TWI685796B - A method for character pattern recognition - Google Patents

Abstract

The present invention provides a method for character pattern recognition. A network of a single hidden layer is chosen; one uses an initialized figure to set the number of hidden layer elements and then generates a parameter randomly. By using a weight of an output layer link calculated by character-training images going through the network, and establishes a new neuron and generates its parameter by random number. Calculating a weight of the output link of the neuron by the character-training images, then determines whether the current neuron should be discarded according to the the errors of the overall recognition after adding the neuron to the current network. One repeats the steps of establishing neurons and determining if the neurons should be discarded until a stopping condition is met, then one inputs the character-training images and begins character pattern recognition. Said network only needs a single hidden layer while the training is for calculating output link parameters and also ensures newly-added neurons to decrease recognition errors. Besides the simplification of the nerve network structure and the acceleration of its training, decrease in recognition time and precision is also achieved.

Description

Intelligent text and figure recognition method

The invention relates to a text and figure recognition method, in particular to a smart text and figure recognition method based on machine learning, which has a simplified computing system architecture, and can improve recognition accuracy and reduce the time required for recognition.

Graphic recognition technology is the origin of artificial intelligence, and it is also one of the main technologies driving various artificial intelligence products. At present, graphic recognition technology has been widely used in various electronic products and large-scale systems, including handwritten text recognition, license plate recognition, portrait recognition, people Face recognition, text recognition, etc.

In the pattern recognition technology, it can be divided into structured pattern recognition and decision-making pattern recognition (or statistical pattern recognition). In the text pattern recognition technology, the commonly used pattern recognition technology includes the eight-connected method and the projection method in the structured pattern recognition. , Template comparison method, edge statistical method, etc., and in decision-making pattern recognition technology, commonly used support vector machine (SVM), radial basis function (RBF) and neural network-like convolutional neural network (Convolutional Neural Network, referred to as CNN).

Structural pattern recognition mainly uses text structure for identification. Before machine learning technology is prevalent, structural pattern recognition is mainly used for text recognition. However, structural pattern recognition technology is mainly designed for text graphic structure, so it is used for different text graphics , All need to design different recognition methods, and each recognition needs to perform multiple image processing, numerical statistics and other steps, so each recognition takes a long time, coupled with the accuracy of noise image recognition, which The overall accuracy is also lower than that of machine learning and neural-like networks. Therefore, it has gradually been recognized by decision graphics Don't replace it.

In decision-making pattern recognition, the text image is pre-processed with various filter data pre-processing steps, and then trained using machine learning methods. Compared with constructing pattern recognition, decision-making pattern recognition has a wider range of applications and is less susceptible to confusion After the training is completed, the machine only needs to store the weight parameters after training and bring the new data points into the parameters for matrix operation to obtain the output result, and the identification takes less time.

Currently the most commonly used convolutional neural network for pattern recognition, its main structure includes a set of convolution layers and a set of flat layers, each containing multiple neurons, and the convolution layer mainly adjusts the weight of image processing. It is the adjustment of neuron weights. In the current technology, superimposing multiple convolutional layers and improving the convolutional layer does help to improve the overall accuracy, but the more multi-layer structure implies that the number of parameters in the network needs to be adjusted. , The required training time will increase, it will also lead to more computing performance and time for text and pattern recognition after training, and there is currently no theoretical basis for planning the number of layers of the neural network, making the structure of the neural network It tends to be complicated, so in some special fields, such as community safety or stolen car tracking, etc., the need to quickly identify the license plate text, the pattern recognition method established by an overly complicated neural network is not applicable, and needs to be Better plan.

In view of the existing pattern recognition technology, the accuracy of constructing pattern recognition is not good, and the method of decision-making pattern recognition overlapping multi-layer neurons will increase the technical defects of recognition time-consuming, the present invention proposes a smart text pattern recognition method, It has a streamlined computing system architecture, reduces training time, and can improve recognition accuracy and reduce the time required for recognition.

作為各神經元的激發函數，且亂數產生高 斯函數的中心點xc與標準差σ；輸入預處理後之文字區域影像的特徵向量至該神經網路，並計算輸出層連結權重與辨識誤差；建立新的神經元，以高斯函數作為新增神經元的激發函數，且亂數產生高斯函數的中心點與標準差；輸入預處理後文字區域影像的特徵向量，計算新增該神經元至網路隱藏層所致整體辨識誤差，若新誤差小於前次誤差，則保留該神經元，若否，則捨去該神經元；重複建立神經元、亂數產生其參數、計算輸出連結權重以及計算辨識誤差之遞迴步驟，直至達到該停止條件；輸入測試影像，進行文字圖形識別。 The technical means adopted to achieve the above purpose make the intelligent text pattern recognition method include: performing training image preprocessing to obtain a text area image and its corresponding feature vector; setting an initialization number and a stopping condition; providing a single The neural network of the hidden layer sets the number of neurons in the hidden layer according to the initialization number

As the excitation function of each neuron, and the random number generates the center point xc and standard deviation σ of the Gaussian function; input the feature vector of the pre-processed text area image to the neural network, and calculate the output layer connection weight and identification error; Create a new neuron, use the Gaussian function as the excitation function of the new neuron, and the random number generates the center point and standard deviation of the Gaussian function; input the feature vector of the pre-processed text area image and calculate the new neuron to the network The overall identification error caused by the hidden layer of the road. If the new error is less than the previous error, the neuron is retained. If not, the neuron is discarded; the neuron is repeatedly created, the parameters are generated by random numbers, and the output link weight and calculation are calculated. Recursive steps for identifying errors until the stop condition is reached; input test images for text and graphic recognition.

The present invention uses a single hidden layer neural network to achieve the nature of non-linear classification. Therefore, a single hidden layer neural network can be used for text and graphic recognition. The main difference from the traditional neural network is that the present invention uses The parameters of the node function of the hidden layer of the neural network, that is, the center point and standard deviation of the Gaussian function, are generated using random numbers and do not need to be adjusted during the training process. Secondly, the number of hidden layers of the neural network used in the visual recognition accuracy needs to be added by itself. The new neuron can ensure the effect of reducing the identification error, so that it can find sufficient accuracy under the premise of a simplified neural network structure. The network model, with this technology, can not only reduce the computational complexity of network training, but also reduce the time required for text and pattern recognition after training.

FIG. 1 is a flowchart of steps according to an embodiment of the invention.

FIG. 2 is a schematic diagram of a neural network according to an embodiment of FIG. 1.

FIG. 3 is a flowchart of steps in another embodiment of FIG. 1.

Fig. 4a is a statistical diagram of the convergence speed of the embodiment of Fig. 1.

FIG. 4b is a statistical diagram of the convergence speed of the embodiment of FIG. 3.

FIG. 5 is a flowchart of the image preprocessing steps of the embodiment of FIG. 1.

Figures 6a~6f: Schematic diagrams of image processing in each step of Figure 5.

FIG. 7 is a flowchart of image preprocessing steps in another embodiment of FIG. 1.

Figures 8a~8f: schematic diagrams of the image processing in each step of Figure 7.

作為各神經元的激發函數，且亂數產生高 斯函數的中心點xc與標準差σ；輸入預處理後之文字區域影像的特徵向量至該神經網路，並計算輸出層連結權重與辨識誤差；建立新的神經元，以高斯函數作為新增神經元的激發函數，且亂 數產生高斯函數的中心點與標準差；利用預處理後文字區域訓練影像計算該神經元之輸出連結權重，進而計算新增該神經元至網路隱藏層所致整體辨識誤差，若新誤差小於前次誤差，則保留該神經元，若否，則捨去該神經元；重複建立神經元、亂數產生其參數、計算輸出連結權重以及計算辨識誤差之遞迴步驟，直至達到該停止條件；輸入測試影像，進行文字圖形識別。 Please refer to FIG. 1, the intelligent text pattern recognition method of the present invention includes: performing image preprocessing to obtain a regional image of a text and its corresponding feature vector; setting an initialization number and a stop condition. In one embodiment, the initialization number can be Set to 0, the stop condition can be set to a neuron number, an error threshold or a maximum number of recursion; provide a single hidden layer neural network, set the number of hidden layer neurons according to the initialization number, and use Gaussian function

As the excitation function of each neuron, and the random number generates the center point xc and standard deviation σ of the Gaussian function; input the feature vector of the pre-processed text area image to the neural network, and calculate the output layer connection weight and identification error; Create a new neuron, use the Gaussian function as the excitation function of the new neuron, and the random number generates the center point and standard deviation of the Gaussian function; use the pre-processed text area training image to calculate the output link weight of the neuron, and then calculate Add the neuron to the overall identification error caused by the hidden layer of the network. If the new error is less than the previous error, the neuron will be retained. If not, the neuron will be discarded; the neuron will be created repeatedly and the random number will generate its parameters 3. Calculate the recursive steps of output link weights and calculation identification errors until the stop condition is reached; input test images for text and graphic recognition.

Please further refer to FIG. 2 to illustrate the above steps with an embodiment. In the above steps, when the error is calculated for the first time, since the initial number of neurons in the hidden layer is 0, it is equivalent to no hidden layer, and the output is calculated at this time ( The actual output) and the desired output (desired output) error, in one embodiment, the error can be calculated using the least square difference, and then add the neuron step to create a neuron with Gaussian function as the excitation function , And generate the center point and standard deviation of the Gaussian function with random numbers, and then calculate the error between the actual output and the desired output. If this error is less than the previous error, this neuron is helpful In order to improve the accuracy, add this neuron to the network and continue to create the next neuron. If not, it means that this neuron is not helpful to improve the accuracy. Abandon it and generate φ ₁ ~ in sequence by the above steps φ _n hidden layer neuron.

，以計算神經元激發量 數學式如下式、ct為新的權重、contri為是否新增進神經網路的基準：hi(X,X _c ,σ)=exp(-σ∥X-X _c ∥²) The above calculation steps to add the neuron to the overall recognition error caused by the hidden layer of the network can be obtained by Mean Squared Error (MSE)

, The mathematical formula for calculating the amount of neuron excitation is as follows, ct is the new weight, and contri is the benchmark for whether to add to the neural network: hi ( X , X _c , σ )=exp(- σ ∥ X - X _c ∥ ² )

When contri<bestcon, add neurons to the neural network and update the new weights. Please further refer to FIG. 3, the above step of adding neurons can be changed to create k neurons, and the random number generates the parameters of the Gaussian function of each neuron, that is, the center point and the standard deviation, and calculates the output link weight of each neuron separately And to identify the amount of change in error, select the neuron with the smallest change and the smallest value to add to the current network, and discard other neurons. Repeat the above steps of creating complex neurons and screening. Overall, this embodiment can speed up the convergence speed, as shown in FIGS. 4a and 4b. The convergence curve of this embodiment is shown in FIG. 4b, which has a faster convergence speed than the previous embodiment.

Please refer to FIG. 5 for further reference. The above-mentioned intelligent text and pattern recognition method of the present invention can be used for water meter digital recognition and license plate recognition. The following will take water meter digital recognition as an example. The image pre-processing steps include: image grayscale and binary Figure 6a and Figure 6b; remove the small pixel connection points of the image, as shown in Figure 6c; erosion and close image processing, as shown in Figure 6d; fill the image area, as shown in Figure 6e, this step can be Find the coordinates of the area that may be text; obtain a text area based on the specified coordinates; capture the corresponding text area image, as shown in Figure 6f; binarize the text area image; use the eight-connected method to find each character and extract the direction of the character Histogram of Oriented Gradient (HOG) is used as the feature vector, and input into the above neural network for training or text and graphics recognition.

Please refer to Figure 7 for further cooperation. The following is an example of license plate recognition. The image preprocessing steps include: Image binarization, as shown in FIG. 8a; image edge detection, in one embodiment, Sobel edge detection can be used, as shown in FIG. 8b; remove small image connection points, as shown in FIG. 8c; erosion and Closed image processing, as shown in Figure 8d, this step can find the coordinates of the area that may be text; get a text area; capture the corresponding text area image, as shown in Figure 8e; text area image binarization, as shown in Figure 8f As shown; after finding each character using the eight-connected method, extract the histogram of orientation gradient (Histogram of Oriented Gradient, HOG) of the character as a feature vector, and input it into the above neural network for training or text graphics recognition.

According to actual comparison, the accuracy of the water meter digital identification according to the present invention is 83%, the identification time is 0.3~0.5 seconds, and using the traditional template comparison method, the accuracy is 54%, and the identification time is 0.9 ~1 second; traditional edge statistics method, accuracy is 29%, identification time is 0.8~1 second; support vector machine, accuracy is 79%, identification time is 1.5~1.8 seconds; radiation basis function, The accuracy is only 20.8%, and the time required for identification is 0.1 to 0.3 seconds; the method of the present invention has higher precision and faster identification speed.

Taking license plate recognition as an example, the accuracy of the two embodiments of the present invention is 77.5% and 87.5%, and the recognition time is 0.4 and 0.3; the traditional template comparison method, the accuracy is 75%, and the recognition time is 3.5 seconds; traditional edge statistics method, accuracy is 25%, identification time is 2.5 seconds; support vector machine, accuracy is 67.5%, identification time is 0.9 seconds; radiation basis function, accuracy is 57.5%, The time required for identification is 1.6 seconds; the method of the present invention also has higher accuracy and faster identification speed.

In summary, the intelligent text and figure recognition method used in the present invention can be improved The accuracy of text and figure recognition, and the recognition system after training has a faster recognition time. In addition to improving the accuracy bottleneck of traditional text recognition, it can also be used in technical fields that require rapid recognition, such as license plate tracking, to overcome traditional text graphics Identified technical defects.

The above are only preferred embodiments of the present invention and are not intended to limit the scope of the claimed rights of the present invention. Any other equivalent changes or modifications made without departing from the spirit of the present invention should be included in the present invention Within the scope of patent application.

Claims (7)

A smart text pattern recognition method, including: performing training image preprocessing to obtain a regional image of a text and its corresponding feature vector; setting an initialization quantity and a stopping condition; providing a single hidden layer neural network, based on the initialization quantity Set the number of neurons in the hidden layer, use the Gaussian function as the excitation function of each neuron, and generate the center point and standard deviation of each Gaussian function in random numbers; input the feature vector of the pre-processed text area image to the neural network and calculate The weight and recognition error of the output layer connection; create a new neuron, use the Gaussian function as the excitation function of the new neuron, and the random number generates the center point and standard deviation of the Gaussian function; input the feature vector of the text area image after preprocessing, Calculate the change in recognition error caused by adding the neuron to the current neural network. If the change is negative, add the neuron to the current network. If the change is positive, discard the neuron ; Repeat the above recursive steps of creating new neurons, generating parameters of random numbers, calculating output link weights, calculating recognition error changes, and choosing neurons until the stop condition is reached; input test images for text and graphic recognition. The intelligent text pattern recognition method as described in claim 1, which includes: image binarization; removal of under-pixel connection points of the image; erosion and closed image processing; filling of the image area; obtaining a text area according to the specified coordinates; retrieving the corresponding text Area image and binarization of text area image; use eight connected methods to find each character; and extract feature vector. The intelligent text pattern recognition method as described in claim 1, which includes: image binarization; image edge detection; removing small pixel connection points of the image; erosion and closed image processing; obtaining a text area; capturing the corresponding text area image ; Binarize the image of the text area; use the eight-connected method to find each character; and extract feature vectors. The intelligent text and pattern recognition method as described in any one of claims 1 to 3, further includes a recursive step of adding a new neuron: creating a complex neuron, and generating the center point and standard deviation of each neuron in random numbers, Calculate the output link weight and recognition error change of each neuron separately, select the neuron with negative change and minimum value to add to the current neural network, and discard other neurons. According to the intelligent text pattern recognition method described in claim 1, the stop condition is a number of neurons. According to the intelligent text pattern recognition method described in claim 1, the stop condition is an error threshold. According to the intelligent text and figure recognition method described in claim 1, the stop condition is a maximum number of recursions.

Images