 13 that applies a noise removal filter that removes noise to the first signal and generates a second signal, a second filter that emphasizes high frequency ranges and a third processing unit 14 that generates a third signal using the second signal, and an output unit 15 that outputs the third signal.

[0066] A noise removal method according to one aspect of the present disclosure includes the steps of acquiring an input signal, generating a first signal using a first filter that attenuates high frequency ranges and the input signal, applying a noise removal filter that removes noise to the first signal to generate a second signal, generating a third signal using a second filter that emphasizes high frequency ranges and the second signal, and outputting the third signal.

[0067] In the noise removal device 10 and the noise removal method of the present disclosure, a first filter is applied to an acquired input signal to generate a first signal. In the first signal, the noise in the signal is reduced by reducing the amount of noise in the high frequency range. A noise removal filter is applied to the first signal to generate a second signal. In noise removal, the greater the amount of noise, the greater the effect of noise removal and the greater the degree of degradation of the signal. In the present disclosure, the noise in the signal is reduced by reducing the amount of noise in the high frequency range in the first signal, so the effect of noise removal by the noise removal filter is reduced. This suppresses signal degradation caused by the noise removal process. In addition, since the noise removal process is applied to the first signal whose edge sharpness has already been reduced, the degree of reduction in edge sharpness caused by the noise removal process is reduced. Then, a second filter is applied to the second signal to generate a third signal, which is output. In the third signal, the edges of the signal are restored. As a result, it is possible to improve the noise removal performance while suppressing signal degradation.

[0068] The second filter has a frequency characteristic that is the inverse of the amplitude ratio of the output signal to the input signal of the first filter. In this case, the first filter and the second filter have inverse frequency characteristics. That is, the high frequency range attenuated by the first filter when the first signal is generated is accurately restored by the second filter.

[0069] The first processing unit 12 performs a Fourier transform on the input signal, applies a first filter in the Fourier transform space, and performs an inverse Fourier transform to generate a first signal. The third processing unit 14 performs a Fourier transform on the second signal, applies a second filter in the Fourier transform space, and performs an inverse Fourier transform to generate a third signal. In this case, the first filter and the second filter are applied in the Fourier transform space. This can reduce the calculation cost associated with the application of the first filter and the second filter.

[0070] The noise removal filter is a deep learning model generated by deep learning using a first data set, which is a plurality of signal data that does not contain noise, and a second data set, which is a plurality of signal data that corresponds to the first data set and has a non-uniform amount of noise. In this case, a noise removal filter trained with various amounts of noise is used. This makes it possible to remove noise according to the amount of noise without specializing in a specific amount of noise. As a result, it is possible to further improve the noise removal performance while suppressing image degradation.

[0071] [Variations] The present disclosure is not necessarily limited to the above-described embodiments, and various modifications are possible without departing from the spirit and scope of the present disclosure.

[0072] In the above embodiment, the case where the noise removal filter is DnCNN or BDnCNN using deep learning has been described, but is not limited thereto. For example, the noise removal filter may be BM3D (Block Matching 3D filtering) (see Non-Patent Document 3). The noise removal filter may be WNNM (Weighted Nuclear Norm Minimization) (see Non-Patent Document 4). Even in these cases, it is possible to improve the noise removal performance while suppressing image degradation.

[0073] In the above embodiment, the functions shown in Equation (3) and Equation (4) are described as low-pass filters, but the low-pass filters are not limited to these. For example, the low-pass filters may be a Hamming window or a Gaussian function.

[0074] [Note] The following is a summary of the present disclosure. [1] An input unit for acquiring an input signal; a first filter that attenuates a high frequency range and a first processing unit that generates a first signal using the input signal; a second processing unit that applies a noise reduction filter to the first signal to reduce noise and generate a second signal; a second filter that emphasizes the high frequency range, and a third processing unit that generates a third signal using the second signal; and an output unit that outputs the third signal. Noise canceller. [2] The noise removal device according to [1], wherein the second filter has a frequency characteristic that is the inverse of an amplitude ratio of an output signal to an input signal of the first filter. [3] The first processing unit generates the first signal by performing a Fourier transform on the input signal, applying the first filter in a Fourier transform space, and performing an inverse Fourier transform; The third processing unit generates the third signal by performing a Fourier transform on the second signal, applying the second filter in a Fourier transform space, and performing an inverse Fourier transform. A noise reduction device according to [1] or [2]. [4] The noise removal filter is a deep learning model generated by deep learning using a first data set which is a plurality of signal data that does not contain noise, and a second data set which is a plurality of signal data that corresponds to the first data set and has a non-uniform amount of noise. The noise removal device according to any one of [1] to [3]. [5] obtaining an input signal; generating a first signal using a first filter that attenuates high frequencies and the input signal; applying a noise reduction filter to the first signal to reduce noise to generate a second signal; generating a third signal using a second filter that emphasizes the high frequency range and the second signal; and outputting the third signal. Noise removal methods. [Explanation of symbols]

[0075] 1... noise elimination system, 10... noise elimination device, 11... input section, 12... first processing section, 13... second processing section, 14... third processing section, 15... output section, 20... imaging device, 30... storage device, 40... terminal.

Claims

1. An input unit for acquiring an input signal; a first filter that attenuates a high frequency range and a first processing unit that generates a first signal using the input signal; a second processing unit that applies a noise reduction filter to the first signal to reduce noise and generate a second signal; a second filter that emphasizes the high frequency range and a third processing unit that generates a third signal using the second signal; and an output unit that outputs the third signal. Noise canceller.

2. 2. The noise removal device according to claim 1, wherein the second filter has a frequency characteristic that is an inverse of an amplitude ratio of an output signal to an input signal of the first filter.

3. The first processing unit generates the first signal by performing a Fourier transform on the input signal, applying the first filter in a Fourier transform space, and performing an inverse Fourier transform. The third processing unit generates the third signal by performing a Fourier transform on the second signal, applying the second filter in a Fourier transform space, and performing an inverse Fourier transform. The noise reduction device according to claim 1 .

4. 2. The noise removal device according to claim 1, wherein the noise removal filter is a deep learning model generated by deep learning using a first data set that is a plurality of signal data that does not contain noise, and a second data set that corresponds to the first data set and is a plurality of signal data with non-uniform amounts of noise.

5. obtaining an input signal; generating a first signal using a first filter that attenuates high frequencies and the input signal; applying a noise reduction filter to the first signal to reduce noise to generate a second signal; generating a third signal using a second filter that emphasizes the high frequency range and the second signal; and outputting the third signal. Noise removal methods.

Images