JPH03121581A - Character reader - Google Patents

Abstract

PURPOSE:To enable parallel arithmetic, to improve speed for recognizing handwritten characters and to make a character reader compact by using a three- dimensional memory cell in the intermediate layer of a neural circuit network. CONSTITUTION:An input layer is provided to input the picture information of the characters to be read and the intermediate layer, which is composed of the three-dimensional memory cell three-dimensionally forming plural memories, is provided to execute the arithmetic processing of the above mentioned character picture information based on a state transition function to be determined by the storing state of each memory cell. Then, an output layer is provided to recognize the prescribed characters from the arithmetic result of this intermediate layer. Namely, for a three-dimensional memory cell 10 forming the intermediate layer, plural laminated memory cells 11 are integrated in the shape of a matrix on a board. Thus, the characters including the handwritten characters can be recognized with a high recognition rate at extremely high speed and the character reader can be made compact.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to a character reading device that can read various characters including handwritten characters using artificial intelligence.

[Conventional technology]

In today's highly developed information society, letters.

There is an extremely high demand for automatically reading information such as alphanumeric characters and inputting it online into information processing systems such as computers. For example, when considering a conversation between Japanese, including kanji, and a computer, the number of characters to be input is enormous, and these characters can be input using a keyboard using the alphabet or hiragana (katakana). This requires extremely complicated work.

Therefore, character reading devices that can automatically read characters, alphanumeric characters, handwritten characters, etc. and input them into an information processing system or the like online have been considered.

When reading handwritten characters using this type of character reading device, the handwritten characters are read by the image reading device, the read character image information is encoded, and this encoded data is temporarily stored in a storage device. The encoded data is sequentially read out and processed by a program to recognize predetermined characters. Generally, character recognition rate can be improved by dividing character image information into smaller pieces and encoding them.

However, in order to increase the character recognition rate, it is necessary to obtain encoded data in which character image information is divided into smaller pieces, and it is necessary to further increase the resolution of the image reading device. Also,
As the character image information is segmented, the amount of encoded data increases, so it is necessary to increase the storage capacity of the storage device. Moreover, since the amount of data to be processed by the program increases, a faster calculation speed than before is required.

By the way, in the field of artificial intelligence, pattern recognition technology that recognizes character patterns using neural networks is known. In particular, the parallel distributed processing model (which is a type of intelligent processing model)
The PDP model (hereinafter referred to as the "PDP model") can learn each individual's handwritten character habits using a backpropagation method (backpropagation learning agent). Therefore, if the FDP model is applied to a character reading device to perform character pattern recognition, a sufficiently high character recognition rate can be obtained even if the encoded data is not segmented.

[Problem to be solved by the invention]

However, in character pattern recognition using the FDP model, encoded data is processed in parallel (matrix calculation processing), which increases the number of calculations and takes a long time to process on a conventional serial computer. There is a problem. Also, conventional computers (Neumann type)
In order to perform parallel calculations in parallel, shorten the calculation processing time, and recognize a practical number of characters such as kanji, it is necessary to use multiple C
Although it is necessary to use a PU, since each CPU does not perform multiple calculation processes at the same time, there is a problem in that the device becomes large in order to actually perform multidimensional parallel calculations.

The present invention has been made in view of the above-mentioned circumstances, and an object of the present invention is to provide a character reading device that can recognize characters including handwritten characters with a high recognition rate and at an extremely high speed, and can also be miniaturized. With the goal.

[Means to solve the problem]

In order to solve the above problems, the present invention uses a neural network configured based on an intelligent processing model with a learning function to input image information of characters to be read in a character reading device that reads characters. an input layer that processes the character image information based on a state transition function determined by the storage state of each memory cell, and an intermediate layer that is composed of a three-dimensional memory element in which a plurality of memory cells are formed in a three-dimensional shape; , and an output layer that recognizes predetermined characters from the calculation results of the intermediate layer.

Further, in the character reading device, there is provided a laminated tunnel switch section in which conductive films and insulating films are alternately laminated and charges are tunnel-conducted through the insulating films, and an insulating film is provided at both ends of the laminated tunnel switch section in the stacking direction. The intermediate layer is a three-dimensional memory element formed by two-dimensionally integrating a stacked memory element consisting of a pair of electrodes provided through the conductive film and a plurality of charge storage capacitors respectively connected to each of the conductive films. I decided to use it.

[Function] The present invention achieves the following effects by taking the above measures. For example, the weights of the connections in the neural network that change when the neural network learns character recognition using the backpropagation method are expressed by changes in the storage states of a plurality of memory cells of the three-dimensional memory element. Therefore, with a simple operation such as changing the storage state of a memory cell,
It can support learning of arithmetic processing models using backpropagation. In addition, the 3D memory element allows simultaneous parallel access to multiple memory cells, so when character image information is input from the input layer and reaches the intermediate layer, it is processed within the 3D memory element (intermediate layer). , arithmetic processing based on a state transition function determined by the storage state of each memory cell is executed in parallel.

In the output layer, a predetermined character is recognized from the calculation result, and the recognition result is output. Characters are read in this way.

Furthermore, if the three-dimensional memory element described above is used, each charge storage capacitor that becomes a memory cell is in a floating state, so each memory cell can be easily accessed, and the memory state can be easily rewritten in accordance with learning results. I can do it. Furthermore, since each memory cell can be accessed in parallel via the write electrode and read electrode, it is possible to support parallel arithmetic processing.

〔Example〕

Embodiments of the present invention will be described below with reference to the drawings.

The character reading device according to the present embodiment includes an input means, such as a line scanner, which takes in character image information such as handwritten characters by photoelectric conversion, and recognizes a predetermined character from the character image information input from this input means. and an output means that processes and outputs the character data recognized by the recognition processing means into a state that can be used in an external device such as an information processing system.

In this embodiment, a neural network of an FDP model is used as a recognition processing means to recognize handwritten characters.As shown in FIG. 1, the FDP model includes an input layer, an intermediate layer,
It consists of a hierarchical network consisting of an output layer.

Here, the input layer of the neural network is composed of 8×8 (64) units and has a function of dividing character image information taken in by the input means into 64 signals.

The intermediate layer is composed of a three-dimensional memory element, and has a function of repeating state transition according to a state transition function described later when a signal is applied from the input layer, and then outputting a signal that reaches a predetermined equilibrium state. The output layer consists of four units that identify four types of characters, and has the function of recognizing a predetermined character from among the four types of characters based on the signal that has reached an equilibrium state in the intermediate layer. There is. Further, each unit of the output layer is connected to all units of the intermediate layer. Some units in the middle layer are connected to all units in the input layer, while others are connected only to specific units in the human power layer. The neural network connected in this way is of a feedforward type in which signals are transmitted in one direction from the human power layer to the output layer.

FIG. 2 is a diagram showing the configuration of a three-dimensional memory element forming an intermediate layer. The three-dimensional memory element 10 shown in the figure is one in which a plurality of stacked memory elements 11 are integrated on the same substrate in a matrix of three rows and three columns. Each stacked memory element 11
A write electrode 12 is provided at one end of each stacked memory element 11, and a read electrode 13 is provided at the other end of each stacked memory element 11.

Further, each write electrode 12 located in the same row is connected to matrix electrode j+-33, and read electrode 13 located in the same column is connected to matrix electrodes L1 to i3.

As shown in FIG. 3, the stacked memory element 11 includes insulating films 22 a to 22 made of LB films (Langmuir-Blodgett films) whose conductivity changes nonlinearly depending on the applied voltage.
A multilayer tunnel switch @H in which conductive films 23a to 23d are alternately laminated, a write electrode 12 provided via an insulating film 22a serving as one end of the multilayer tunnel switch section H in the stacking direction, and a multilayer tunnel switch section H. , a readout electrode 13 provided at the other end of H with an insulating film 22e interposed therebetween;
The charge storage capacitors 01 to C4 are connected to the conductors H23a to 23d, respectively, and the read pulse application terminal 18 is connected to the readout electrode 13. readout capacitor C5,
It is composed of a readout circuit including a switch Sl, an ammeter 19, and the like. FIG. 3 shows a stacked memory element 11 on a substrate 2.
0 is shown.

In the stacked memory element 11 configured in this way, charges are tunnel-transferred from the write electrode 13 to the conductive film 23a by applying, for example, a positive pulse from the write electrode 13 and a negative pulse from the terminal 16. . Then, by applying transfer pulses VA, V having mutually different phases from the terminals 16 and 17, the charges written from the write electrode 12 are sequentially transferred to the read electrode 13 side,
The charges conducted to each conductive film 23 are stored in a charge storage capacitor C connected to the conductive film.

Then, by applying a predetermined pulse voltage to the read pulse application terminal 18, the charge stored in the charge storage capacitor C4 is read out to the capacitor C5, and by closing the switch S1, the stored information is read out. Output. Note that the stacked memory element 11 shown in FIGS. 3 and 4 can store a 4-bit signal.

Next, the operation of this embodiment will be explained.

Character image information of handwritten characters read by the input means is divided by 64 units of the input layer, and signals are given to a plurality of units of the intermediate layer connected to each unit of the input layer. In the intermediate layer, each unit of the intermediate layer is weighted according to the charge storage state of the charge storage capacitors C1 to C4 of each stacked memory element 11, and a signal transmitted between each unit is amplified or attenuated by the weight. , when the voltage at each unit exceeds a threshold, it outputs a signal to the other connected units. Here, in the FDP model, the weight of the connection from unit l to j is Wij, and the voltage (1,5
V) is the threshold value θj of unit j, the state transition function of the intermediate layer is Xi-ΣYLWji+θj...(1)Y
j = 1/ (1+ e-”) ・=
It can be expressed as (2). Note that Xj represents the input sum of unit j, and Yj represents the output of unit j.

In the intermediate layer, signals from the input layer are transmitted to each stacked memory element 11.
This input signal is input from the write electrode 12 of the stacked memory element 11, processed according to a state transition function according to the charge storage state of the stacked memory element 11, and outputted from the read electrode 13 to the output layer.

That is, when a signal is given to each unit in the middle layer,
A signal propagates within the neural network according to the state transition function, and a signal that has reached a predetermined equilibrium state is output from each unit of the output layer.

Next, the neural network is trained using backpropagation. Therefore, a teacher signal is given to the output signal output from the output layer as described above, and the performance of the neural network in the current connection weight state is evaluated using the function shown in equation (3).

E-(1/2) ・Σ(Yj-tj)'...(3) Note that E is the sum of errors (sum of squared errors), Yj is the output signal of unit j in the output layer, and tj is the output signal of unit j in the output layer. It shows the teacher signal given to unit j.

Then, using equation (4), the amount of change ΔW in the connection weight that minimizes the sum of squared errors E is determined.

AW --e a E/ i3W j i
-(4) Note that ε is a constant indicating the learning rate. Further, aE/aWji can be obtained by calculating 'a E/a Y j. In reality, since there is an intermediate layer, 'aE/9Yj cannot be determined directly, but for any unit j in the intermediate layer, only a specific unit in the output layer connected to that unit is affected. do not have. Therefore, 9 E / e Y j can be obtained from equation (5).

... (5) According to the change amount ΔW obtained in this way, the weight W of the connection is
Change ij. The connection weight Wij is changed by storing or erasing charges in each of the charge storage capacitors 01 to C4 of the stacked memory element 11.

By using the backpropagation method as described above, it is possible to make a neural network learn, for example, the characteristics of an individual's handwritten characters.

Next, character recognition was performed using the trained neural network of the FDP model. As a result, 98,5 for the training sample
showed a kanji recognition rate of over %.

As described above, according to this embodiment, the write electrodes 12 provided on each of the plurality of stacked memory elements 11 that are two-dimensionally integrated are
and readout electrode 13, matrix electrode j1""J3
111 to 111i, each stacked memory element 11 can be accessed independently.

In addition, since the three-dimensional memory element 10 in which a plurality of stacked memory elements 11 are integrated is used in the intermediate layer of the neural network configured based on the FDP model, the connection weight Wij of each unit that becomes a neuron is This can be changed by an extremely simple operation such as changing the charge storage state of each stacked memory element 11. Also, state transition functions (1), (2)
can be calculated simultaneously and in parallel for all values of j,
Therefore, compared to the conventional method of serial calculation, calculation speed can be dramatically improved, and the time required to read handwritten characters can be shortened.

In addition, since it uses a three-dimensional memory element 10 with an extremely high density per unit area, it is possible to process large amounts of information at high speed, allowing for the recognition of a large number of kanji characters, which could previously only be done with large computers, into a compact computer. This can be done using a device.

〔Effect of the invention〕

As detailed above, according to the present invention, since a three-dimensional memory element is used in the middle layer of the neural network, parallel calculations are possible, the recognition speed of handwritten characters can be dramatically improved, and the device can be made smaller. It is possible to aim for

[Brief explanation of drawings]

Figure 1 is a diagram showing the conceptual configuration of a neural network, Figure 2 is a diagram showing the configuration of the intermediate layer, Figure 3 is a diagram showing the configuration of a stacked memory element, and Figure 4 is a diagram showing the configuration of the stacked memory element shown in Figure 3. FIG. DESCRIPTION OF SYMBOLS 10... Three-dimensional memory element, 11... Laminated memory element, 12... Write electrode, 13... Read electrode, 2
2 a to 22 e-insulating film, 23 a to 23 d
--- Conductive film.

Claims (3)

[Claims] (1) A character reading device that reads characters using a neural network configured based on an intelligent processing model with a learning function has an input layer that inputs image information of the characters to be read, and multiple memories. an intermediate layer consisting of a three-dimensional memory element in which cells are formed in a three-dimensional shape, and which performs arithmetic processing on the character image information based on a state transition function determined by the storage state of each memory cell; 1. A character reading device comprising: an output layer that recognizes predetermined characters. (2) In the character reading device according to claim 1, the three-dimensional memory element includes a laminated tunnel switch section in which conductive films and insulating films are alternately laminated and charges are tunnel-conducted through the insulating films, and the laminated tunnel switch. A multilayer memory element is two-dimensionally constructed, which includes a write electrode and a read electrode provided at both ends in the stacking direction of the section through an insulating film, and a plurality of charge storage capacitors connected to each of the conductive films. A character reading device characterized by being integrated into. (3) In the character reading device according to claim 1, the neural network is constructed based on a parallel distributed processing model, and the neural network is trained by a backpropagation method (error backpropagation method). A character reading device featuring: