JP7202578B1 - Image processing device, image processing method, program and storage medium - Google Patents

Abstract

Kind Code: A1 A brightness dynamic range is expanded while maintaining the image quality of bright and dark areas in an image.
Kind Code: A1 An image processing apparatus includes a pixel extraction unit for extracting image pixels from a color original image including a plurality of image pixels, and a conversion unit for converting an RGB color space of each image pixel into an HSV color space. a noise applying unit 14 for applying applied noise having an intensity corresponding to the brightness of each image pixel to the V component of the image pixel; and comparing the V component of each image pixel to which the applied noise is applied with a predetermined threshold. It includes a threshold application unit 16 that outputs a result, and a restored image generation unit 18 that generates a restored image from the comparison result.
[Selection drawing] Fig. 6

Description

Article 30, Paragraph 2 of the Patent Act applies Kanazawa Institute of Technology 2020 Proceedings of the Project Design III presentation review meeting "Project Design III Project Report Research on image processing to reduce the effect of light distribution" 2021 1 25th of the month

The present invention relates to an image processing device, an image processing method, a program, and a storage medium.

A stochastic resonance phenomenon is known as a technique for extracting a hidden signal by applying noise to a weak signal buried in noise (see, for example, Non-Patent Document 1).

Koji Tanaka: "Fundamental Concept of Stochastic Resonance and Its Application", Osaka University of Commerce, Vol. 5 (No. 1), 2009, pp. 379-391 Noriko Kojima et al.: "Proposal of Autotuning Image Enhancement Method Using Stochastic Resonance", The Institute of Electrical Engineers of Japan Transactions C, Vol. 138 (No. 11), 2018, pp. 1425-1434 Raji Kumar Jha et al. : “Dark and low-contrast image enhancement using dynamic stochastic resonance in discrete cosine transform domain,” APSIPA Transactions on Signal and Information, 1-6, Processing, Vol. J. J. Collins et al. : "Stochastic resonance without tuning," Nature, Vol. 376, 1995, pp. 236-238

In order to improve the visibility of images in which bright and dark areas coexist, it is necessary to increase the brightness of dark areas (hereinafter also referred to as "brightness") to expand the dynamic range of brightness. Become. Conventionally, in order to prevent deterioration of image quality in bright areas, locally different image processing has been performed, or a single image has been generated from a plurality of images. In this case, since image processing such as DCT (Discrete Cosine Transform) is performed for each block containing a plurality of pixels, there is a problem such as the occurrence of a halo effect due to the influence of peripheral information in the image.

As an image enhancement method, a method using a stochastic resonance phenomenon, which will be described later, has been proposed (see, for example, Non-Patent Documents 2 and 3).

In the technique described in Non-Patent Document 2, the degree of enhancement of an image is adjusted by determining the number of elements based on the average value and median value of the image. At this time, the relationship between the noise intensity and the threshold is constant. In addition, since a grayscale image is used as an object of the experiment, there is a problem that application to a color image cannot be realized.

Non-Patent Document 3 proposes an algorithm that applies stochastic resonance of a bistable system to enhance a dark, low-contrast image. The algorithm adjusts image enhancement by using a quantitative performance metric to determine the number of iterations. In this case, only internal noise is used without external noise application. At this time, DCT is used to determine parameters according to internal noise. This algorithm can achieve contrast restoration for color images as well. However, since spatial processing using DCT is required, there is a problem that, in principle, deterioration of image quality due to determination of the optimum block size and spatial processing cannot be avoided.

SUMMARY OF THE INVENTION An object of the present invention is to expand the dynamic range of lightness while maintaining the image quality of bright and dark areas in an image.

In order to solve the above problems, an image processing apparatus according to one aspect of the present invention includes a pixel extracting unit that extracts image pixels from a color original image that includes a plurality of image pixels; a transforming unit for transforming into space; a noise applying unit for applying applied noise having an intensity corresponding to the brightness of each image pixel to the V component of the image pixel; A threshold applying unit that compares with a threshold and outputs a comparison result, and a restored image generating unit that generates a restored image from the comparison result.

According to this aspect, it is possible to expand the dynamic range of lightness while maintaining the image quality of the bright and dark areas in the image.

In one embodiment, the noise applicator and the threshold applicator are implemented by a summing network.

In one embodiment, the relationship between the intensity of each image pixel and the applied noise intensity is positively correlated both when the intensity of each image pixel is below the threshold and when the intensity of each image pixel is above the threshold.

In one embodiment, the relationship between the intensity of each image pixel and the applied noise intensity is negatively correlated when the intensity of each image pixel is below the threshold and when the intensity of each image pixel is above the threshold.

In one embodiment, the relationship between the intensity of each image pixel and the applied noise intensity is positively correlated when the intensity of each image pixel is below a threshold and negatively correlated when the intensity of each image pixel is above the threshold. have.

In one embodiment, the relationship between the intensity of each image pixel and the applied noise intensity is negatively correlated when the intensity of each image pixel is below a threshold and positively correlated when the intensity of each image pixel is above the threshold. have.

In one embodiment, the relationship between the intensity of each image pixel and the applied noise intensity is constant when the intensity of each image pixel is below a threshold and is positively correlated when the intensity of each image pixel is above the threshold. .

In one embodiment, the relationship between the intensity of each image pixel and the applied noise intensity is positively correlated when the intensity of each image pixel is below a threshold and is constant when the intensity of each image pixel is above the threshold. .

In one embodiment, the relationship between the intensity of each image pixel and the applied noise intensity is negatively correlated when the intensity of each image pixel is below a threshold and is constant when the intensity of each image pixel is above the threshold. .

In one embodiment, the relationship between the intensity of each image pixel and the applied noise intensity is constant when the intensity of each image pixel is below the threshold and negatively correlated when the intensity of each image pixel is above the threshold. .

An image processing apparatus according to one embodiment includes a storage unit that stores the relationship between the brightness of each image pixel and the applied noise intensity.

The image processing apparatus of one embodiment performs machine learning using the brightness of each image pixel, the applied noise intensity, and the restored image as learning data, and outputs the applied noise intensity when the original image is input. It has a learning unit that

An image processing apparatus according to one embodiment includes a threshold determination unit that determines a threshold based on the distribution of the V component in the HSV color space of the original image.

Another aspect of the present invention is an image processing method. This method is an image processing method for performing image processing using an image processing device including a pixel extraction unit, a conversion unit, a noise application unit, a threshold application unit, and a restored image generation unit, wherein the pixel an extraction unit for extracting image pixels from a color original image including a plurality of image pixels; a transformation unit for transforming the RGB color space of each image pixel into an HSV color space; applying an applied noise having an intensity corresponding to the brightness of the image pixel to the V component of the image pixel; and comparing the V component of each image pixel to which the applied noise is applied with a predetermined threshold in a threshold application unit. The step of outputting the result, and the step of generating a restored image from the comparison result in a restored image generation unit.

According to this aspect, the image processing device can be used to expand the dynamic range of lightness while maintaining the image quality of bright and dark areas in an image.

Yet another aspect of the present invention is a program. This program is a program that causes a computer to execute image processing using an image processing device that includes a pixel extraction unit, a conversion unit, a noise application unit, a threshold application unit, and a restored image generation unit, a pixel extraction unit extracting image pixels from a color original image including a plurality of image pixels; a transformation unit transforming the RGB color space of each image pixel into an HSV color space; applying an applied noise having an intensity corresponding to the brightness of each image pixel to the V component of the image pixel; A step of outputting a comparison result and a step of generating a restored image from the comparison result in a restored image generating section are executed by a computer.

According to this aspect, a computer can execute a program for expanding the dynamic range of lightness while maintaining the image quality of bright and dark areas in an image.

Yet another aspect of the invention is a storage medium. This storage medium stores a program for causing a computer to execute image processing using an image processing device that includes a pixel extraction unit, a conversion unit, a noise application unit, a threshold application unit, and a restored image generation unit. a storage medium, a pixel extractor for extracting image pixels from a color original image containing a plurality of image pixels; and a transform unit for converting the RGB color space of each image pixel to an HSV color space. a step of applying, in a noise application unit, applied noise having an intensity corresponding to the brightness of each image pixel to the V component of the image pixel; and outputting a comparison result, and generating a restored image from the comparison result in a restored image generation unit.

According to this aspect, it is possible to store a program for expanding the dynamic range of lightness while maintaining the image quality of bright and dark areas in the image.

Any combination of the above constituent elements, and conversion of expressions of the present invention between methods, devices, systems, etc. are also effective as aspects of the present invention. Any suitable combination of the above elements may also fall within the scope of the invention for which patent protection is sought by this patent application.

According to the present invention, it is possible to expand the dynamic range of lightness while maintaining the image quality of bright and dark areas in an image.

It is a figure which shows an input signal and an output signal when the noise intensity to apply is changed. 1 is a schematic diagram showing the configuration of a stochastic resonance phenomenon system; FIG. FIG. 4 is a diagram showing the relationship between noise intensity and correlation coefficient for a linear system and a stochastic resonance system; 1 is a schematic diagram showing a configuration of a summing network; FIG. FIG. 4 is a diagram showing the relationship between noise intensity and correlation coefficient for summing networks with different numbers of channels; 1 is a functional block diagram of an image processing device according to a first embodiment; FIG. FIG. 5 is a diagram showing an example of applied noise characteristics; 8 is a diagram showing an original image and a restored image when the applied noise characteristics of FIG. 7 are given; FIG. FIG. 4 is a diagram showing examples of various applied noise characteristics; FIG. 10 is a diagram showing a restored image corresponding to the applied noise characteristics of FIG. 9; FIG. 10 is a functional block diagram of an image processing device according to a second embodiment; FIG. FIG. 11 is a functional block diagram of an image processing apparatus according to a third embodiment; FIG. FIG. 11 is a functional block diagram of an image processing device according to a fourth embodiment; FIG. 10 is a flow chart of an image processing method according to a fifth embodiment;

BEST MODE FOR CARRYING OUT THE INVENTION The present invention will be described below based on preferred embodiments with reference to the drawings. In the embodiment and modified examples, the same or equivalent constituent elements and members are denoted by the same reference numerals, and duplication of description will be omitted as appropriate. In addition, the dimensions of the members in each drawing are appropriately enlarged or reduced for easy understanding. Also, in each drawing, some of the members that are not important for explaining the embodiments are omitted. Also, although terms containing ordinal numbers such as first, second, etc. are used to describe various components, such terms are only used to distinguish one component from another, and the terms The constituent elements are not limited by

Before describing specific embodiments, first, basic knowledge will be described. The stochastic resonance phenomenon is a phenomenon in which the response to a weak signal is improved by adding noise to the weak signal. FIG. 1 shows an input pulse signal, an output pulse signal and a threshold for a certain time-series signal. The left figure shows the input signal, ie the original input signal, and the output signal when no noise is applied. In this case, since the input signal is weak and the entire signal is below the threshold, no waveform is observed in the output signal. The upper right figure shows the input and output signals when noise of appropriate intensity (eg, Gaussian white noise) is applied to the threshold. In this case, a waveform similar to that of the input signal is observed in the output signal, indicating that the stochastic resonance phenomenon has occurred. This makes it possible to obtain information hidden in noise from weak signals. The lower right diagram shows the input and output signals when noise is applied that is too strong (ie, excessive) relative to the threshold. In this case, since the entire input signal exceeds the threshold, the correlation between the input signal and the output signal is lost. Therefore, hidden information cannot be obtained from weak signals. Thus, it can be seen that it is basically necessary to adjust the applied noise intensity to an appropriate value in order to properly realize the stochastic resonance phenomenon.

FIG. 2 schematically shows the configuration of a system 100 for realizing the stochastic resonance phenomenon. The system 100 comprises a noise application section 102 to an input signal and a threshold application section 104 . The threshold application unit 104 is configured including a nonlinear element. The term "nonlinear element" used in this specification refers to an element that outputs 0 or 1 regardless of the magnitude of the input (the same applies hereinafter). The non-linear element of the threshold application unit 104 compares the input signal (to which noise is added) and the threshold, and outputs 1 as an output signal if the input signal is larger, and outputs 0 if it is smaller.

FIG. 3 shows both using a stochastic resonance system of the type of FIG. The relationship between the noise intensity and the correlation coefficient between the input and the output (hereinafter abbreviated as "correlation coefficient") is shown. In a linear system, the correlation coefficient decreases monotonically with increasing applied noise intensity. On the other hand, in a stochastic resonance system, the correlation coefficient increases as the noise intensity is increased from 0. After that, when the noise intensity is _IP , the correlation coefficient shows the maximum value, and when the noise intensity is further increased, the correlation coefficient decreases. From this, it can be seen that for a stochastic resonance system of the type of FIG. 2, it is advantageous to adjust the noise power to be approximately _IP .

A summing network has been proposed as a technique capable of obtaining a high input-output correlation without requiring noise intensity adjustment for the stochastic resonance system of the type shown in FIG. 4). FIG. 4 schematically shows the configuration of the summing network.

The summing network of FIG. 4 includes a plurality (N) of noise application units, a threshold application unit corresponding to each noise application unit, and an addition unit. Each noise application unit and each corresponding threshold application unit function as the stochastic resonance phenomenon system of FIG. 2, and process input signals in parallel. Outputs for N channels obtained from each threshold application unit are added by an adder and normalized by N.

FIG. 5 shows the relationship between the applied noise power and the correlation coefficient when using a summing network of the type of FIG. Here, cases where the number of channels N is 10, 100, and 100 are shown. As shown in the figure, as the number of channels increases, the peak value of the correlation coefficient rises, and the decrease in the correlation coefficient accompanying the increase in noise intensity becomes gentle. Therefore, by using a summing network with a sufficient number of channels, it can be expected that high input-output correlation can be obtained without requiring adjustment of noise intensity.

[First embodiment]
FIG. 6 shows a functional block diagram of the image processing apparatus 1 according to the first embodiment. The image processing device 1 includes a pixel extraction unit 10, a conversion unit 12, a noise application unit 14, a threshold application unit 16, and a restored image generation unit 18.

A pixel extraction unit 10 extracts image pixels from a color original image including a plurality of image pixels. It is effective for the original image to have a mixture of bright areas and dark areas, but it is not limited to this. The following processing is performed for each extracted image pixel.

The conversion unit 12 converts the RGB color space of each image pixel into the HSV color space. The RGB color space is a color space consisting of three primary colors of red (R), green (G), and blue (B). In the RGB color model, colors are expressed by additively mixing these three primary colors. On the other hand, the HSV color space is a color space consisting of three components of hue (Hue), saturation (Saturation), and lightness (Value). Since the HSV color space is similar to human color recognition, it is possible to perform processing closer to human sensitivity than the RGB color space.

Conversion from the RGB color space to the HSV color space is performed by equations (1), (2), (3). Here, MIN is the minimum value in [R, G, B], and MAX is the maximum value in [R, G, B]. where [R, G, B] are in the range of 0 to 1, respectively. In the HSV color space resulting from this transformation, H ranges from 0 to 360, and S and V range from 0 to 1. Since the hue (H) is represented by a circle, H=0 and H=360 are equal.

Conversion from the HSV color space to the RGB color space is done by equation (4). R, G, B] and [H, S, V] ranges are the same as above.

The noise applying unit 14 applies applied noise having an intensity corresponding to the brightness of each image pixel to the V component of the image pixel. A specific relationship between the brightness of an image pixel and the applied noise intensity will be described later.

A threshold applying unit 16 compares the V component of each image pixel to which the applied noise is applied with a predetermined threshold and outputs the comparison result. The nonlinear element shown in FIG. 2 can be used for the threshold application unit 16 .

A restored image generator 18 generates a restored image from the comparison result.

FIG. 7 shows an example of the relationship between the brightness of an image pixel and the applied noise intensity. Here, Gaussian white noise with an average value of 0 is used as applied noise. The noise intensity value is represented by the variance of this white Gaussian noise. Three points, P1 (the point at which the brightness of the image pixel is minimum), P2 (the point at which the brightness of the image pixel is the threshold value), and Point P3 (the point at which the brightness of the image pixel is maximum) are defined below. The applied noise intensity shall be given by connecting the points with a straight line. The noise intensities at P1, P2 and P3 are denoted as VP1, VP2 and VP3, respectively. Further, the threshold value VP2 is fixed as VP2=50/255 (≈0.2) (the same applies to other examples below).

In FIG. 7, VP1=0.02 and VP3=0.3. In this case, the relationship between the brightness of the image pixel and the applied noise intensity (hereinafter also referred to as "applied noise characteristic") is indicated by a straight line with a positive slope having the point P2 as an inflection point. That is, the applied noise intensity has a positive correlation with the brightness of the image pixel both when the brightness of the image pixel is less than the threshold and when the brightness of the image pixel is greater than or equal to the threshold.

FIG. 8 shows an example of an original image and a restored image when given the applied noise characteristics of FIG. (a) is an original image, (b) is a restored image in which the stochastic resonance phenomenon is expressed in all RGB components, and (c) is a restored image in which the stochastic resonance phenomenon is expressed only in the V component. Bright areas and dark areas are remarkably mixed in the original image of (a). In particular, an image of a strawberry in a dark area is completely blackened and is almost invisible. This indicates that the image information in the dark portion is buried in noise. In the restored image of (b), the visibility is improved by raising the brightness of the dark region. However, the entire image is tinged with white, and the image quality is degraded especially in bright areas. The hue is also different from the original image. In the restored image of (c), the brightness of the dark region is increased, and visibility is improved, and the entire image is not tinged with white. The hue is almost the same as the original image. In this way, in the restored image of (c), the dynamic range of brightness is expanded while maintaining the image quality of the bright and dark areas in the image.

The noise applying section 14 and the threshold applying section 16 may be configured by a summing network. In this case, a high input-output correlation can be obtained without requiring noise intensity adjustment, so that the image quality of the restored image can be easily improved.

FIG. 9 shows examples of nine different applied noise characteristics obtained by varying the values of VP1 and VP3.
(a) is for VP1=0.02 and VP3=0.02. This characteristic has a positive correlation when the brightness of the image pixel is below the threshold and a negative correlation when it is above the threshold.
(b) is for VP1=0.10 and VP3=0.02. This characteristic is constant when the brightness of the image pixel is below the threshold and has a negative correlation when it is above the threshold.
(c) is for VP1=0.35 and VP3=0.02. This characteristic has a negative correlation both when the brightness of the image pixel is below the threshold and when it is above the threshold.
(d) is for VP1=0.02 and VP3=0.10. This property has a positive correlation when the brightness of the image pixels is below the threshold and is constant when it is above the threshold.
(e) is for VP1=0.10 and VP3=0.10. This characteristic is constant whether the brightness of the image pixel is below or above the threshold.
(f) is for VP1=0.35 and VP3=0.10. This characteristic has a negative correlation when the brightness of the image pixel is below the threshold and is constant when it is above the threshold.
(g) is for VP1=0.02 and VP3=0.35 (same as the applied noise characteristic in FIG. 7). This characteristic has a positive correlation both when the brightness of the image pixel is below the threshold and when it is above the threshold.
(h) is for VP1=0.10 and VP3=0.35. This characteristic is constant when the brightness of the image pixel is below the threshold and has a positive correlation when it is above the threshold.
(i) is for VP1=0.35 and VP3=0.35. This characteristic has a negative correlation when the brightness of the image pixel is below the threshold and a positive correlation when it is above the threshold.

FIG. 10 shows restored images corresponding to the applied noise characteristics of (a) to (i) of FIG. It can be seen that, depending on how the applied noise characteristics are given, various effects can be obtained regarding the expansion of the dynamic range of lightness. In FIG. 10, the tendency is that the darker regions are emphasized in the restored image on the right, and the overall brightness is increased in the restored image on the upper side.

[Second embodiment]
FIG. 11 shows a functional block diagram of the image processing device 2 according to the second embodiment. The image processing device 2 includes a storage unit 20 in addition to the configuration of the image processing device 1 shown in FIG. The storage unit 20 stores the relationship between the brightness of each image pixel and the applied noise intensity (applied noise characteristic). For example, the storage unit 20 may store various applied noise characteristics as shown in FIG. The storage unit 20 may store applied noise characteristics in any manner. For example, the user may manually input new applied noise characteristics to the storage unit 20 . Alternatively, data of applied noise characteristics may be automatically uploaded to the storage unit 20 from an external database.

According to this embodiment, noise with different applied noise characteristics can be applied to the original image.

[Third Embodiment]
FIG. 12 shows a functional block diagram of the image processing device 3 according to the third embodiment. The image processing device 3 has a learning unit 22 in addition to the configuration of the image processing device 1 in FIG. The learning unit 22 performs machine learning using the brightness of each image pixel, the applied noise intensity, and the restored image as learning data, and outputs the applied noise intensity when the original image is input.

It is believed that there is some relationship between the brightness of each image pixel, the applied noise intensity, and the reconstructed image. Therefore, it is considered that the optimal brightness vs. noise intensity profile can be determined by performing machine learning using the brightness of each image pixel, the applied noise intensity, and the restored image as teacher data. Machine learning may be performed by known AI. A specific AI technique is not particularly limited, but for example, deep learning may be applied. In this embodiment, a large number of pairs of brightness and applied noise intensity of each image pixel for generating a restored image are prepared, and these are used as learning data for AI to learn. Accordingly, AI can calculate and output the applied noise characteristics that generate the optimum restored image when the original image is input.

According to this embodiment, machine learning can be used to predict applied noise characteristics that can generate an optimal restored image when an original image is given.

[Fourth Embodiment]
FIG. 13 shows a functional block diagram of the image processing device 4 according to the fourth embodiment. The image processing device 4 includes a threshold determination unit 24 in addition to the configuration of the image processing device 1 in FIG. The threshold determination unit 24 determines the threshold based on the distribution of the V component in the HSV color space of the original image.

According to this embodiment, since the threshold value is determined based on the distribution of the V component in the HSV color space of the original image, the present invention can provide a suitable threshold value for the binarization processing that is normally performed. .

[Fifth Embodiment]
FIG. 14 shows a flowchart of an image processing method according to the fifth embodiment. This method is an image processing method that performs image processing using an image processing apparatus that includes a pixel extraction unit, a conversion unit, a noise application unit, a threshold application unit, and a restored image generation unit. In step S1, the method extracts image pixels from a color source image containing a plurality of image pixels. At step S2, the method converts the RGB color space of each image pixel to the HSV color space. In step S3, the method applies an applied noise whose intensity depends on the brightness of each image pixel to the V component of that image pixel. In step S4, the method compares the V component of each image pixel with applied noise to a predetermined threshold and outputs a comparison result. At step S5, the method generates a decompressed image from the comparison result.

According to this embodiment, the image processing apparatus can be used to expand the dynamic range of lightness while maintaining the image quality of bright and dark areas in an image.

[Sixth Embodiment]
A sixth embodiment is a program. This program is a program that causes a computer to execute image processing using an image processing device that includes a pixel extraction unit, a conversion unit, a noise application unit, a threshold application unit, and a restored image generation unit. This program causes the computer to execute the following steps S1 to S5. In step S1, image pixels are extracted from a color original image containing a plurality of image pixels. At step S2, the RGB color space of each image pixel is converted to the HSV color space. In step S3, applied noise having an intensity corresponding to the brightness of each image pixel is applied to the V component of the image pixel. In step S4, the V component of each image pixel to which the applied noise is applied is compared with a predetermined threshold, and the comparison result is output. In step S5, a restored image is generated from the comparison result.

According to this embodiment, a computer can execute a program for expanding the dynamic range of lightness while maintaining the image quality of bright and dark areas in an image.

[Seventh Embodiment]
A seventh embodiment is a storage medium. This storage medium stores a program that causes a computer to execute image processing using an image processing device that includes a pixel extraction unit, a conversion unit, a noise application unit, a threshold application unit, and a restored image generation unit. . This program causes the computer to execute the following steps S1 to S5. In step S1, image pixels are extracted from a color original image containing a plurality of image pixels. At step S2, the RGB color space of each image pixel is converted to the HSV color space. In step S3, applied noise having an intensity corresponding to the brightness of each image pixel is applied to the V component of the image pixel. In step S4, the V component of each image pixel to which the applied noise is applied is compared with a predetermined threshold, and the comparison result is output. In step S5, a restored image is generated from the comparison result.

According to this embodiment, it is possible to store a program for expanding the dynamic range of lightness while maintaining the image quality of bright and dark areas in the image.

The calculation process in the present invention has a high affinity with accelerators for deep learning. In fact, it should be noted that a GPU (Graphics Processing Unit) can be used to increase the speed, and that the process of bundling signals via nonlinear elements is a neural network itself.

Although the present invention has been described with reference to the above-described embodiments, the present invention is not limited to the above-described embodiments. It is also included in the present invention. In addition, it is also possible to appropriately rearrange the combination and the order of steps in each embodiment based on the knowledge of a person skilled in the art, and to add modifications such as various design changes to each embodiment. Embodiments in which is added may also be included in the scope of the present invention.

In the embodiment, three points P1, P2, and P3 are determined, and applied noise intensity is given by connecting these points with a straight line. The applied noise intensity is not limited to this, and may be given by drawing any suitable curve on the image pixel brightness (x-axis)-applied noise intensity (y-axis) plane. According to this modification, the degree of freedom in determining the applied noise characteristics is improved.

Nonlinear elements in systems that exhibit stochastic resonance are related to activation functions in deep learning. In the above embodiments, a simple threshold system (step function) has been described. However, it is not limited to this, and it has been confirmed that the stochastic resonance phenomenon occurs even when a sigmoid function or the like is used as a nonlinear element, and there is a possibility that various activation functions used in recent deep learning can be applied. .

Such a modified example has the same function and effect as the embodiment.

Any combination of the above-described embodiments and modifications is also useful as an embodiment of the present invention. A new embodiment resulting from the combination has the effects of each of the combined embodiments and modifications.

1.. An image processing device.
2.. An image processing device.
3.. An image processing device.
4: Image processing device.
10... pixel extractor.
12... Conversion part.
14. Noise applying unit.
16... A threshold application unit.
18: Restored image generator.
20... A storage unit.
22... Learning part.
24 .. a threshold determination unit.
100... Stochastic resonance phenomenon system.
102... Noise application unit.
104... A threshold application unit.
S1: Step of extracting image pixels.
S2: Step of converting RGB color space to HSV color space.
S3: Applying an applied noise whose intensity depends on the brightness of each image pixel to the V component of that image pixel.
S4: Step of comparing the V component of each image pixel with a predetermined threshold value and outputting the comparison result.
S5: A step of generating a restored image.

Claims (16)

a pixel extractor for extracting image pixels from a color source image containing a plurality of image pixels;
a conversion unit that converts the RGB color space of each image pixel to an HSV color space;
a noise applying unit that applies applied noise having an intensity corresponding to the brightness of each image pixel to the V component of the image pixel;
a threshold application unit that compares the V component of each image pixel to which the applied noise is applied with a predetermined threshold and outputs a comparison result;
a restored image generation unit that generates a restored image from the comparison result ,
The relationship between the brightness of each image pixel and the applied noise intensity is:
The brightness of each image pixel has a positive or negative correlation or is constant when the brightness is less than a threshold, and has a positive or negative correlation or is constant when the brightness of each image pixel is greater than or equal to the threshold. image processing device. 2. The image processing apparatus according to claim 1, wherein said noise applying section and said threshold applying section are configured by a summing network. The relationship between the brightness of each image pixel and the applied noise intensity is:
3. The image processing apparatus according to claim 1, wherein there is a positive correlation both when the brightness of each image pixel is less than the threshold and when the brightness of each image pixel is greater than or equal to the threshold. The relationship between the brightness of each image pixel and the applied noise intensity is:
3. The image processing apparatus according to claim 1, wherein the image pixels have a negative correlation both when the brightness of each image pixel is less than a threshold value and when the brightness of each image pixel is greater than or equal to the threshold value. The relationship between the brightness of each image pixel and the applied noise intensity is:
having a positive correlation when the brightness of each image pixel is less than a threshold;
3. The image processing apparatus according to claim 1, wherein when the lightness of each image pixel is equal to or higher than a threshold value, it has a negative correlation. The relationship between the brightness of each image pixel and the applied noise intensity is:
having a negative correlation when the brightness of each image pixel is less than a threshold;
3. The image processing apparatus according to claim 1, wherein the brightness of each image pixel has a positive correlation when it is equal to or higher than a threshold value. The relationship between the brightness of each image pixel and the applied noise intensity is:
is constant when the brightness of each image pixel is less than a threshold;
3. The image processing apparatus according to claim 1, wherein the brightness of each image pixel has a positive correlation when it is equal to or higher than a threshold value. The relationship between the brightness of each image pixel and the applied noise intensity is:
having a positive correlation when the brightness of each image pixel is less than a threshold;
3. The image processing apparatus according to claim 1, wherein the brightness of each image pixel is constant when it is equal to or higher than a threshold. The relationship between the brightness of each image pixel and the applied noise intensity is:
having a negative correlation when the brightness of each image pixel is less than a threshold;
3. The image processing apparatus according to claim 1, wherein the brightness of each image pixel is constant when it is equal to or higher than a threshold. The relationship between the brightness of each image pixel and the applied noise intensity is:
is constant when the brightness of each image pixel is less than a threshold;
3. The image processing apparatus according to claim 1, wherein when the lightness of each image pixel is equal to or higher than a threshold value, it has a negative correlation. 11. The image processing apparatus according to any one of claims 1 to 10, further comprising a storage unit that stores the relationship between the brightness of each image pixel and the applied noise intensity. A learning unit that outputs the applied noise intensity when the original image is input by performing machine learning using the brightness of each image pixel, the applied noise intensity, and the restored image as learning data. The image processing apparatus according to claim 1 or 2. 13. The image processing apparatus according to any one of claims 1 to 12, further comprising a threshold determination unit that determines a threshold based on a distribution of V components in the HSV color space of the original image. An image processing method for performing image processing using an image processing device comprising a pixel extraction unit, a conversion unit, a noise application unit, a threshold application unit, and a restored image generation unit,
extracting image pixels from a color source image containing a plurality of image pixels in the pixel extractor;
converting the RGB color space of each image pixel to an HSV color space in the conversion unit;
applying, in the noise applying unit, applied noise having an intensity corresponding to the brightness of each image pixel to the V component of the image pixel;
comparing the V component of each image pixel to which the applied noise is applied with a predetermined threshold in the threshold applying unit and outputting a comparison result;
generating a restored image from the comparison result in the restored image generation unit ;
The relationship between the brightness of each image pixel and the applied noise intensity is:
The brightness of each image pixel has a positive or negative correlation or is constant when the brightness is less than a threshold, and has a positive or negative correlation or is constant when the brightness of each image pixel is greater than or equal to the threshold. image processing method. A program for causing a computer to execute image processing using an image processing device comprising a pixel extraction unit, a conversion unit, a noise application unit, a threshold application unit, and a restored image generation unit,
extracting image pixels from a color source image containing a plurality of image pixels in the pixel extractor;
converting the RGB color space of each image pixel to an HSV color space in the conversion unit;
applying, in the noise applying unit, applied noise having an intensity corresponding to the brightness of each image pixel to the V component of the image pixel;
comparing the V component of each image pixel to which the applied noise is applied with a predetermined threshold in the threshold applying unit and outputting a comparison result;
causing a computer to execute a step of generating a restored image from the comparison result in the restored image generating unit ;
The relationship between the brightness of each image pixel and the applied noise intensity is:
The brightness of each image pixel has a positive or negative correlation or is constant when the brightness is less than a threshold, and has a positive or negative correlation or is constant when the brightness of each image pixel is greater than or equal to the threshold. A program to A storage medium storing a program for causing a computer to execute image processing using an image processing device comprising a pixel extraction unit, a conversion unit, a noise application unit, a threshold application unit, and a restored image generation unit, ,
extracting image pixels from a color source image containing a plurality of image pixels in the pixel extractor;
converting the RGB color space of each image pixel to an HSV color space in the conversion unit;
applying, in the noise applying unit, applied noise having an intensity corresponding to the brightness of each image pixel to the V component of the image pixel;
comparing the V component of each image pixel to which the applied noise is applied with a predetermined threshold in the threshold applying unit and outputting a comparison result;
causing a computer to execute a step of generating a restored image from the comparison result in the restored image generating unit ;
The relationship between the brightness of each image pixel and the applied noise intensity is:
storing a program that has a positive or negative correlation or is constant when the brightness of each image pixel is less than a threshold, and has a positive or negative correlation or is constant when the brightness of each image pixel is greater than or equal to the threshold; storage medium.

Images