KR19990081569A - Feature Vector Value Extraction Method for Character Recognition - Google Patents

Abstract

Detecting an outline of a pixel array having a predetermined size by normalizing an image of a character block, and then retrieving a pixel unit and storing the outline point coordinate group consisting of point coordinates of each pixel existing on the connected contour path; Storing a normalized point coordinate group which is a point coordinate group in which the coordinate number of the stored outline point coordinate group is normalized to a predetermined number of point coordinates; Calculating and storing adjacent point coordinate angles, which are angles of line segments connecting adjacent point coordinates on contour paths of the normalized point coordinate group; And extracting an output of the artificial neural network as a feature vector value by applying the adjacent point coordinate angle as an input of the artificial neural network.

According to the present invention, in a character recognition method using an artificial neural network, after detecting a character outline, a feature value for character recognition is extracted by extracting an angle between the contour pixels along a path of the character outline as a feature vector value of the character. The advantage is that feature vector values can be extracted more accurately and faster.

Description

Feature Vector Value Extraction Method for Character Recognition

The present invention relates to a feature vector value extraction method for character recognition. More particularly, in a character recognition method using an artificial neural network, after detecting a character outline, the angle between the contour pixels along the path of the character outline is determined. The present invention relates to a feature vector value extraction method for character recognition that extracts a feature vector value.

The characteristics of the human brain are represented by a high degree of parallelism, fault tolerant, distributed representation of knowledge, and distributed control, which are based on the unique neural tissues that humans possess. Human neural tissue consists of nerve cells called neurons, as shown in FIG. 1.

Neurons are cells that have been specially differentiated for information processing in a living body. They have a cell body (soma), which is a main body, and a part called complex dendrites. It is divided into three parts called axons.

Axons are fibers that carry signals from the body of the cell to other neurons, and dendrites are the parts that receive signals from other neurons. That is, the end of the axon of another neuron is connected to it. This connection is called a synapse.

The potential inside the neuron is usually lower than it is outside, but when it passes through the axon and the input signal arrives from the outside, the neuron 'excitory' under certain conditions and the internal potential suddenly rises. In this case, the neurons are said to be 'fire', and a pulse of about 0.1 m is transmitted through the axon as a time width and a voltage of 1 msec, which is transmitted to other neurons as a signal of the neuron. When electric pulses from other neurons reach the dendrites, the potential here fluctuates slightly. When a pulse comes in several places of the dendrite, the voltage of several places fluctuates slightly. This variation reaches and adds to the body, and when the sum exceeds a certain threshold, this triggers the neuron to ignite, and if no threshold is reached, no reaction occurs.

Numerical models of neurons have large linear and nonlinear threshold characteristics, which are what neurons multiply and add to signals from other neurons, and nonlinear threshold characteristics do nothing unless their weighted sum exceeds the threshold. On the other hand, if the weighted sum exceeds a threshold, it is said to be a nonlinear operation that emits a pulse. This is explained through the well-known MP (McCulloch and Pitts) model as follows.

In the MPCulloch and Pitts (MP) model, the neuron outputs 1 if it satisfies Equation (1), and 0 otherwise.

w ₁ x ₁ + w ₂ x ₂ + ⃛ + w _M x _M ＞ T

D = w ₀ x ₀ + w ₁ x ₁ + ⃛ + w _M x _M

here, x ₁ , x _2, ⃛, x _M Is the input pattern, w ₁ , w ₂ , ⃛, w _M Is the weights. At this time, w ₀ = -T, x ₀ = 1 When is of formula (2) D> 0 If the output is 1 D <0 If it is, the output is zero. only, w ₀ Becomes a bias weight.

Connectionist models, parallel distributed processing models, neural morphological systems, or artificial neural network models, sometimes referred to simply as "neural networks", provide high classification performance through the interconnection between simple computational elements (ie, nodes). The purpose is to obtain.

The difficulty with the use of artificial neural networks is to find weights that solve the problem with proper operation. Once the weights are found correctly, it is easy to classify the samples using artificial neural networks.

Artificial neural networks are defined by engineering human models of biological neurons, reflecting the concepts described above, and are often arranged as sensory layers (input layers), connection layers (hidden layers), and reaction layers (output layers). Learning to use a computational model and modifying the state of your brain based on these computational models can be regarded as a model that can solve your problems.

So far, the proposed artificial neural network has various forms according to the activation function of the neuron, the structure of the network, and learning rules for adjusting the connection strength of the neuron. Representative artificial neural networks include the basic neural network, the perceptron, and the hopfield. Hopfield network, self-organizing maps, neocognitron, backpropagation, Adaptive Resonance Theory, Boltzman machine, etc. There is this.

General character recognition method will be described in order to help the understanding of the present invention to recognize a character by learning in relation to the artificial neural network as described above.

2 is a flowchart illustrating a general character recognition method.

First, in the original image input step S10, an original image is input from a scanning image obtained by scanning an original through a document reading means such as a scanner or an image file previously stored in a storage medium, and then, in the preprocessing step S20. Preprocessing is performed to remove the noise and to separate the background and text areas to make the original image suitable for text recognition.

Subsequently, in the feature vector value extraction step (S30), feature vector values defined to reflect characteristics of the character are extracted by using various methods for the corresponding character separated into the character area.

In the recognition algorithm step (S40), the artificial neural network or template matching, fuzzy algorithm, semantic method, genetic algorithm, and dynamic programming described above are used. A character corresponding to this feature vector value is recognized using a pattern recognition algorithm such as programming.

Subsequently, in the post-processing step (S50), the above process is repeatedly performed to restore the original image for the recognized character portion and to process the next character.

As with most recognition algorithms including artificial neural networks described above, in order to provide a relatively high reliability for recognition results, a relatively large number of feature vector values need to be extracted, which naturally results in delaying the convergence speed. The number of vector values and the reliability of the recognition result can be seen as having a mutual tradeoff.

However, in the case of the artificial neural network, if too many feature vector values are applied to the input layer to the artificial neural network, a relatively high number of nodes of the hidden layer is required to output a highly accurate feature vector value. Not only does it take a great deal of time to learn artificial neural networks, but the complexity increases and the probability that an incorrect feature vector value is output increases, which leads to a problem of lowering the actual recognition rate and recognition speed.

Accordingly, in order to optimize the number of feature vector values to reduce the complexity of the artificial neural network, improve the recognition speed, and at the same time obtain a high recognition rate, not only the uncontaminated printed character pattern but also the geometrically deformed or somewhat damaged recognition object pattern Even if is inputted, it should be possible to extract a feature vector value that can fully reflect the characteristics of the character string to be recognized.

Accordingly, the present invention has been made to reflect such a necessity in the character recognition process. In the character recognition method using an artificial neural network, after detecting a character outline, the angle between the contour pixels along the path of the character outline is converted into a character. It is an object of the present invention to provide a feature vector value extraction method for character recognition that can extract a feature vector value more accurately and quickly by extracting the feature vector value.

1 is an exemplary view showing the structure of neurons of the human nerve,

2 is a flowchart illustrating a general character recognition method;

3 is a block diagram showing a preferred embodiment of a feature vector value extraction method for character recognition according to the present invention;

4 is an exemplary view showing an example of extracting a feature vector value using the feature vector value extraction method for character recognition according to the present invention.

In order to achieve the above object, the feature vector value extraction method for character recognition according to the present invention detects an outline of a pixel array having a predetermined size normalizing an image of a character block, and then searches by pixel unit to search for the contour. Storing outline point coordinate groups composed of point coordinates of each pixel existing on the contour path;

Storing a normalized point coordinate group which is a point coordinate group in which the coordinate number of the stored outline point coordinate group is normalized to a predetermined number of point coordinates;

Calculating and storing adjacent point coordinate angles, which are angles of line segments connecting adjacent point coordinates on contour paths of the normalized point coordinate group; And

And applying the adjacent point coordinate angle as an input of the artificial neural network to extract the output of the artificial neural network as a feature vector value.

Here, the adjacent point coordinate angle is preferably trigonometrically calculated for the pair of adjacent point coordinates.

Hereinafter, a preferred embodiment of a feature vector value extraction method for character recognition according to the present invention will be described with reference to the accompanying drawings.

3 is a block diagram illustrating a preferred embodiment of a feature vector value extraction method for character recognition according to the present invention. For the sake of consistent description, the same reference numerals will be given to the same process as the prior art.

As shown in FIG. 3, in the original image input step S10, a scanning image obtained by scanning an original through a document reading means such as a scanner, is illustrated in FIG. 3. After receiving the original image from the image file or the like previously stored in the storage medium, the preprocessing step (S20) performs a preprocessing process to remove the miscellaneous images or to separate the background region and the character region so that the original image is suitable for character recognition. .

Thereafter, the feature vector extraction process of the present invention is performed. First, in step S100, as shown in FIGS. 4A, 4B, and C, N × N normalizing an image of a character block is performed. After the contour is detected by the pixel array of the size, the contour is searched in units of pixels, and the contour point coordinate group consisting of the point coordinates of each pixel existing on the connected contour path is stored.

Thereafter, in step S110, a normalized point coordinate group, which is a point coordinate group in which the number of coordinates of the stored outline point coordinate group is normalized to a predetermined number of point coordinates, is stored, and in step S120, point coordinates adjacent to each other on a contour path of the normalized point coordinate group The angle of adjacent point coordinates, which is the angle of the line connecting them, is calculated and stored trigonometrically.

In more detail, the coordinates of adjacent pairs of point coordinates (x _1, y ₁ ) , P2 (x _2, y ₂ ) When, the adjacent point coordinate angle is as shown in equation (3).

For example, when the coordinates of the pair of point coordinates adjacent to each other are (3, 4) and (5, 6), the adjacent point coordinate angles can be easily calculated as equation (45).

Subsequently, in step S130, the adjacent point coordinate angle is applied as an input of the artificial neural network, and in step S140, the output of the artificial neural network is extracted as a feature vector value.

Then, as in the prior art, in the recognition algorithm step S40, an artificial neural network or template matching, fuzzy algorithm, semantic method, genetic algorithm, dynamic A pattern recognition algorithm such as dynamic programming is used to recognize a character corresponding to this feature vector value. At this time, it is preferable to use an artificial neural network.

Finally, in the post-processing step (S50), the above process is repeatedly performed to restore the original image for the recognized character portion and process the next character.

Terminologies used herein are terms defined in consideration of functions in the present invention, which may vary according to the intention or customs of those skilled in the art, and the definitions should be made based on the contents throughout the present application. will be.

In addition, since the present invention has been described through the preferred embodiment of the present invention, in view of the technical difficulty aspects of the present invention, those having ordinary skill in the art can easily be different from another embodiment of the present invention. Since modifications may be made, it is obvious that both the embodiments and modifications cited in the above description belong to the claims of the present invention.

As described in detail above, in the character recognition method using an artificial neural network, after detecting the character outline, the character according to the present invention extracts the angle between the contour pixels along the path of the character outline as a feature vector value of the character. According to the feature vector value extraction method for recognition, there is an advantage that the feature vector value can be extracted at a more accurate and faster speed when the feature value for character recognition is extracted.

Claims (2)

Detecting an outline of a pixel array having a predetermined size by normalizing an image of a character block, and then retrieving a pixel unit and storing the outline point coordinate group consisting of point coordinates of each pixel existing on the connected contour path; Storing a normalized point coordinate group which is a point coordinate group in which the coordinate number of the stored outline point coordinate group is normalized to a predetermined number of point coordinates; Calculating and storing adjacent point coordinate angles, which are angles of line segments connecting adjacent point coordinates on contour paths of the normalized point coordinate group; And And extracting the output of the artificial neural network as a feature vector value by applying the adjacent point coordinate angle as an input of the artificial neural network. The method of claim 1, wherein the adjacent point coordinate angle, Characteristic vector value extraction method for character recognition, characterized in that the trigonometric operation on the adjacent pair of point coordinates.