Design and implementation of a marking system based on image recognition

The final examination of college students is an important task in the management of ordinary colleges and universities. It is based on the national requirements for the all-round development of college students' moral, intellectual and physical qualities, and in accordance with unified principles, methods and procedures, to conduct stage and whole-process quality assessment, recording, evaluation and processing of students' learning and behavior performance. At present, English exams or computer basic exams with more objective questions have realized automated marking, without the intervention of teachers, which can not only speed up the speed of test paper feedback, but also ensure fairness and justice. The traditional marking system uses the photoelectric conversion principle of the optical mark reader to collect the content filled in on the answer sheet at high speed, and then enter the computer for processing [1]. However, it has too high requirements on the paper and printing quality of the answer sheet, such as the overprint error and shear error must be controlled within 0.1 mm and 0.2 mm respectively; the requirements for users to write are too high and there are too many restrictions, such as the need to use a special pencil to fill in the rectangular strips, and the filling outside the rectangular strips must not be filled, and the entire answer sheet must be filled with the same depth; the mechanical transmission mechanism is complex, the service life is short, the maintenance is large, and the equipment consistency is poor [2].
This paper proposes a marking system based on image recognition, which allows users to use any color of ballpoint pen, fountain pen or pencil to write handwritten symbols such as "√", "╳", "○" and so on on fixed positions on ordinary paper to complete the test. The marking system uses image recognition technology to automatically identify the information on the answer sheet and realize automatic marking. The difference between this system and the traditional optical mark reader marking system is that this marking system has no special requirements for the answer sheet paper and the symbols filled in. No special answer sheet is required, and no special writing requirements are required [3].
1 System implementation process and module composition
(1) Design and define the answer sheet, and limit the handwritten symbols that users can fill in to a certain rectangular area;
(2) Collect the answer sheet image through the CCD imaging device, and after scanning, sampling, quantization and other processes, save the answer sheet image in JPG format in the host memory and hard disk;
(3) Preprocess the answer sheet image, including adhesion character segmentation, grayscale transformation, binarization, image skew correction, smoothing and refinement, so as to eliminate the interference factors caused by the inherent paper problems, irregular writing, and the positioning accuracy of the answer sheet transmission mechanism, and prepare for the subsequent recognition work [4];
(4) Recognize the answer sheet information. The image is extracted through image analysis, and after comprehensive feature extraction, combined with the logical structure and geometric structure of the answer sheet table, the handwritten symbols such as "√", "╳", "○" filled in the rectangular block position are accurately identified to obtain the answer information. The quality of the answer sheet recognition algorithm determines the performance of the entire system (such as recognition accuracy, reliability, etc.), and is the most important part of the entire software system;
(5) Process, organize, analyze and count the recognized answer information, and score the answer sheet information in combination with the software settings.
In summary, the system implementation process is shown in Figure 1.

In order to facilitate students and teachers to query and analyze the marking results, this marking system also sets up a user login module, a student information management module, a course basic information management module, a test paper analysis module, a score query module, an answer sheet image query module, a report printing module and a make-up exam management module. The test paper analysis module analyzes and counts various test paper data, including the highest score, the lowest score, the average, the pass rate variance, the standard deviation, etc., to complete the test paper analysis function; the answer sheet image query module allows students to query the answer sheet images stored in the hard disk by name; the make-up exam management module counts the information of students who fail, miss or cheat, and generates a make-up exam form; the report printing module completes the printing function of answer sheet images, score reports, test paper analysis reports and make-up exam forms. The system module composition is shown in Figure 2.

2 Main technologies of the system
2.1 Answer sheet design
The answer sheet is designed in the form of a table, which consists of basic information of the test paper answer sheet, basic information of the examinee, instructions for filling out the answer sheet and a list of question options. The examinee uses handwritten symbols such as "√", "╳", "○" to select the answer in the corresponding rectangular box. An example of the answer sheet is shown in Figure 3.

The answer sheet layout has geometric structure and logical structure like ordinary tables. The geometric structure reflects the position and size of the information area filled in, and the logical structure represents the actual meaning of the information filled in the answer sheet and the corresponding relationship between the filled information and the filled items.
In the traditional optical mark reader grading system, the geometric structure of the answer sheet is mostly described by positioning mark blocks. This description method has many disadvantages, such as wasting layout space, lack of flexibility, complex layout design, high requirements for printing quality, inconvenient modification, etc., and the positioning mark blocks do not look beautiful. In the answer sheet layout of this system, the filled information can be regarded as a number of non-intersecting rectangular blocks in terms of geometric structure. They constitute the smallest unit of the answer sheet layout [5]. A two-dimensional coordinate system can be established with the border line of the answer sheet, and the position and size of the rectangular block can be described with the coordinates of the diagonal vertices of the rectangle to complete the geometric structure description of the answer sheet. This method is concise, flexible and easy to identify.
The logical structure description of the answer sheet is to define the attributes of the rectangular block. The attributes of the rectangular block include the selection or non-selection of the filled item represented by "√", "╳" and "○".
This system uses document structure description language to describe both geometric structure and logical structure. Assume that an answer sheet contains n rectangular blocks B1, B2,..., Bn filled with information. There is a geometric position relationship between the rectangular blocks, with upper and lower structures and left and right structures. The logical order is generally from top to bottom and from left to right. This order is also used when describing the document structure. The document structure description language DDL is expressed as follows:

where i is the serial number of the answer sheet; n is the total number of rectangular blocks; xi is the horizontal position of the rectangular block; yi is the vertical position of the rectangular block; li is the length of the rectangular block; wi is the width of the rectangular block; attri represents the attribute of the rectangular block. When attri is 0, it means that the content of the rectangular block is the character to be recognized. When attri is 1, the rectangular block is saved as an image.
The answer sheet is designed in a common table format. On the one hand, it is similar to the answer sheet used in general exams, which conforms to people's usage habits and writing habits; on the other hand, its geometric structure and logical structure can be described using document description language, so as to better extract and recognize the character features in the rectangular frame, identify the characters in the rectangular frame, compare them with the standard answers, and judge the objective questions of the examinees.
2.2 Answer sheet information recognition
This system uses the statistical decision method in the Chinese character recognition method to recognize handwritten characters, as shown in Figure 4. First, the character features are extracted, and the characters are classified and identified. After two stages of training and recognition, the characters are finally recognized[6]. In the training stage, the handwritten symbols that people are used to in daily life are collected as samples, and a sample library is established after screening and classification, so as to classify and recognize the characters to be recognized. In the recognition stage, the features of the symbol to be recognized are compared with the standard sample features established in the training stage, and the maximum similarity is calculated to determine the category to which the handwritten symbol belongs[7].

2.2.1 Character feature extraction
Due to nervousness and personal reasons, students often write various symbols such as √, ╳, ○, etc. in the process of answering questions. In order to accurately recognize these handwritten symbols, it is necessary to extract the character structure features and extract the structural features that best reflect the characteristics of the character and the differences between characters. The features extracted by this system are a combination of the following features.
(1) Point feature
Point feature is an important structural feature, which refers to the endpoints in the character strokes. The endpoint reflects the starting and ending information of the strokes in the character, and the number of points connected to the point is 1.
(2) Stroke density feature
The stroke density feature is to obtain the stroke density function d(x) in the horizontal direction and the stroke density function d(y) in the vertical direction of the symbol, and then merge the same terms [1]. As shown in Figure 5, the horizontal stroke density function d(x) and the vertical stroke density function d(y) of the symbol "○" are: d(x)=(1,...,1,2,...,2,1,...,1), d(y)=(1,...,1,2,...,2,1,...,1), then the merged stroke density can be expressed as d(x)=d(y)=(1,2,1).

(3) Structural features based on the chain code method
There are 8 possible directions of the pixels connected to the starting point of the curve: k×45° (k=0,1,…,7), as shown in Figure 6. If the direction of the line between two pixels is k×45°, "k" is used as the code of this line, and a curve can be approximately represented by the following formula:
An=a1a2…an,ai∈{0,1,2,…,7}, i=1,2,…,n

(4) Hole feature
In a binary image, the background pixels 0 (set of pixels) surrounded by the target pixel 1 are called holes. In the process of forming the chain code of the skeleton line of a character, if the next point found is the search starting point of the skeleton line, and the length of the skeleton chain code formed exceeds a certain threshold, it is considered that a hole has been found [5].
(5) Horizontal and vertical intersection feature
Scanning characters horizontally or vertically, the number of times the pixels in a row or column change from white to black is the horizontal or vertical intersection feature of the row or column. This system applies 7 lines of unequal distance in the horizontal and vertical directions to the characters and calculates the number of intersections with the characters in the horizontal and vertical directions, as shown in Figure 7.

2.2.2 Establishment of symbol model library
The training phase of answer sheet information recognition requires the establishment of a symbol model library in order to classify and identify the handwritten symbols to be recognized. The quality of the symbol model library directly affects the application of the classifier, thereby affecting the handwritten symbol recognition effect [5].
Due to the diversity of handwritten symbols, it is necessary to select multiple representative samples of a certain class of handwritten symbols to construct standard samples. This system uses the mean of the feature vector of handwritten character samples to describe the class target. There are n symbol classes, each symbol class has a training samples, each sample has b symbol features, and the features of the samples in each symbol class are recorded as fkj, k is the sample feature number, j is the sample number of each handwritten symbol, then the mean of the i-th target class feature is P(i), that is:

each time Pik is calculated, k is a fixed value greater than or equal to 1 and less than or equal to b. Pik is the mean of the k-th feature values ​​corresponding to each sample in the a samples in the i-th target class.
2.2.3 Handwritten symbol recognition
The identification of the rectangular box information in the answer sheet includes two aspects: one is to identify whether there are characters in the rectangular box, and the other is to identify the specific type of characters. Among them, it is relatively simple to identify whether there are characters. As long as the mean square error of the rectangular box image after contrast enhancement is compared with the known rectangular box filled with information, it can be identified whether there are characters, because the mean square error of the blank rectangular box and the written rectangular box is very different. The following mainly introduces how to identify the specific characters in the rectangular box.
Handwritten symbol recognition is to determine which handwritten symbol a certain graphic dot matrix represents based on a certain discriminant function after extracting the feature vector of the symbol.
The discriminant function can be simply defined as follows: Consider P1, P2, ..., Pm symbol categories. If each category has a standard sample, there are m standard samples in total, which are represented as k1, k2, ..., km respectively. The "similarity" between any symbol feature vector X and the i-th (i=1, 2, ..., m) standard sample is Ri. Calculate the "similarity" between the feature vector X of the symbol to be recognized and each type of standard sample [7], and classify X into the category with the greatest "similarity" to it, that is, for all j not equal to i, if Di>Dj, then X belongs to the Pi type symbol.
The system uses a template matching algorithm based on the nearest neighbor classifier to recognize handwritten symbols.
First, define the character feature vector. After the previous feature extraction analysis, the feature vector is a 16-dimensional vector, X={x1,x2,..,x16}, which is specifically defined as:
x1: number of holes;
x2: number of endpoints;
x3~x9: number of intersections between the 7 horizontal lines and the character;
x10~x16: number of intersections between the 7 vertical lines and the character.
By measuring the proximity between the character to be recognized and the sample characters in the sample library, a criterion for the nearest classification is established. In the nearest neighbor classification, similarity is often used. As shown in Figure 8, after extracting the feature vector of the character to be recognized and establishing the character library, the similarity R(X, G) between the feature vector of the character to be recognized and the i-th sample in the sample library is calculated. R(X, G) is defined as follows:

In the formula, xi is the i-th component of the feature vector of the symbol to be identified, gik is the k-th component of the i-th standard sample in the sample library, and m is the number of sample categories. The numerator is the inner product between vectors X and G, and the denominator is the modulus of vectors X and G respectively. α is the angle between vectors X and G in m-dimensional space. Obviously, when vectors X and G are exactly the same, their angle is 0, R(X,G)=1, and their distance D(X,G)=0, that is, the similarity is maximum. Find the maximum Rr(X,G). If Rr(X,G)≥the given threshold, the sample category closest to the character to be identified can be found. Otherwise, manual intervention and modification of the sample library are performed[7].
3 Experimental results and analysis
The experiment uses a CCD camera to collect the answer sheet image. After image preprocessing, several feature extractions, information recognition and other processes, it is determined whether there are characters in the rectangular frame and what characters they are. Finally, the answer sheet information is analyzed and counted. The experiment used 100 test papers as samples and tested 20 test papers. The results showed that the handwritten symbols that were misrecognized were mainly "√" and "╳" because they are similar in structure, while the recognition rate of the symbol "○" reached 100%.
This system applies image preprocessing, character feature extraction and image recognition technologies to the development of the marking system, realizes the automation of marking, speeds up the assessment of grades, and improves the teaching management environment. Compared with the traditional marking system based on optical mark readers, this system uses image recognition technology to realize the automation of marking. It does not require special answer sheets. Candidates can also use various handwritten symbols such as "√", "╳", "○" to answer questions at will, and do not need to use designated 2B pencils to fill in rectangular blocks, which is more in line with people's habits.
References
[1] Wang Hu. Research on mark reader and election counting system based on image recognition [D]. Hefei: Anhui University, 2006.
[2] Zhang Ting. Research and application of optical mark reader based on image recognition technology [D]. Hefei: Anhui University, 2007.
[3] Wu Yuanjun, Zhang Ting, Lei Jingpeng. Application of an improved OMR technology in standardized examinations [J]. Computer Education, 2007(13):250-272.
[4] Ding Huidong. Research on offline handwritten Chinese character recognition [D]. Changchun: Northeast Normal University, 2006.
[5] Pang Donghu, Jin Weijie. English character feature extraction system [J]. Computer Simulation, 2007, 24(12):208-210.
[6] Yang Ling, Mao Yifang, Wu Tianai. Research on offline handwritten Chinese character recognition based on multi-feature multi-classifier [J]. Computer and Network, 2008(01):217-217.
[7] Qin Sheng, Liu Xiaoming. Implementation of image-based OMR technology [J]. Application of Electronic Technology, 2003, 29(10):17-19.
[8] Weng Gongping. Design and implementation of OMR principle of OMR reader [J]. Industrial Control Computer, 2010, 23(04):61-62.