CN106203417A - A kind of adhesion character alienable RMB crown word number identification method - Google Patents

Abstract

The invention discloses a method for recognizing RMB serial numbers with divisible sticky characters. Through horizontal fine-segmentation, it is judged that those whose width is greater than the maximum width D _max of a single character image are abnormal character graphics, that is, sticky character images. Vertically project the curve, find the minimum value point in the set sub-interval as the character segmentation point, and then perform vertical segmentation. If the character image of the segmentation process is greater than the minimum character image width D _min , keep the character image, so that the characters will be glued Segmentation can improve the accuracy rate of character segmentation in the preprocessing process and reduce the misrecognition rate of prefix numbers.

Description

A Recognition Method of Renminbi Serial Numbers with Segmentable Sticky Characters

technical field

The invention belongs to the field of pattern recognition and optical character recognition (OCR) technology, and in particular, relates to a method for recognizing a Renminbi serial number with detachable sticky characters, which can be applied to the Renminbi counting, checking, sorting and sorting machine system. Serial number identification.

Background technique

In 2009, the National Standard Revision and Drafting Working Group of "General Technical Conditions for RMB Discriminators" published a new version of the banknote detector standard GB 16999-2010, requiring banknote detectors to increase the banknote serial number recognition function, so the banknote serial number is embedded in the banknote counting machine Recognition function has become the latest research topic.

Affected by the performance of the machine or the defacement of the banknote itself, when the money detector shoots or scans the serial number on the banknote, there will be some interference, which will cause the characters to stick together. Unable to meet the requirement of the serial number misrecognition rate of less than 0.03% stipulated in the "General Technical Conditions for Renminbi Verifiers", where the serial number misrecognition rate refers to the number of banknotes with serial number misrecognition and the actual number of banknotes recognized ratio.

Contents of the invention

The object of the present invention is to overcome the deficiencies of the prior art, and propose a method for recognizing the serial number of RMB in which the glued characters can be divided, so as to effectively segment the glued characters and reduce the misrecognition rate of the serial number.

In order to realize the above-mentioned purpose of the invention, the renminbi serial number recognition method that the sticky characters of the present invention can be divided is characterized in that, comprises the following steps:

(1), character segmentation

1.1), character horizontal rough segmentation

The input is the serial number grayscale image collected by the image sensor, and binarization is performed, and then vertical projection analysis is performed, and vertical segmentation is performed at the horizontal position projection value of 0 to obtain a rough segmentation image of the character;

1.2), character horizontal subdivision

1.2.1), detect the width of the character image obtained by coarse segmentation: if the width is greater than the maximum width D _max of a single character image, the character image is regarded as an abnormal character image, and step 1.2.2) is turned; otherwise, the character image is regarded as For a single character image, continue to detect the next character image until the last character image is detected and character segmentation ends;

1.2.2), count the vertical projection curve of the abnormal character image, find the minimum value point of the vertical projection curve in the subinterval [x _l +s, x _r -s] of the character image horizontal interval containing the horizontal center coordinates, and This point is used as the segmentation point of glued characters, where x _l and x _r are the left and right coordinates in the horizontal direction of the character image obtained by rough segmentation, and s is the subinterval setting parameter, which is set according to the specific situation. The larger the value, the smaller the subinterval ;

1.2.3), with the sticky character segmentation point that step 1.2.2) obtains, vertically segment the abnormal character image, and judge the character width of the left and right character images after the segmentation, if the character image width is less than the minimum character image width D _min , discard the character image, otherwise keep the character image, then return to step 1.2), and continue to detect the next character image;

1.3), character vertical division

The character image obtained in step 1.2) is analyzed horizontally, and the character image is removed when the projection value is 0, and the upper and lower blank parts of the character image are removed;

(2), character recognition

After the single character image obtained by the character segmentation in step (1) is normalized by the size of the character image, it is sequentially sent to the trained convolutional neural network according to the order of horizontal segmentation to identify the corresponding character;

(3), all recognized characters are combined according to the order of segmentation to obtain the serial number (code).

The purpose of the present invention is achieved like this.

The method for recognizing the RMB serial number with the divisible sticky characters of the present invention, through horizontal subdivision, that is: by determining the width greater than the maximum width D _max of a single character image, it is an abnormal character figure, that is, a sticky character image, and through the vertical projection curve of the abnormal character image, Find the minimum value point in the set sub-interval as the character segmentation point, and then perform vertical segmentation. If the character image of the segmentation process is greater than the minimum character image width D _min , then keep the character image, so that the glued characters are separated. Improve the correct rate of character segmentation in the preprocessing process, and reduce the misrecognition rate of prefix numbers.

Description of drawings

Fig. 1 is a kind of specific implementation flow chart of the Renminbi serial number identification method that sticking character can be divided of the present invention;

Fig. 2 is a specific example figure of a serial number grayscale image with glued characters;

Fig. 3 is the vertical projection curve diagram of serial number grayscale image shown in Fig. 2;

Fig. 4 is the character image that the gray scale image of serial number shown in Fig. 2 carries out horizontal subdivision and obtains;

FIG. 5 is a structural diagram of a specific implementation of the convolutional neural network shown in FIG. 1 .

detailed description

Specific embodiments of the present invention will be described below in conjunction with the accompanying drawings, so that those skilled in the art can better understand the present invention. It should be noted that in the following description, when detailed descriptions of known functions and designs may dilute the main content of the present invention, these descriptions will be omitted here.

Fig. 1 is a flow chart of a specific embodiment of the recognition method for RMB serial numbers with detachable sticky characters of the present invention.

In the present embodiment, as shown in Fig. 1, the method for identifying the RMB serial number with detachable sticky characters of the present invention comprises the following steps:

1. Character segmentation

S101, the input is the 8-bit grayscale serial number image (input image) collected by the image sensor, and binarization processing is performed to obtain a binarization image; in this embodiment, after the binarization processing, a negation operation is performed , that is, 0 becomes 255, and 255 becomes 0, so that the characters become white and the background becomes black;

S102. Perform vertical projection analysis on the binarized image. At the place where the horizontal position projection value is 0, that is, there is no pixel point number of 255 in the vertical direction, the number of pixels is 0, and the vertical segmentation is performed to obtain a rough segmentation image of the character;

In this embodiment, steps S101 and S102 are horizontal rough segmentation of characters.

S103, detection and segmentation of sticky characters, i.e. character horizontal subdivision;

In this embodiment, as shown in Figure 2, during the image acquisition process of serial number images, there may be adhesion between characters due to noise and contamination. If the traditional vertical projection segmentation method, that is, the method of steps S101 and S102, is used to directly segment , U and F will be regarded as a character for segmentation, obviously the segmentation result is wrong, which will lead to unrecognized or incorrectly recognized. At the same time, the last character 0 is also regarded as an abnormal character because the character width is greater than D _max due to noise.

In the present invention, the following method is adopted to carry out character level subdivision, namely:

S1031, detect the width of the character image obtained by coarse segmentation: if the width is greater than the maximum width D _max of a single character image, the character image is regarded as an abnormal character image, and step S1032 is turned; otherwise, the character image is regarded as a single character image, and continues Detect the next character image until the last character image is detected and character segmentation ends;

S1032, count the vertical projection curve of the abnormal character image, and find the minimum value point of the vertical projection curve in the subinterval [x _l +s, x _r -s] of the character image horizontal interval including the horizontal center coordinate, that is, the vertical direction of the point is The pixel number of 255 is the least, and this point is used as the sticky character segmentation point, wherein, x _l and x _r are the left and right coordinates of the horizontal direction of the character image obtained by rough segmentation, and s is the sub-interval setting parameter, which is set according to the specific situation , the larger its value, the smaller the sub-interval; in this embodiment, the size of the character image is represented by pixels, and the parameter s=3, that is, 3 pixels

S1033. Vertically segment the abnormal character image with the sticky character segmentation point obtained in step S1032, and determine the character width of the segmented left and right character images. If the character image width is smaller than the minimum character image width D _min , discard the character. image, otherwise keep the character image, then return to step 1.2), and continue to detect the next character image;

After adopting the above character horizontal fine segmentation method, the segmentation coordinates of the two cohesive character images marked with black thick lines in Figure 3 are obtained. After the first abnormal character image is segmented, two characters are obtained, and the character width is greater than the minimum character width D _min , so They are all preserved, and the second abnormal character is split to obtain two characters. The width of the area on the left side of the sticky split point is greater than D _min , the character is reserved, and the width of the area on the right is smaller than D _min , so it is discarded. Finally, the ten characters correctly segmented are obtained, as shown in Fig. 4 .

S104, character vertical division

Perform horizontal projection analysis on the character image obtained in step S103, and remove the upper and lower blank parts of the character image when the projection value is 0.

2. Character recognition

S201. Normalize the individual character images obtained by character segmentation in Step 1 through character image size normalization. In this example, the normalized size is 14x14.

S202. The character image after the normalized size is sequentially sent to the trained convolutional neural network for recognition according to the order of horizontal segmentation, and corresponding characters are obtained.

Compared with the recognition method of traditional template matching and manual extraction of character features, convolutional neural network has the characteristics of high recognition rate and automatic extraction of various deep-level features of character images. In this embodiment, in order to reduce the consumption of resources, increase the recognition speed, and take into account the correct rate of recognized serial numbers (the misrecognition rate of serial numbers), the structure of the traditional convolutional neural network is simplified.

In this embodiment, as shown in Figure 5, the convolutional neural network structure is based on the LeNet5 architecture, and the simplified operation is:

1. Does not include the input layer. The convolutional neural network of the LeNet5 architecture has 7 layers in total. In this embodiment, the convolutional neural network has 5 layers in total, that is, the C5 layer and the F6 layer are reduced, and the C5 layer is composed of 120 feature maps. The convolutional layer of F6 is a fully connected layer. Removing these two layers can reduce a large number of training parameters and save computing resources and memory resources. The S4 layer feature map is directly connected to the output layer in a fully connected manner without an intermediate hidden layer;

2. Reduce the number of feature maps of each layer. The feature map of the convolutional neural network C3 layer of the LeNet5 architecture contains 16 feature maps, while the feature map of the C3 layer of the present invention only contains 6 feature maps;

3. The feature map of the S2 layer is fully connected to the C3 feature map, instead of the partial connection method in the convolutional neural network of the LeNet5 architecture. Although this method increases some connection parameters, it reduces the difficulty of FPGA implementation.

4. In this embodiment, the downsampling layer adopts average pooling, no training parameters are required, and only one bias parameter needs to be trained for each layer.

Specifically, in this embodiment, the convolutional neural network is identified as:

S2021. Convolving the 14x14 single character image normalized in size to obtain a C1 layer feature map containing 6 feature maps, each of which has a size of 12x12;

S2022. Down-sampling the C1 layer feature map to obtain an S2 layer feature map containing 6 feature maps, each of which has a size of 6x6;

S2023. Convolving the S2 layer feature map to obtain a C3 layer feature map containing 6 feature maps, each feature map has a size of 4x4;

S2024. Down-sampling the feature map of layer C3 to obtain feature maps of layer S4 comprising 6 feature maps, each of which has a size of 2x2;

S2025, the feature map of layer S4 is directly connected to the output layer in a fully connected manner, and the corresponding characters are output.

3. Combine all the recognized characters according to the order of segmentation to obtain the serial number (code). Crown number.

In this embodiment, the convolutional neural network is first constructed and trained on the PC host computer software platform, and then the trained convolutional neural network is downloaded to the FPGA chip of the cash detector to recognize the character images obtained by collecting and segmenting.

Although the illustrative specific embodiments of the present invention have been described above, so that those skilled in the art can understand the present invention, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those of ordinary skill in the art, As long as various changes are within the spirit and scope of the present invention defined and determined by the appended claims, these changes are obvious, and all inventions and creations using the concept of the present invention are included in the protection list.

Claims (2)

1. A method for recognizing the renminbi serial number that can be divided by sticking characters, is characterized in that, comprises the following steps: (1), character segmentation 1.1), character horizontal rough segmentation The input is the serial number grayscale image collected by the image sensor, and binarization is performed, and then vertical projection analysis is performed, and vertical segmentation is performed at the horizontal position projection value of 0 to obtain a rough segmentation image of the character; 1.2), character horizontal subdivision 1.2.1), detect the width of the character image obtained by coarse segmentation: if the width is greater than the maximum width D _max of a single character image, the character image is regarded as an abnormal character image, and step 1.2.2) is turned; otherwise, the character image is regarded as For a single character image, continue to detect the next character image until the last character image is detected and character segmentation ends; 1.2.2), count the vertical projection curve of the abnormal character image, find the minimum value point of the vertical projection curve in the subinterval [x _l +s, x _r -s] of the character image horizontal interval containing the horizontal center coordinates, and This point is used as the segmentation point of glued characters, where x _l and x _r are the left and right coordinates in the horizontal direction of the character image obtained by rough segmentation, and s is the subinterval setting parameter, which is set according to the specific situation. The larger the value, the smaller the subinterval ; 1.2.3), with the sticky character segmentation point that step 1.2.2) obtains, vertically segment the abnormal character image, and judge the character width of the left and right character images after the segmentation, if the character image width is less than the minimum character image width D _min , discard the character image, otherwise keep the character image, then return to step 1.2), and continue to detect the next character image; 1.3), character vertical division The character image obtained in step 1.2) is analyzed horizontally, and the character image is removed when the projection value is 0, and the upper and lower blank parts of the character image are removed; (2), character recognition After the single character image obtained by the character segmentation in step (1) is normalized by the size of the character image, it is sequentially sent to the trained convolutional neural network according to the order of horizontal segmentation to identify the corresponding character; (3), all recognized characters are combined according to the order of segmentation to obtain the serial number (code). 2. identification method according to claim 1, is characterized in that, in step (2), described convolutional neural network is identified as: 2.1), the size normalized 14x14 single character image is convolved to obtain a C1 layer feature map comprising 6 feature maps, each of which has a size of 12x12; 2.2), the C1 layer feature map is down-sampled to obtain the S2 layer feature map containing 6 feature maps, and the size of each feature map is 6x6; 2.3) Convolving the S2 layer feature map to obtain a C3 layer feature map containing 6 feature maps, the size of each feature map is 4x4; 2.4), the C3 layer feature map is down-sampled, and the S4 layer feature map comprising 6 feature maps is obtained, and the size of each feature map is 2x2; 2.5), the S4 layer feature map is directly connected to the output layer in a fully connected manner, and the corresponding characters are output.