JP3574246B2 - Pattern recognition device - Google Patents

Description

[0001]
TECHNICAL FIELD OF THE INVENTION
The present invention relates to a pattern recognition device, and more particularly to an improvement in a pattern recognition rate.
[0002]
[Prior art]
In the pen-type pattern recognition device, characters are input using a tablet board. When there is a character input, the pen-type pattern recognition device obtains the xy coordinates of the pen tip, creates stroke information, and performs pattern matching or the like to recognize the input character.
[0003]
In the handwritten character recognition device disclosed in Japanese Patent Application Laid-Open No. 7-175893, instead of performing pattern matching, a feature value of a handwritten character input to a tablet board is given to a neural network operation unit, and the neural network operation unit recognizes the handwritten character. I am aware.
[0004]
[Problems to be solved by the invention]
However, tablet boards require a large amount of space and are therefore inconvenient to use, and there has been a demand for an apparatus capable of accurately recognizing an input pattern without using a tablet board.
[0005]
SUMMARY OF THE INVENTION The present invention has been made to satisfy such a demand, and it has been made possible to input characters and graphics without using a tablet board, and to recognize a pattern of characters and the like input using a neural network. The aim is to improve the quality.
[0006]
[Means for Solving the Problems]
A pattern recognition device according to the present invention includes a direction detection unit, a pressure sensing unit, a neural network operation unit, and a switching unit. The direction detection unit has an acceleration detection unit that detects the acceleration of the pen tip, and detects the moving direction of the pen tip on the writing surface. The pressure sensing section detects whether the pen tip is in contact with the writing surface. Further, the neural network operation unit analyzes and inputs information on the movement direction of the pen tip detected by the acceleration detection unit when the pressure sensing unit detects that the pen tip is in contact with the writing surface. Outputs information indicating the characteristics of the pattern. The switching unit switches the neural network coupling coefficient of the neural network operation unit according to the detection result of the pressure sensing unit. Therefore, with a simple configuration, the pattern drawn by the pen tip when the pen tip is in contact with the writing surface and when the pen tip is not in contact is detected, and the two patterns are compared to determine the pattern recognition rate when a graphic is input. Make it higher.
[0017]
BEST MODE FOR CARRYING OUT THE INVENTION
A pattern recognition device according to the present invention has a pen-like shape, recognizes an input pattern based on a moving direction and a moving amount of a pen tip moving on a writing surface, and outputs a recognition result to a host computer or the like. A pattern recognition device of the type, which detects a force in two directions applied to a pen tip so that the pattern recognition device can input characters and figures without using a tablet board or the like. Is used to improve the recognition accuracy of the pattern.
[0018]
The pattern recognition device has a direction detection unit and a calculation unit. The direction detecting unit detects a moving direction of the pen tip on the writing surface. The direction detection unit may include, for example, a strain detection unit having a strain sensor in two different directions perpendicular to the pen axis or an acceleration detection unit having an acceleration sensor in two different directions perpendicular to the pen axis at the pen tip. It detects the distortion or acceleration of the pen tip in two different directions perpendicular to the pen axis. Since the distortion or acceleration of the pen tip with respect to the pen tip is detected in two different directions perpendicular to the pen axis, the pattern recognition device detects the direction of movement of the pen tip by the direction detection unit, and thus moves the pen tip. By detecting the direction and the amount of movement and tracing the detected direction and amount of movement, characters and figures can be input. The amount of movement of the pen tip is obtained by, for example, integrating the acceleration detected by the acceleration detection unit twice.
[0019]
The calculation unit includes, for example, a preprocessing unit, a neural network calculation unit, and a pattern recognition unit. The preprocessing unit amplifies or binarizes data from the direction detection unit according to the configuration of the neural network operation unit, and converts the data into a signal suitable for input to the neural network operation unit. The neural network operation unit analyzes the input signal into several features based on coordination, competition, and self-organizing ability, and combines them to form a neural network, which is detected by the direction detection unit. Information about the direction of movement of the pen tip is input to analyze the input pattern and output information indicating the characteristics of the input pattern. The neural network operation unit, for example, samples data from the direction detection unit when a pattern to be recognized in advance is actually input, inputs these data to the input layer, and prepares neurons for the number of patterns to be classified into the output layer. Then, each pattern is associated with a neuron. When a certain pattern is input, learning is performed by, for example, a back propagation method or the like so that the output of the corresponding neuron becomes “1” and the output of other neurons becomes “0”. The neural network operation unit performs the neural network operation using the same coupling coefficient as the coupling coefficient thus obtained. The pattern recognition unit inputs the information indicating the feature output by the neural network operation unit, searches for the one with the largest output from the neuron, and outputs the corresponding pattern as a recognition result.
[0020]
In addition, pattern recognition is performed only when the pen tip is in contact with the writing surface to prevent unnecessary calculations, or to trace the pen tip when the pen tip is not in contact with the writing surface. Alternatively, the pattern recognition rate may be increased by comparing with the locus of the pen tip when the pen tip is in contact with the writing surface. Whether or not the pen tip is in contact with the writing surface is determined by using a pressure-sensitive portion having a pressure sensor or the like, judging from the detection result of the strain sensor, or by using an acceleration sensor or the like provided in the pen axis direction. . When judging from the detection result of the strain sensor, the neural network operation unit analyzes information on the force applied to the pen tip detected by the strain detection unit when the output of the strain detection unit is equal to or more than a certain value, and analyzes the characteristics of the input pattern. Is output. Here, the constant value is an output value of the distortion sensor when the pen tip portion is not in contact with the writing surface and a value that blocks noise.
[0021]
Further, in order to simplify the configuration of the neural network operation unit, the detection result of the direction detection unit is temporarily stored in the storage unit, and the neural network operation unit analyzes the input pattern based on the information stored in the storage unit. You may use it. If the detection result of the direction detection unit is directly input to the neural network operation unit, the neural network operation unit must be of a recurrent type, which complicates the configuration. However, once stored in the storage unit, the neural network has a simple configuration. The operation unit can analyze the input pattern.
[0022]
【Example】
FIG. 1 is a configuration diagram of a pattern recognition device according to one embodiment of the present invention. The pattern recognition device 1 has a pen shape as shown in the figure, and has a pressure sensing unit 2, a direction detection unit 3, a calculation unit 4, a storage unit 5, and a power supply unit 6, and a pen that moves on a writing surface. The input pattern is recognized based on the moving direction and the moving amount of the tip section 8, and the recognition result is output to a host computer or the like. The pressure sensing unit 2 includes a pressure sensor 2a. The pressure sensor 2a is provided on the pen tip 8, and detects pressure applied to the pen tip 8 from a writing surface. The movement detecting means 3 includes, for example, strain sensors 3a and 3b. The distortion sensors 3a and 3b are provided on the pen tip 8 of the pattern recognition device 1, and detect the distortion of the pen tip 8 in two directions orthogonal to the pen axis 9 when the pen tip 8 moves on the writing surface. Then, the direction of movement of the pen tip 8 is detected.
[0023]
The calculation unit 4 includes a preprocessing unit 41, a neural network calculation unit 42, and a pattern recognition unit 43, as shown in FIG. The preprocessing unit 41 converts the signals from the distortion sensors 3a and 3b into signals suitable for input to the neural network operation unit 42 by performing an amplification process, a binarization process, and the like. The pre-processing unit 41 amplifies an analog signal indicating the distortion of the pen tip 8 from the distortion sensors 3a and 3b and then converts the signal into a digital signal. In this case, the neural network operation unit 42 analyzes the digital signal indicating the moving direction of the pen tip 8 input from the input layer 421 and outputs information indicating the characteristics of the input pattern. Thereby, the pattern recognition device 1 can know the input timing of the pattern without providing a switch or the like for starting the pattern input in the pen tip portion 8.
[0024]
The neural network operation unit 42 performs a neural network operation based on the signal from the pre-processing unit 41. The neural network operation unit 42 analyzes the input signal into several features based on coordination, competition, and self-organizing ability. Are integrated to form a neural network as shown in FIG. 3. In this embodiment, the neural network is provided in the arithmetic unit 4. However, the neural network is composed of an analog neural network chip or a digital neural network chip. Is also good. For example, eight types of straight lines, two types of curves, and two types of bent lines are prepared as patterns to be recognized in advance, and data from the strain sensors 3a and 3b when these patterns are actually input are sampled. These data are input to the input layer 421, and neurons for the number of patterns to be classified are prepared in the output layer 423, and each pattern is associated with a neuron. When a certain pattern is input, learning is performed by, for example, a back propagation method so that the output of the corresponding neuron becomes “1” and the output of the other neurons becomes “0”. The neural network operation unit 42 performs a neural network operation using the same coupling coefficient as the coupling coefficient thus obtained. Since the data from the strain sensors 3a and 3b in this embodiment is time-series data, the neural network is configured as a recurrent type.
[0025]
The pattern recognition unit 43 is a unit that recognizes an input pattern based on an output signal from the neural network operation unit 42. The pattern recognition unit 43 searches for the largest output of the neuron of the output layer 423 of the neural network operation unit 42 learned in advance as described above, determines the corresponding pattern as the input pattern, and determines the determination result. Output to host device etc. As described above, by detecting the distortion of the pen tip 8, a pattern can be input from the pattern recognition device 1 without using a tablet or the like. The pattern to be input may be a figure or a basic component constituting a character. In the case of inputting a pattern of basic constituent elements constituting a character, recognition results of the basic constituent elements are sequentially obtained, and character recognition is performed by matching the obtained combination of the basic constituent elements with a dictionary. be able to. The operation unit 4, the storage unit 5, and the power supply unit 6 may be provided inside a host computer or the like.
[0026]
The operation of the pattern recognition device 1 having the above configuration will be described with reference to the flowchart of FIG.
[0027]
The preprocessing unit 42 amplifies an analog signal indicating the distortion of the pen tip 8 detected by the distortion sensors 3a and 3b, converts the signal into a digital signal, and inputs the digital signal to the input layer 421 of the neural network operation unit 42 (step S1). When the neural network operation unit 42 detects that the pen tip 8 has touched the writing surface and the pattern input has started based on the signal from the pressure sensor 2 (step S2), the input is input from the preprocessing unit 41 to the input layer 421. The output of the neuron corresponding to the obtained signal is set to “1”, and the outputs of the other neurons are set to “0”, the input signal pattern is largely classified, and the classification result is output to the intermediate layer 422 (step S3). After that, the neural network operation unit 42 classifies the input signal pattern into sub-classes by setting the output of the neuron corresponding to the large classification result input to the intermediate layer 422 to “1” and the other neurons to “0”, Output to the pattern recognition unit 43 (Step S4). The neural network operation unit 42 determines that the pattern input is continued while determining that the pen tip 8 is in contact with the writing surface, and repeats the above processing (steps S2 to S4) (step S2). S5). The pattern recognition unit 43 searches for a neuron having the largest output value based on the analysis result of the input pattern input from the neural network operation unit 42 and recognizes the input pattern (step S6).
[0028]
Here, since the neural network operation unit 42 performs the neural network operation, for example, based on the input pattern in which the components of the character as shown in FIG. 5A are separated from each other, as shown in FIG. A correct input can be recalled by itself, and the recognition rate of the input pattern can be increased. As described above, since the forces applied to the pen tip 8 in two different directions perpendicular to the pen axis 9 are detected using the strain sensors 3a and 3b, characters, figures, and the like can be input on an arbitrary plane. In addition, by using the neural network operation unit 42, the recognition rate of the input pattern can be increased. Also, when the pen tip 8 is not in contact with the writing surface, the neural network operation unit 42 does not perform the neural network operation, so there is no need to perform useless operation, and the processing speed is improved particularly when complicated processing is performed. can do.
[0029]
In the above embodiment, the strain sensors 3a and 3b are used to detect the force applied to the pen tip to determine the moving direction and the amount of movement of the pen tip 6, but instead of the strain sensors 3a and 3b, FIG. The acceleration sensors 3c and 3d may be used as shown in FIG. By using the acceleration sensors 3c and 3d, the movement trajectory of the pen tip 8 can be traced even when the pen tip 8 is not in contact with the writing surface.
[0030]
The pen-axis-direction acceleration detection unit 10 having an acceleration sensor for detecting acceleration in the pen-axis direction is provided, and it is determined whether or not the pen tip 8 is in contact with the writing surface based on a signal from the acceleration sensor. You may do it. This makes it possible to detect whether or not the pen tip 8 is in contact with the writing surface, and to detect the inclination of the pen shaft 9 and correct the detection result of the direction detection unit 3 to more accurately detect the pen. The direction and amount of movement of the tip 8 can be detected.
[0031]
In addition, for example, when a character cannot be written with one stroke as shown in FIG. 5A, it may be difficult to recognize the entire character pattern. Therefore, as shown in FIG. A neural network operation unit 42a for recognizing a pattern drawn by the pen tip 8 when in contact with the user, and a neural network operation unit 42b for recognizing a pattern drawn by the pen tip 8 when not in contact with the writing surface are provided. You may do it. In this way, by separately classifying the input patterns when the pen tip 8 is in contact with the writing surface and when the pen tip 8 is not in contact with the writing surface, the two classification results can be compared and compared. By making them continuous, the shape of the entire input pattern can be easily known, and the recognition rate of the input pattern can be further increased. Whether the pattern recognized by the neural network operation units 42a and 42b is a pattern divided within one character component or a different character component is determined by using a timer or the like for a certain period of time. When the tip part 8 is separated from the writing surface, it is determined that the character is different.
[0032]
In addition, since the configuration becomes complicated when the neural network operation unit 42a and the neural network operation unit 42b are provided, a switching unit 44 as shown in FIG. 8 may be provided. The neural network operation unit 42 learns in advance the pattern of the trajectory drawn by the pen tip unit 8 to form one character in each case when the pen tip unit 8 is in contact with the writing surface and when it is apart from the writing surface. In this case, the neural network coupling coefficient is obtained in advance, and the switching unit 44 switches the neural network coupling coefficient when the pen tip 8 is in contact with the writing surface and when it is apart from the writing surface. Thus, the pattern recognition rate of the pattern recognition device 1 can be further increased with a simple configuration.
[0033]
In the above embodiment, a recurrent type neural network is configured. However, information indicating the moving direction of the pen tip 8 may be temporarily stored in the storage unit 5, and a neural network other than the recurrent type may be used. good. As shown in FIG. 9, after the signals from the strain sensors 3a, 3b or the acceleration sensors 3c, 3d are processed by the pre-processing unit 41, for example, only when the pen tip 8 is in contact with the writing surface, the strain sensors 3a, 3b The signal from the acceleration sensor 3b or the acceleration sensors 3c and 3d is temporarily stored in the storage unit 5. For example, the pen tip unit 8 divides the data from the pre-processing unit 41 from the time it comes into contact with the writing surface to the time it separates from the writing surface into 32 pieces, and inputs the data to the neural network operation unit 42. The neural network operation unit 42 outputs an analysis of the input pattern based on the data stored in the storage unit 5. Thus, the neural network operation unit 42 can employ a neural network other than the recurrent type, and the configuration of the neural network operation unit 42 can be simplified.
[0035]
Further, as shown in FIG. 10 , a thinning-out interpolation unit 45 may be provided to adjust the influence of the moving speed of the pen tip unit 8. Thinning interpolation section 45 is adjusted to the data relating to the force applied to the pen tip portion 8 stored in the storage unit 5, for example, the data amount in the writing of the same pattern becomes constant. For example, the thinning-out interpolation unit 45 performs the thinning-out processing when the data amount stored in the storage unit 5 is large, and performs the interpolation processing when the data amount stored in the storage unit 5 is small.
[0036]
In the above embodiment, the pressure sensor 2a is used for the pressure sensing unit 2, but a micro switch or the like may be used.
[0037]
【The invention's effect】
As described above, the present invention provides a pattern recognition device including a direction detecting unit, a pressure sensing unit, a neural network operation unit, and a switching unit. The direction detection unit has an acceleration detection unit that detects the acceleration of the pen tip, and detects the moving direction of the pen tip on the writing surface. The pressure sensing section detects whether the pen tip is in contact with the writing surface. Further, the neural network operation unit analyzes and inputs information on the movement direction of the pen tip detected by the acceleration detection unit when the pressure sensing unit detects that the pen tip is in contact with the writing surface. Outputs information indicating the characteristics of the pattern. The switching unit switches the neural network coupling coefficient of the neural network operation unit according to the detection result of the pressure sensing unit. Therefore, with a simple configuration, the pattern drawn by the pen tip when the pen tip is in contact with the writing surface and when the pen tip is not in contact is detected, and the two patterns are compared to determine the pattern recognition rate when a graphic is input. Make it higher.
[Brief description of the drawings]
FIG. 1 is a configuration diagram showing an embodiment of the present invention.
FIG. 2 is a configuration diagram illustrating a configuration of a calculation unit.
FIG. 3 is a configuration diagram of a neural network.
FIG. 4 is a flowchart showing the operation of the present invention.
FIG. 5 is an explanatory diagram illustrating an example of self-recall of a neural network operation unit.
FIG. 6 is a configuration diagram when an acceleration sensor is used for a direction detection unit.
FIG. 7 is a configuration diagram of a calculation unit when two neural network calculation units are provided.
FIG. 8 is a diagram illustrating a configuration part of an arithmetic unit when a switching unit is provided.
FIG. 9 is a configuration diagram of a calculation unit when a detection result of a direction detection unit is stored in a storage unit.
FIG. 10 is a configuration diagram of a calculation unit in a case where a pattern having a different input pattern size is determined .
[Explanation of symbols]
Reference Signs List 1 pattern input device 2 pressure sensing unit 2a pressure sensor 3 direction detection unit 3a distortion sensor 3b distortion sensor 3c acceleration sensor 3d acceleration sensor 4 operation unit 41 preprocessing unit 42 neural network operation unit 43 pattern recognition unit 44 switching unit 45 thinning interpolation Unit 5 storage unit 8 pen tip 10 pen axis direction acceleration detection unit

Claims (1)

In a pen-type pattern recognition device that recognizes an input pattern based on a moving direction and a moving amount of a pen tip moving on a writing surface,
A direction detection unit that has an acceleration detection unit that detects the acceleration of the pen tip unit and detects a moving direction of the pen tip unit on the writing surface,
A pressure-sensitive unit that detects whether the pen tip is in contact with the writing surface,
When the pressure-sensitive portion detects that the pen tip portion and the writing surface are in contact with each other, the information on the movement direction of the pen tip portion detected by the direction detection portion is analyzed, and the characteristics of the input pattern are analyzed. A neural network operation unit that outputs information indicating
A switching unit that switches a neural network coupling coefficient of the neural network operation unit according to a detection result of the pressure sensing unit;
Pattern recognition apparatus comprising: a.

Images