JPH11101944A - Light source device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To make it possible to quickly response to a change in a spectral distribution and to improve efficiency timewise and costwise, by letting a neural network means output a control signal for a spatial light modulating means for providing spatially distributed light with a desired spectral distribution in a light source device which transforms white light into light of a predetermined spectral distribution and outputs it. SOLUTION: Plural desired spectral distribution data are stored in the memory in a microcomputer 16, and when a selected spectral distribution data is inputted to the neural network 17 constructed in the memory of the microcomputer 16, this is converted into a control signal for a liquid-crystal display panel control to achieve light with the desired spectral distribution. Thereafter, this control signal is forwarded to a panel driver 18, and the panel driver 18 drives a liquid-crystal display panel 5 to achieve light with the desired spectral distribution based on this control signal.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a light source device for illuminating a subject in an instrument for performing various inspections and analyzes, such as a microscope and an endoscope, and more particularly, to a light source device having a desired spectrum for the subject. The present invention relates to a light source device that irradiates light having a distribution and enables highly accurate spectral image analysis.

[0002]

2. Description of the Related Art In a device such as a microscope or an endoscope that illuminates a subject and performs inspection and analysis, the color information of the subject is faithfully reproduced in order to improve the inspection accuracy and analysis accuracy. Such reproduction is strongly desired, and it has been required to freely control the spectral distribution of the illumination light source.

In recent years, especially in microscopes and endoscopes, the use of photoelectric conversion means such as a CCD has often taken an image of a subject, and the wavelength characteristics of a lamp as a light source, Various filters (color filters, mosaic filters, filters for cutting off infrared light, etc.) placed in front of the solid-state image sensor, the wavelength characteristics of the light guide, and the sensitivity characteristics of the solid-state image sensor Since it is becoming difficult to faithfully reproduce the color information of the subject, the subject is not illuminated with light from a lamp as a light source, but is illuminated with light having a specific spectral distribution. Then, it is intended to ensure faithful reproduction of the color information of the subject.

Of course, optical filters that transmit only light in a specific wavelength range have been widely known, but such optical filters selectively transmit only light in a specific wavelength range. However, it is difficult to faithfully reproduce the color information of the subject by giving the illumination light an arbitrary spectral distribution.

In order to meet such a demand, the present inventors have already disclosed a light source device as shown in FIG. 10 (see Japanese Patent Application Laid-Open No. 4-297225). In other words, this device converts white illumination light emitted from the light source 101 into light having an arbitrary spectral distribution and can irradiate the light toward the subject. The light is collimated by a lens 102, diffracted by a diffraction grating 103 and spatially dispersed for each wavelength component, and the dispersed light is imaged on a liquid crystal display panel 105 by an imaging lens 104. By controlling the transmittance for each wavelength region, white light is converted into light having a target spectral distribution, and then the light of each wavelength component is converted by a light-mixing fiber bundle 108 through a condenser lens 107. Output evenly mixed.

[0006]

By the way, in order to control the transmittance for each wavelength region of light by the liquid crystal panel 105, a control signal obtained by calculation by a computer is input to an LCD driver. , A predetermined drive signal is applied to the liquid crystal display panel 105.

As a calculation method for deriving the control signal, Optics published on September 15, 1995
Iterative Feedback in Communications Magazine
Method (Iterative feedback method to make a spatial fi
lter on a liquid crystalspatial light modulator fo
r 2D spectroscopic pattern recognition; N.Hayasaka,
S. Toyooka, et al.) Is known. This technique is
When the liquid crystal display panel has a transmittance distribution of a predetermined function, feedback is repeated many times so that the spectral distribution of light output from the liquid crystal display panel approaches a desired spectral distribution. According to this method, it is possible to bring the spectrum distribution close to the target.

However, according to this method, each time light having a desired spectral distribution is produced, the above-mentioned feedback must be repeatedly performed in a similar manner, and the desired spectral distribution frequently changes. Was inefficient in terms of time and cost.

The present invention has been made in view of such circumstances, and when the spatial light modulating means produces light having a desired spectral distribution, the modulation characteristics of the spatial light modulating means are reduced in time and cost. It is an object of the present invention to provide a light source device that can be efficiently controlled in terms of efficiency.

[0010]

A light source device according to the present invention comprises: a spectral means for spatially dispersing light from a light emitting source in accordance with a wavelength; and a light having a predetermined spectral distribution as a whole. A spatial light modulator for controlling a modulation amount for each wavelength region, a light mixer for spatially mixing light for each wavelength component output from the spatial light modulator, and the spatial light A neural network for outputting a control signal for controlling a modulation amount of each of the wavelength components in the modulation means.

The neural network means receives a function signal representing a desired spectral intensity distribution as an output signal of the light source device, calculates the input signal based on a predetermined transfer function, and converts the control signal into a control signal. It is configured to output. Further, the transfer function can be a linear function.

The neural network means corresponds to a spectral intensity distribution output from the light source device when a learning teacher signal represented by a function f (λ) is input to the spatial light modulation means. A signal represented by a function g (λ) is input. At this time, the difference between the signal a (λ) output from the neural network means and the teacher signal represented by the function f (λ) is reduced. It is preferable that the weighting function ω (j, i) and the bias value b (j) are corrected, and this correction operation is set by repeatedly performing a plurality of the teacher signals.

Further, it is possible that the teacher signal is a sine function. Further, the spatial light modulating means can be constituted by an area code type filter for controlling a transmission cross-sectional area of light spatially dispersed by the spectral means. Further, the spatial light modulating means can be constituted by a gradation code type filter for controlling the transmittance of light spatially dispersed by the spectral means.

Preferably, the spatial light modulator is a liquid crystal display panel. In addition, it is desirable that the dispersing unit includes a lens for converting light from the light source lamp into parallel light, and a diffraction grating or a prism for decomposing the parallel light into light of each wavelength component. Further, it is preferable that the light mixing means is constituted by a diffraction grating, an integrating sphere, or a fiber bundle in which the output end is random with respect to the input end.

[0015]

Embodiments of the present invention will be described below with reference to the drawings. Here, FIG. 1 is a schematic diagram showing the configuration of the entire light source device. In FIG. 1, white light from a white light source lamp 1 is converted into parallel light by a collimator lens 2 and is incident on a diffraction grating 3, where it is diffracted in different directions for each wavelength. The diffracted light passes through the cylindrical lens 4 to form a spectral spectrum image spatially dispersed on the liquid crystal display panel 5 according to the wavelength band. The liquid crystal display panel 5 has two polarizing plates 6
a, 6b, the amount of transmitted light changes in each wavelength region according to the applied voltage. Polarizing plate 6a,
The light transmitted through the liquid crystal panel 5 and the polarizing plate 6b is focused on the same area of the diffraction grating 8 by the cylindrical lens 7. The light condensed in this manner is mixed by the diffraction grating 8, condensed on the light incident part 10 of the light guide 11 by the condensing lens 9, transmitted by the light guide 11, and transmitted by the light guide 11. The light is emitted from the emission unit 12 and becomes illumination light for the subject 13. The light that enters from the back surface of the subject 13 and passes through the subject 13 and carries the image information of the subject 13 is magnified by the microscope 14 and formed on the image plane of the CCD camera 15.

The subject image data picked up by the CCD camera 15 is sent to a microcomputer 16 and used for spectral image analysis.

On the other hand, a plurality of desired spectral distribution data (target data T) is stored in a memory in the microcomputer 16 and is selected by the neural network 17 constructed in the memory of the microcomputer 16. When the spectral distribution data is input, it is converted into a control signal for controlling a liquid crystal display panel for realizing light having a desired spectral distribution. Thereafter, the control signal is sent to the panel driver 18, and the panel driver 18 drives the liquid crystal display panel 5 based on the control signal.

A condensing lens 19 and a slit 20 are provided between the white light source lamp 1 and the collimator lens 2 for forming a linear light source by the light from the white light source lamp 1.

Here, as the light source lamp 1, a lamp that emits white light, such as a xenon lamp or a halogen lamp, is used. The white light from the light source lamp 1 is converted into parallel light by a collimator lens 2, diffracted by a diffraction grating 3, and transmitted to a light transmitting area of a liquid crystal display panel 5 by a cylindrical lens 4, and a visible light area of about 400
A dispersion image of about 700 nm to about 700 nm is projected. The liquid crystal display panel 5 can control the transmittance of light in, for example, a vertical stripe shape, and determines the transmittance of the light projected on the liquid crystal display panel 5 for each predetermined wavelength band. That is, as shown in FIG.
8, a panel driver 18 is connected to a neural network 17, and the light transmittance is controlled for each wavelength band based on a control signal from the neural network 17. Thus, the liquid crystal display panel 5 can convert white light into light having a desired spectral distribution.

As described above, the neural network 17
Is used to control the liquid crystal display panel 5,
According to the neural network 17, the transfer function of the entire light source device including the liquid crystal display panel 5 can be uniquely specified,
A filter function for generating light having a desired spectral distribution can be immediately generated, and the IterativeFee described above is used.
This is because a complicated operation of turning a feedback loop every time the filter function is generated as in the related art such as the dback method may be unnecessary.

By the way, in order to determine the transfer function of the whole system of the apparatus and to construct a neural network corresponding to the transfer function, it is necessary to previously train the neural network 17 using a predetermined learning algorithm. Hereinafter, a specific learning method of the neural network 17 will be described.

FIG. 2 shows the configuration of the device when learning is performed on the neural network 17. Compared with the configuration shown in FIG. 1 when the device is actually operated, the subject 13, the microscope 14, and the CCD A monochromator 31 and an AD converter 32 are inserted instead of the camera 15. Hereinafter, a learning method for training the neural network 17 using the device shown in FIG. 2 will be described. The learning method used in this embodiment is Widrow-Ho
f This is called a learning rule.

First, the network architecture of the neural network 17 will be described with reference to FIG. This network is a linear network that outputs an output vector group a (j) corresponding to a given input vector group p (i) (since the transfer function of the entire apparatus is assumed to be a linear function). , One layer having S neurons coupled to R input vectors p (i) via a matrix of weighting functions ω (j, i).
In this embodiment, a linear transfer function is used as each neuron. That is, in each element n (j) of the neuron, each input vector p (i) is assigned a weighting function ω
A bias b (j) is further added to the sum obtained by multiplying by (j, i), and this is multiplied by a transfer function specific to a neuron to output an output vector group a (j).

At the time of learning, this output vector group a
(J) is compared with a teacher function L serving as a control signal of the liquid crystal display panel 5, and a weighting function ω is set so that the value of the output vector group a (j) approaches the value of the teacher function L.
The parameter values of (j, i) and the bias b (j) are modified. The above parameters are converged by repeatedly performing the same training using a large number of teacher functions L for learning, and a neural network 17 capable of immediately generating a desired function in actual use can be constructed. .

More specifically, in this embodiment, training of the neural network 17 is performed by selecting 30 types of learning function L for learning. Fig. 4 shows three examples.
(A), 5 (A) and 6 (A). That is, FIG.
5A shows the first teacher function L1, which is 0 at all positions on the horizontal axis, and FIG. 5A shows the second teacher function L2, which is a half of the sine wave. FIG. 6A shows the characteristics of the wavelength shape, and FIG.
3 shows the characteristic of a continuous sine wave shape. Here, FIGS. 4 (A), 5 (A), 6
In the graph shown in (A), the horizontal axis is the position in the light dispersion direction of the liquid crystal display panel 5 (since the liquid crystal display panel 5 of the present embodiment has a configuration in which 512 stripes are arranged in parallel in the light dispersion direction, And the vertical axis is the signal input value. The reason why the sine function is selected as the teacher signal L in the present embodiment is that
This is for the purpose of facilitating subsequent arithmetic processing such as Fourier transformation. Of course, various teacher functions such as other commonly used binary functions and Chebyshev functions may be used instead of or together with this. It is possible.

Such a teacher function L is stored in a memory (not shown) provided in the microcomputer 16, and is output one by one at the time of learning of the neural network 17, and is output via the panel driver 18. It is input to the liquid crystal display panel 5. As a result, in the liquid crystal display panel 5, the transmittance is set for each stripe so that a transmittance distribution according to the input teacher function L is formed.

In this manner, when the transmittance of each stripe is set according to the position of the liquid crystal display panel 5 in the light dispersion direction, the light transmitted from the white light source lamp 1 through the liquid crystal display panel 5 becomes Rather than having the same spectral distribution as the teacher function L, it becomes significantly distorted due to the characteristics of the system. As described above, the deformation is caused by the characteristics of the entire system of the apparatus, such as the lamp characteristics of the light source and the characteristics of various filters.

Each of the obtained teacher functions L
FIGS. 4B, 5B, and 6 show the spectral distributions measured by the monochromator 31 corresponding to 1, L2, and L3, respectively.
It is shown in (B). In each of these figures, the horizontal axis indicates wavelength (nm), and the vertical axis indicates relative intensity.

That is, in the apparatus shown in FIG. 2, light having a spectral distribution modulated by the liquid crystal display panel 5 is mixed with light of each wavelength region by the diffraction grating 8 and input to the monochromator 31 via the light guide 11. In the monochromator 31, the signal is again separated into level signals for each wavelength region, and is converted into a digital value by the AD converter 32, as shown in FIGS. 4 (B), 5 (B) and 6 (B). It is output as a signal representing the input function g (λ) to the neural network 17 having such characteristics.

The function g (λ) constitutes the input vector group p (i) to the neural network 17 described above, and the output vector group a (j) at this time.
And the teacher function L output from the microcomputer 16 for obtaining the function g (λ), and the weighting function ω (j,
i) and the bias b (j) will be modified. Thereafter, as described above, the training is repeated using a number of teacher functions L, and the weighting function ω (j, i) and the bias b
(J) is made to converge to an optimal value.

Next, when general test data (target data T) other than learning data is input to the neural network 17 using the neural network trained in this manner, the apparatus shown in FIG. The extent to which the obtained output signal shows a value close to the target data T was verified. 7 (A), 8 (A), 9
The dotted lines in (A) are graphs showing target data T (filter function optimized to display the Munsell color target 1257 colors) input to the neural network 17, and are shown in FIGS. 7 (B) and 8 (B). , 9 (B) are graphs showing output signals (control signals of the liquid crystal display panel for realizing a spectrum distribution corresponding to the target data T) output from the neural network 17, respectively, as shown in FIGS. Solid lines 8 (A) and 9 (A) are graphs each showing the finally obtained spectrum distribution.
In (A) of each figure, the horizontal axis represents the wavelength (nm), the vertical axis represents the normalized intensity (each curve and the horizontal axis (in units of nm)), and the light intensity is normalized by the area enclosed by the vertical axis. In each figure (B), the horizontal axis indicates the position of the liquid crystal display panel 5 in the light dispersion direction, and the vertical axis indicates the input value.

The norm error is 3.36% for the target data T shown in FIG. 7A, 7.77% for the target data T shown in FIG. 8A, and the target data shown in FIG. 9A. At T is 4.36%,
A spectrum distribution that coincided to a practically acceptable level was obtained. In the light source device shown in FIG. 1, a target data T corresponding to a desired spectral distribution is converted into a neural network 17 trained as described above.
A light having a desired spectral distribution can be output simply by inputting the light to the object, and this can be used as illumination light for the subject. Therefore, spectral image analysis with a microscope can be performed easily and with high accuracy. It can respond quickly to changes in distribution.

In the above embodiment, the case where the light source device of the present invention is applied to an illumination light source of a microscope has been described. However, by applying the light source device to an illumination light source of an endoscope or the like, color reproducibility of a subject image can be improved. It is also possible to
By applying the present invention to various inspection light sources and arbitrarily controlling the wavelength modulation characteristics of the spatial light modulator, it is possible to obtain a simple and highly accurate inspection result.

Further, the neural network means is not limited to the above-described embodiment, but a nonlinear network can be used. Further, it is possible to use a multilayered neuron layer to ensure consistency with a target spectral distribution. It is possible to further increase. In addition, the neural network learning algorithm described above
It is not limited to the widrow-Hoff learning rule, and various other learning algorithms can be used. In the above-described embodiment, the diffraction grating is used as the spectral unit. However, even if a prism is used, the light from the light source lamp can be spatially dispersed for each wavelength component.

As the spatial light modulating means, a liquid crystal display panel is used as a so-called gradation code type spatial filter for controlling the transmittance of light at each wavelength. It is also possible to use a so-called area code type spatial filter that provides a window through which light passes, and controls the cross-sectional area of this window through which light of each wavelength passes. The spatial light modulating means is not limited to a spatial filter for controlling the light transmittance.
MD (Digital Micro Mirror Device)
It is also possible to use light modulation means capable of controlling the reflectance of light in a predetermined direction.

Further, the light mixing means is not limited to the above-described diffraction grating, but may be a fiber bundle, an integrating sphere, frosted glass, or the like. Further, the wavelength region to be spectrally dispersed is not necessarily limited to visible light. For example, when illumination is performed with light in a wavelength region such as an infrared region or an ultraviolet region, a desired spectral distribution characteristic can be provided. is there.

[0037]

As described above, according to the light source device of the present invention, the white light from the white light source is spatially dispersed for each wavelength region, and the dispersed light is dispersed by the spatial light modulating means. A desired modulation is applied to each wavelength region to give a desired spectral intensity distribution, and when generating light having a required spectral intensity distribution, the light modulation of the spatial light modulating means is controlled using a neural network means. If the neural network means is trained in advance, it is not necessary to perform complicated control signal generation processing every time the required spectral intensity distribution changes. Even when the required spectral intensity distribution of light is changed, quick response is possible, and the efficiency can be significantly improved in terms of time and cost.

[Brief description of the drawings]

FIG. 1 is a schematic view showing a light source device according to an embodiment of the present invention.

FIG. 2 is a diagram for explaining a case where learning is performed on the neural network shown in FIG. 1;

FIG. 3 is a schematic diagram showing the architecture of the neural network shown in FIG. 1;

FIG. 4 is a graph (A) showing a teacher function L1 input to the spatial light modulator during learning of a neural network, and a graph showing a spectrum distribution of light output from the system corresponding to the teacher function L1 ( B)

FIG. 5 is a graph (A) showing a teacher function L2 input to the spatial light modulator during learning of the neural network, and a graph showing a spectrum distribution of light output from the system corresponding to the teacher function L2 ( B)

FIG. 6 is a graph (A) showing a teacher function L3 input to the spatial light modulator during learning of the neural network, and a graph showing a spectrum distribution of light output from the system corresponding to the teacher function L3 ( B)

FIG. 7 shows target data T1 input to the neural network, a corresponding graph (A) showing the spectral distribution of light output from the system, and output from the neural network inputting the target data T1. (B) showing the signal applied to the liquid crystal display panel

FIG. 8 shows target data T2 input to the neural network, a corresponding graph (A) showing the spectral distribution of light output from the system, and output from the neural network inputting the target data T2. (B) showing the signal applied to the liquid crystal display panel

FIG. 9 shows target data T3 input to the neural network, a corresponding graph (A) showing the spectral distribution of light output from the system, and output from the neural network inputting the target data T3. (B) showing the signal applied to the liquid crystal display panel

FIG. 10 is a schematic view showing a light source device according to the related art.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1,101 Light source lamp 2,102 Collimator lens 3,8,103 Diffraction grating 4,7,104,107 Cylindrical lens 5,105 Liquid crystal display panel 6a, 6b, 106a, 106b Polarizer 9,107 Condenser lens 11 Light guide 13 Subject 14 Microscope 15 CCD Camera 16 Microcomputer 17 Neural Network 18 Panel Driver 31 Monochromator 32 A / D Converter 108 Fiber Bundle

Claims (10)

[Claims] 1. A spectral means for spatially dispersing light from a light emitting source in accordance with a wavelength, and a modulation for each wavelength region so that the light dispersed by the spectral means has a predetermined spectral distribution as a whole. A spatial light modulator for controlling the amount; a light mixer for spatially mixing the light of each wavelength component output from the spatial light modulator; a modulation amount for each of the wavelength components in the spatial light modulator. And a neural network means for outputting a control signal for controlling the light source device. 2. The neural network means receives a function signal representing a desired spectral intensity distribution as an output signal of the light source device, calculates the input signal based on a predetermined transfer function, and calculates the control signal. The light source device according to claim 1, wherein the light source device is configured to output the light. 3. The light source device according to claim 2, wherein the transfer function is a linear function. 4. The neural network means corresponds to a spectral intensity distribution output from the light source device when a learning teacher signal represented by a function f (λ) is input to the spatial light modulation means. A signal represented by a function g (λ) is input. At this time, the difference between the signal a (λ) output from the neural network means and the teacher signal represented by the function f (λ) is reduced. Weighting function ω
4. The method according to claim 1, wherein (j, i) and the bias value b (j) are corrected, and the correction operation is set by repeatedly performing the plurality of teacher signals. The light source device according to claim 1. 5. The light source device according to claim 4, wherein the teacher signal is a sine function. 6. The spatial light modulator according to claim 1, wherein said spatial light modulator comprises an area code type filter for controlling a transmission cross-sectional area of the light spatially dispersed by said spectral unit. Light source device. 7. The spatial light modulator according to claim 1, wherein said spatial light modulator comprises a tone code type filter for controlling the transmittance of the light spatially dispersed by said spectral unit. Light source device. 8. The light source device according to claim 1, wherein the spatial light modulator is a liquid crystal display panel. 9. The apparatus according to claim 1, wherein the light splitting means comprises a lens for converting light from the light source lamp into parallel light, and a diffraction grating or a prism for decomposing the parallel light into light of each wavelength component. The light source device according to claim 1, wherein 10. The light source device according to claim 1, wherein said light mixing means is constituted by a diffraction grating, an integrating sphere, or a fiber bundle whose output end is random with respect to the input end.