JP7008448B2 - Hologram recording / playback device - Google Patents

Description

The present invention relates to a hologram recording / playback device, and more particularly to a hologram recording / playback device that demodulates bit string data from data modulated into a symbol of a two-dimensional image.

In recent years, a hologram optical memory medium (hologram recording medium) has been attracting attention as a medium capable of efficiently recording a large amount of data. Holographic memory is expected to be used as a large-capacity memory for images, sounds, computers, etc.

FIG. 6 shows an example of the optical system of the hologram recording / reproducing device 100. In hologram recording, a data pattern (page data) in which modulated (encoded) bit strings are arranged two-dimensionally is used.

The laser light output from the laser light source 101 and passing through the lens 102 and the spatial filter 103 becomes parallel rays in the collimating lens 104, and after the polarization plane is rotated to 45 degree polarization by the 1/2 wave plate 105, the mirror 106 It is reflected and separated into signal light and reference light by PBS (polarized beam splitter) 107. At the time of recording, the signal light passes through the PBS 108 and is irradiated on the SLM (spatial light modulation element) 109 composed of a reflective liquid crystal element or the like. The irradiated light is reflected by carrying the recording page data of the two-dimensional image by the black and white bit pattern projected on the element surface of the SLM 109, and returns to the PBS 108 as the object light. The generation of the recording page data is controlled by the recording processing unit (described later) which is the arithmetic unit 150. The object light returned from the SLM 109 is reflected by the PBS 108 and irradiated onto the hologram recording medium 116 via the FT (Fourier transform) lens 110.

On the other hand, the reference light reflected by the PBS 107 passes through the 1/2 wave plate 111, is reflected by the mirror 112, is reflected by the PBS 113, is further reflected by the galvano mirror 114, passes through the relay lens 115, and then passes through the relay lens 115, and then the hologram recording medium 116. Illuminated above. The reference light and the object light radiated on the hologram recording medium 116 in this way interfere with each other on the hologram recording medium 116 to form interference fringes, and the interference fringes are written on the hologram recording medium 116. ..

Next, the time of reproduction will be described. The laser light source 101 to PBS 107 are the same as at the time of recording, but the reflected reference light passes through the 1/2 wavelength plate 111, is reflected by the mirror 112, passes through the PBS 113, and passes through the mirror 117 and the galvano mirror. It is reflected by 118, passes through the relay lens 119, and then enters the hologram recording medium 116. The signal light diffracted by the interference fringes written on the hologram recording medium 116 has the information of the recording page data. After passing through the FT lens 110, the signal light passes through the PBS 108 and is imaged by the two-dimensional image sensor (image sensor) 120, and is processed by the reproduction processing unit (described later) which is the arithmetic unit 160 to restore the digital data. Will be.

In general, a hologram recording device can record multiple holograms at the same location by using various multiplexing methods such as angle multiplexing, phase code multiplexing, and spherical reference optical shift multiplexing, thereby recording information at high density. be able to.

As a modulation code, a method of determining a bright spot and a dark spot within a certain range in which page data is divided has been proposed (Patent Document 1). This is because the brightness conditions of the page data reproduced in the peripheral part and the central part of the page data differ depending on various conditions at the time of reproduction and the uneven brightness in the in-plane direction of the laser beam. This is because it is difficult to determine the bright spot and the dark spot. Examples of codes include a difference code that subtracts the diffracted light intensities of adjacent pixels and associates them with "0" if the result is negative and "1" if the result is positive, and 5: 9 modulation.

FIG. 7 describes data conversion (modulation) by 5: 9 modulation. Here, n: r modulation refers to a modulation method that converts a 1,0 bit string of n bits into a symbol (modulation symbol) composed of r pixels. 5: 9 modulation is a modulation method that expresses ⁵ -bit (25 = 32 ways) data with 3 × 3 9 pixels. As shown in FIG. 7A, the 3 × 3 9-pixel symbol is composed of 2 bright spots and 7 dark spots, and has 32 different symbols (# 1 to # 32) in which the bright spots are arranged differently. ) Is created. Each of the 32 symbols is associated with 5-bit data and used as a modulation code. Next, as shown in FIG. 7 (b), the data to be recorded is divided into 5 bits each and modulated by replacing each with the corresponding symbol. For example, if the first 1 to 5 bits are "10010", they are replaced with the corresponding symbol # 18, and if the next 6 to 10 bits are "10111", they are replaced with the corresponding symbol # 23. After that, page data is formed by laying out 3 × 3 symbols and arranging them. For example, in the case of page data of 1,740 pixels × 1,044 pixels, 1740 × 1044 ÷ (3 × 3) = 201,840 symbols are arranged, and 5 bits × 201840 = 1,009,200 bits. The data is modulated.

Next, data demodulation in 5: 9 modulation will be described with reference to FIG. The page data can be demodulated by determining the bright spot and the dark spot within a certain range (3 × 3 pixels) corresponding to the modulation symbol. As a data demodulation method, a process by hard determination, in which the brightness of each pixel in the symbol extracted from the page data is compared, is generally used. For example, in the case of 5: 9 modulation, since two bright spots are included, the brightness of the nine pixels is measured, and the two highest brightness points (in FIG. 8, the pixel having an 8-bit luminance value of 232). A pixel having a luminance value of 212) is regarded as a bright spot, and a symbol # 31 having a bright spot is associated with this position (upper right and lower left), and a data string is demoted from this symbol.

Next, the entire configuration of the hologram recording / reproducing device will be shown. FIG. 9 is an example of a block diagram of a conventional hologram recording / reproducing device using 5: 9 modulation. The hologram recording / reproducing device includes a recording processing unit 10, an optical system 20, and a reproduction processing unit 30.

The recording processing unit 10 performs processing for recording the recording bit string as a hologram. The recording processing unit 10 arranges the 5: 9 modulation unit 11 that performs the above-mentioned 5: 9 modulation and the symbol string generated by the 5: 9 modulation unit with respect to the input recording bit string, and creates page data. It is provided with a page data conversion unit 12 to be created.

The optical system 20 includes a recording optical system 21, a recording medium 22, and a reproduction optical system 23, and is composed of, for example, the optical elements shown in FIG. The recording optical system 21 irradiates the recording medium 22 with the object light and the reference light carrying the page data generated by the recording processing unit 10.

Page data is recorded on the recording medium (hologram recording medium 116 of FIG. 6) 22 due to the interference between the object light and the reference light.

The reproduction optical system 23 irradiates the recording medium 22 with reference light, reproduces signal light having page data information, acquires image data by an image pickup device or the like, and reproduces page data.

The reproduction processing unit 30 includes a symbol extraction unit 31, a luminance comparison unit 32, a bright spot position estimation unit 33, and a 5: 9 demodulation unit 34. The symbol extraction unit 31 cuts out the reproduced page data for each predetermined area and extracts a symbol (3 × 3 pixels). As described with reference to FIG. 8, the luminance comparison unit 32 measures the luminance of each pixel constituting the symbol and compares the luminance. When comparing the brightness, if necessary, processing for reducing noise caused by the optical system may be performed. As a result of the brightness comparison, the bright spot position estimation unit 33 estimates that the two pixels having the highest brightness are bright spots, identifies the pixel pattern, and determines the corresponding symbol number. The 5: 9 demodulation unit 34 converts the determined symbol number into a bit string according to the correspondence between the symbol and the bit string, demodulates the 5-bit data, and generates a reproduction bit string.

In this way, the hologram recording / reproducing device records / reproduces a bit string using 5: 9 modulation.

Patent No. 3209493 Patent No. 4863913

G. Burr et al., "Noise reduction of page-oriented data storage by inverse filtering during recording," Optics letters, 1998, Vol.23, No.4, pp.289-291 H. Osawa et al., "Neural Network Equalizer Matched to Recording Code in Holographic Data Storage," Jpn. J. Appl. Phys., 50, 09MB05, 2011

As described above, a hologram recording / reproducing device that determines a bit string by comparing the luminance of a certain pixel area (for example, 3 × 3 pixels in the case of 5: 9 modulation) has relatively few reading errors. Even in this case, there are many causes for lowering the signal-to-noise ratio. FIG. 10 shows some of the causes of lowering the SN ratio depending on the position in the page data. In FIG. 10, the left side is recorded data (page data), and the laser beam 140 having the recorded data information is transmitted via an optical element (lens, PBS, etc.) 130, and the reproduced data on the right side is read out. Shows.

FIG. 10A is an example in which the unevenness of the brightness of the laser beam reduces the SN ratio. Even within the range of 3 × 3 pixels, the in-plane luminance distribution 141 occurs in the laser beam when viewed locally, and the uneven luminance of the reproduced data misleads the identification of the bright spot pixel and lowers the SN ratio. Will let you. Although the effect can be suppressed in binary recording in which only the bright spot and the dark spot are determined by the difference code, for example, in multi-value recording, deterioration of the data error rate is unavoidable due to this decrease in the SN ratio.

FIG. 10B is an example in which dirt such as dust 131 adhering to the optical system 130 and a moire pattern generated by reflection in an optical path cause bit chipping. As shown in the figure, if a bit is missing in one pixel of a symbol using 5: 9 modulation, one bright spot can be detected during demodulation, but the brightness of the remaining eight spots is almost the same, and the camera is dark. The second bright spot is estimated by a slight difference in luminance caused by current noise or the like, and a data error occurs.

Further, as shown in FIG. 10 (c), the aberration 132 due to the optical component 130 such as a lens blurs the page data image and causes intersymbol interference. All of these noises are generated depending on the position of the page data.

Various methods for improving the signal-to-noise ratio have been proposed for these problems. In Non-Patent Document 1, the process of creating an inverse filter based on the optical characteristics acquired from the reproduced page data and re-recording the page data adjusted by the filter is repeated to reduce the unevenness of brightness in the page data surface. It is said that the suppression SN ratio can be improved. Further, in Patent Document 2, the moire pattern can be removed by recording the page data of all white pixels together and dividing the reproduced page data by the page data. Non-Patent Document 2 states that signal processing can be performed by an equalizer based on a neural network model configured for each pixel to suppress intersymbol interference.

However, even if this method is introduced in a place where a pixel defect is generated due to dust 131 or the like in the optical system using Non-Patent Document 1, the reproduced page data still has a defect, and the defect remains at the time of demodulation. Bit error rate does not improve. Further, in Patent Document 2, intersymbol interference cannot be suppressed. Non-Patent Document 2 also requires a camera whose pixel size is half that of SLM, and at the same time, since a neural network is constructed for each pixel of page data, the amount of calculation and time for creating an equalizer are required. Will be enormous. Then, after performing the processing for improving the SN ratio by these methods, it is necessary to demodulate the page data (symbol), which is a two-step processing.

The present invention has been made to solve the above-mentioned conventional problems, and the purpose of the present invention is simply to suppress noise and demodulate symbols based on optical characteristics of data reproduced from a hologram recording / reproducing device. It is an object of the present invention to provide a hologram recording / reproducing device capable of improving the SN ratio while performing one process.

In order to solve the above problems, the hologram recording / reproducing device according to the present invention is obtained from a hologram recording medium recorded by a modulation method that converts n (n> 1) bit bit string data into a modulation symbol consisting of r pixels. It is a hologram recording / playback device that restores data, and the page data reproduced from the hologram recording medium is subjected to the modulation symbol by image recognition processing using an algorithm configured by machine learning using the past reproduction page data in advance. It is characterized in that it is demoted to the bit string data from the above.

Further, the hologram recording / reproducing device can divide the reproduction page data into modulation symbols, machine-learn the optical characteristics different for each position in the page data for each modulation symbol, and construct a demodulation algorithm different for each position. desirable.

Further, the hologram recording / reproducing device is a neural network model convoluted by the machine learning, and it is desirable that the neural network model does not perform pooling to reduce the dimension.

According to the present invention, it is possible to remove noise and the like contained in the optical system and demodulate the symbols in the page data by a single process. Therefore, since signal processing and data demodulation for reproduction with a low error rate are executed at the same time, high-speed data demodulation can be realized. Further, by creating a plurality of models in which page data is divided into fixed ranges, it is possible to suppress deterioration of data errors during demodulation due to noise generated depending on the in-plane position.

It is a block diagram which shows an example of the hologram recording / reproduction apparatus of this invention. It is a figure which shows the system example of the image recognition by a convolutional neural network. It is a figure which shows the outline of the demodulation system of the reproduction page data. It is a figure explaining the correlation between the bit error by the conventional demodulation method and the learning frequency of a convolutional neural network. It is a figure which shows the comparison of bit error by a demodulation method. It is a figure which shows an example of the optical system of the hologram recording / reproduction apparatus. It is a figure explaining the data modulation method by 5: 9 modulation. It is a figure explaining the data demodulation method by 5: 9 modulation. It is a block diagram which shows an example of the conventional hologram recording / reproduction apparatus. It is a figure explaining the cause which reduces the SN ratio depending on the position in a page data.

Hereinafter, embodiments of the present invention will be described with reference to the drawings.

FIG. 1 shows an example of the hologram recording / reproducing device of the present invention in a block diagram. Here, 5: 9 modulation was used as a modulation / demodulation method for constructing a symbol (modulation symbol). The modulation / demodulation method is not limited to 5: 9 modulation, and any modulation / demodulation method that converts a bit string into a symbol (for example, 2: 4 modulation, 12:16 modulation in which a symbol is composed of 4 × 4 pixels, etc.) is used. be able to.

The same configuration as the conventional hologram recording / reproducing device shown in FIG. 9 has the same reference numerals. The hologram recording / reproducing device of the present invention includes a recording processing unit 10, an optical system 20, and a reproduction processing unit 40.

The recording processing unit 10 is basically the same as a conventional device, and includes a 5: 9 modulation unit 11 and a page data conversion unit 12, and performs processing for recording a recording bit string as a hologram.

The 5: 9 modulation unit 11 divides the input recording bit string (data string to be recorded) into 5 bits, and selects a symbol (3 × 3 pixels) corresponding to the 5 bits by using, for example, a conversion table or the like. .. This symbol is sequentially output to the page data conversion unit 12 as a modulation code.

The page data conversion unit 12 arranges the symbol strings generated by the 5: 9 modulation unit 11 to generate page data. The page data is, for example, 1740 pixels × 1044 pixels, and 1740 × 1044 ÷ (3 × 3) = 201840 symbols are arranged to create one page data.

Although the processing required for hologram recording of the recording bit string will be described here, an error correction function may be added to the hologram recording / reproducing device as needed. That is, the recording processing unit 10 may also include an error correction coding unit (not shown) that performs error correction coding for the recording bit string.

The optical system 20 is basically the same as the conventional device, and is composed of, for example, the optical elements shown in FIG. 6, and includes a recording optical system 21, a recording medium 22, and a reproduction optical system 23.

The recording optical system 21 displays the page data generated by the recording processing unit 10 on the SLM (spatial light modulation element) 109, and as described above, generates the object light and the reference light on which the page data is carried to generate the recording medium. 22 (hologram recording medium 116) is irradiated.

In the recording medium (hologram recording medium) 22, page data is recorded as interference fringes due to the interference between the object light and the reference light. Generally, a plurality of holograms are recorded on the recording medium 22, and high-density information recording is possible.

The reproduction optical system 23 irradiates the recording medium 22 with the reference light, and reproduces the signal light having the information of the page data as described above. This signal light is imaged by an image sensor 120 or the like, becomes image data, and reproduces page data.

The reproduction processing unit 40 includes a symbol extraction unit 41 and an image recognition processing unit 42, and has a different configuration from the conventional one.

The symbol extraction unit 41 is the same as the conventional one, and the reproduced page data is cut out for each predetermined area, and the symbol (3 × 3 pixels) is extracted.

The image recognition processing unit 42 directly democratizes 5-bit data from the symbol extracted by the symbol extraction unit 41 by image recognition (or pattern recognition) using machine learning (for example, deep learning), and reproduces a bit string. To generate. That is, the image recognition processing unit 42 removes noise caused by the optical system, which has been conventionally performed, and determines an appropriate symbol, and demodulates a reproduction bit string (bit string data) from the determined symbol. It is done all at once.

Although the demodulation process of the reproduction bit string has been described here, the reproduction processing unit 40 includes an error correction decoding unit (not shown) or the like, which corrects and decodes the error based on the demodulated bit string, if necessary. You can also.

As described above, the hologram recording / reproducing device of the present invention records / reproduces a bit string using 5: 9 modulation.

As a machine learning algorithm used for image recognition in the image recognition processing unit 42, a convolutional neural network was used in this embodiment. FIG. 2 shows a system example of image recognition by a convolutional neural network. Each symbol image (here, 3 x 3 pixels) in the reproduced page data is used as an input signal, and from a plurality of neurons (here, 9 neurons whose input is the brightness value of each pixel) that receives the signal. One or more intermediate layers consisting of a group of neurons that receive the excitement pattern of the neurons in the previous layer and perform pattern conversion, and then output the excitement pattern to the next stage, and the final layer. It consists of an output layer that receives, converts, and outputs the excitement pattern of neurons in the middle layer.

The intermediate layer contains one or more convolution layers for detecting features between adjacent pixels, and here, a 3 × 3 convolution filter processing layer and a 2 × 2 convolution filter processing layer are shown. There is. Further, if necessary, zero padding may be performed around 3 × 3 pixels to perform a 5 × 5 pixel convolution process. This neural network usually does not perform dimension-reducing pooling to extract features.

If the number of neurons in the output layer is 32 (= 5 bits), 32 types of likelihood for the input symbol (described in the image of the bar graph in Fig. 2) will be output, which is the highest likelihood. The data with the highest value is immediately demodulated data (determined symbol = demodulated 5-bit data string).

FIG. 3 shows an outline of the demodulation system of the reproduction page data. The image recognition processing unit 42 in FIG. 1 is actually an aggregate of a plurality of image recognition processing units corresponding to each area (3 × 3 pixels) of page data, and is an image recognition processing unit (convolutional neural network) to be constructed. The number of models) is the same as the number of symbols contained in the playback page data. The size of the page data used in this embodiment is 1740 pixels × 1044 pixels, and since there are 201,840 symbols in one page data, each symbol 1 to 201,840 corresponds to 201, 840 convolutional neural network models were constructed. With this configuration, the reproduction bit strings 1 to 201840 can be simultaneously reproduced from each of the image recognition processing units 42-1 to 42-18840, and one page data can be read at a time. The reproduction bit strings output in parallel are arranged in a predetermined order, and the input 1,009,200 bit recording bit strings are demodulated.

As the teacher data used for learning the Algo rubber rhythm, page data reproduced from the same hologram recording / reproducing device (past reproduction page data) and recording bit strings (recorded original data) are used. A large number of reproduced page data are aggregated for each symbol position in the page data and input to the convolution neural network. The result output from the convolutional neural network is compared with the recorded bit string, and each parameter of the convolutional neural network is learned using the error backpropagation method. By dividing the page data and performing machine learning for each position and for each modulation symbol, a 5: 9 demodulation algorithm that takes into account the noise source and optical transfer function that is generated depending on each position in the page data. It can be acquired mechanically (that is, the optical characteristics such as noise and aberration included in the optical system of the hologram recording / playback device and the page data demodulation algorithm are simultaneously acquired), and the data can be demolished with high accuracy.

In learning each parameter of the convolutional neural network, the optical characteristics are different for each position in the page data, and therefore the number of learnings required for optimizing the parameters is different depending on the optical characteristics. If the number of learnings is set uniformly for all convolutional neural networks without considering this, in the convolutional neural network that requires a large number of learnings for optimization, the parameters are not optimized. Will be completed, causing data errors. In addition, in a convolutional neural network that can optimize parameters with a small number of learnings, extra learning time may be required and at the same time overfitting may occur.

The region that requires a large number of learnings for optimization is a region that contains a large amount of optical noise, and it is necessary to learn both the demodulation algorithm of 5: 9 modulation and the optical noise characteristics. On the other hand, the region that can be optimized with a small number of learnings is a region with little optical noise, and only the demodulation algorithm of 5: 9 modulation needs to be trained. Therefore, by setting the learning frequency of each convolutional neural network with reference to the number of bit errors when demodulation is performed by the conventional method of hard judgment, the entire page data can be demodulated with a low error rate, and at the same time, learning is performed. The time required for the operation can be streamlined.

FIG. 4 shows the relationship between the bit error in demodulation by the hardness determination process and the number of learnings required for parameter optimization. The horizontal axis is the symbol number in the page data (corresponding to the symbol position, part of the area is extracted), △ indicates the number of bit errors (right axis), and the number of learnings required for the bar graph (contour display) (left axis). .. It should be noted that the thick line with 0 times is the overlap of the Δ display of the symbols without bit errors. It can be seen that the number of learnings required for parameter optimization tends to increase for symbols with many bit errors in the demodulation by the hard determination process, and it can be seen from FIG. 4 that there is a correlation between the two. Therefore, it is possible to efficiently complete the optimization of the convolutional neural network by learning according to the following procedure, for example.
(1) The entire reproduction page data is demodulated by the hardness determination, and the bit error for each area is calculated.
(2) It is necessary to learn the convolutional neural network using the data of several areas including the area with the most bit errors and the area with the least bit errors in the playback page data, and to optimize the parameters. Find the number of learning times.
(3) From the result of (2) above, the number of learnings according to the bit error rate in the hard determination demodulation is set for the entire area in the reproduction page data.
(4) The convolutional neural network is learned for the entire playback page data.

In the present invention, in order to construct an image recognition processing unit (neural network model) based on machine learning for page data in units of symbols (3 × 3 pixels), a neural network model is constructed for each pixel as in Non-Patent Document 2. Compared with the conventional technology, the system scale and the amount of calculation can be reduced. Further, the process of removing noise caused by the optical system, the process of determining an appropriate symbol, and the process of reproducing from the determined symbol, which are performed by a plurality of processing blocks as in the conventional hologram recording / reproducing device (FIG. 9). A series of processes such as a process of demodulating a bit string can be collectively performed by the image recognition processing unit 42, and the processing speed is improved.

FIG. 5 shows a comparison of bit errors by the demodulation method. FIG. 5 shows the result of actually demodulating 164 page data, the horizontal axis shows the symbol number (corresponding to the symbol position) in the page data, and the vertical axis shows the number of bit errors. The results of the hard determination method in which two points with high luminance are regarded as bright spots and demodulated are indicated by points Δ, and the results of the image recognition process of the present invention are indicated by ●. The image recognition processing algorithm used here is machine learning based on 700 page data. Compared with the conventional rigid determination method, the image recognition process of the present invention can suppress errors during demodulation in the entire page data, and the overall bit error rate is 1.59 × 10 from the conventional 4.03 × 10 ^-4 . The bit error was reduced to ^-4 by about 61%. The symbol with a lot of bit errors was a symbol in which dust was attached to the optical system. Even for such a pattern having constant noise, the image recognition process of the present invention can reduce bit errors by accumulating learning.

In this embodiment, an example of processing by a convolutional neural network is given as machine learning, but algorithms using a neural network other than the convolutional neural network, SVM (support vector machine), and other image recognition and pattern recognition are available. Machine learning may be used.

In the above embodiment, the configuration and operation of the hologram recording / reproducing device have been described, but the present invention is not limited to this, and may be configured as a method for reproducing a hologram. That is, in the process of reproducing the page data recorded on the hologram recording medium according to the data flow of FIG. 1, the process of extracting the symbol from the page data, and the image recognition process using an algorithm composed of machine learning from the symbol. It may be configured as a method of reproducing a hologram including a step of demolishing data.

Although the above embodiments have been described as typical examples, it will be apparent to those skilled in the art that many modifications and substitutions can be made within the spirit and scope of the present invention. Therefore, the present invention should not be construed as being limited by the above-described embodiments, and various modifications and modifications can be made without departing from the scope of claims. For example, it is possible to combine the plurality of constituent blocks described in the embodiment into one, or to divide one constituent block into one.

10 Recording processing unit 11 5: 9 Modulation unit 12 Page data conversion unit 20 Optical system 21 Recording optical system 22 Recording medium 23 Reproduction optical system 30 Reproduction processing unit 31 Symbol extraction unit 32 Brightness comparison unit 33 Bright spot position estimation unit 34 5: 9 Demodition unit 40 Reproduction processing unit 41 Symbol extraction unit 42 Image recognition processing unit 100 Hologram recording / playback device 101 Laser light source 102 Lens 103 Spatial filter 104 Collimating lens 105 1/2 wave plate 106 Mirror 107 PBS (polarized beam splitter)
108 PBS (Polarized Beam Splitter)
109 SLM (Spatial Light Modulation Element)
110 FT lens 111 1/2 wave plate 112 mirror 113 PBS (polarization beam splitter)
114 Galvano mirror 115 Relay lens 116 Hologram recording medium 117 Mirror 118 Galvano mirror 119 Relay lens 120 Two-dimensional image pickup element 150 Arithmetic logic unit (recording processing unit)
160 Arithmetic logic unit (reproduction processing unit)

Claims (3)

A hologram recording / playback device that restores data from a hologram recording medium recorded by a modulation method that converts n (n> 1) bit bit string data into a modulation symbol consisting of r pixels.
The page data reproduced from the hologram recording medium is demodulated from the modulated symbol to the bit string data by an image recognition process using an algorithm configured by machine learning using the past reproduced page data in advance. Hologram recording / playback device. In the hologram recording / reproduction apparatus according to claim 1, the reproduction page data is divided for each modulation symbol, the optical characteristics different for each position in the page data are machine-learned for each modulation symbol, and a demodulation algorithm different for each position is constructed. A hologram recording / playback device characterized by The hologram recording / reproducing device according to claim 1 or 2, wherein the machine learning is a convolutional neural network model, and the neural network model does not perform pooling to reduce dimensions.

Images