Graduate employment questionnaire scanning identification method and system

Abstract

According to the scanning identification method and system of the graduate employment questionnaire, paper test paper is converted into image data by using an image scanning technology, the filled test paper is further analyzed by using a deep learning method of machine vision, the validity of the test paper is judged, the test paper which does not meet the standard is rapidly screened out, and the accuracy of the test paper is improved. A lot of time is saved for later survey and analysis, and the efficiency of questionnaire survey and analysis is improved.

Description

Technical Field

[0001] The present invention belongs to the field of image processing in automatic pattern recognition technology, and in particular, relates to a scanning and recognition system for graduate employment questionnaires. Background Art

[0002] As students' educational awareness improves, the number of graduates in my country increases dramatically each year, and employers are approaching saturation. This situation has led to a more severe employment situation. When graduates and employers make a two-way choice, "errors" are prone to occur in the employment direction of graduates, talent market demand, actual admission situation, etc. In this case, it is necessary to conduct an effective analysis based on the employment situation of graduates and strictly control the employment situation.

[0003] Among them, the collection of employment information is a key step, which involves collecting information about graduates' employment status through questionnaires. For anonymous, non-directional, and large-scale employment questionnaire surveys, the validity of the questionnaire is usually a problem. The respondents may fill in the questionnaire at will, making the questionnaire basically worthless. Moreover, such situations often occur, making it a time-consuming and labor-intensive task to screen out these questionnaires during later statistics. Therefore, how to quickly and efficiently screen effective questionnaires is an important step in the questionnaire survey process.

[0004] There are a lot of technologies in the prior art for reading effective information from questionnaires and conducting statistical analysis on questionnaires, but there are few technologies for screening the effectiveness of questionnaires, and even fewer technologies for accurate screening based on this. Summary of the invention

[0005] The present invention proposes a questionnaire scanning, recognition and screening method and system based on visual analysis, which uses image scanning technology to convert paper test papers into image data, and further uses the deep learning method of machine vision to analyze the filled-in test papers, judge the validity of the test papers, and quickly screen out test papers that do not meet the standards, saving a lot of time for subsequent investigation and analysis and improving the efficiency of questionnaire investigation and analysis.

[0006] A scanning and recognition method for graduate employment questionnaires.

[0007] Collect a scanned image of a blank questionnaire, denoted as B; collect a scanned image of the answered questionnaire to be identified, denoted as S;

[0008] The B and S images are filtered by selecting different window sizes σ; where σ is selected as σ 0 , 2σ 0 , 4σ 0 ; Three groups of B image filtering results can be obtained respectively And three sets of S image filtering results

[0009] Compute the matrix: We get 8 matrices, among which:

[0010]

[0011] The above three groups of S image filtering results are filtered again using the above 8 matrices to obtain three groups of filtering results.

[0012] calculate: in

[0013] Will Divide into several sub-areas of equal size: D s (x, y); and feed it into the first layer of the neural network as input:

[0014] in, represents the maximum value of the corresponding position in the equal-sized sub-image, α 1c is the linear convolution kernel, β 0c is the corresponding bias, c represents the number of independent convolution kernels, and λ is the activation function;

[0015] Neural network is used to judge whether the questionnaire meets the standards.

[0016] c=64.

[0017] The above filtering is Gaussian filtering.

[0018] When training a neural network, the cost function is

[0019]

[0020] According to the above model output recognition results, when the model output ∈>0, it means that the questionnaire corresponding to the input meets the standard, otherwise, it means that it does not meet the standard.

[0021] An automatic scanning and recognition system for graduate employment questionnaires comprises a camera and a processor, wherein the processor is used to execute the above method.

[0022] The camera is used to capture questionnaire images.

[0023] The camera and the processor are connected via network communication.

[0024] A scanning and processing system for graduate employment questionnaires includes a standard determination module, which is used to execute the above method.

[0025] This module is embedded in the existing questionnaire processing system.

[0026] This module is connected with the questionnaire processing and statistical analysis module.

[0027] The invention and technical effects of the present invention are as follows:

[0028] 1. The invention of the present invention is to use image scanning technology to convert paper test papers into image data, and further use the deep learning method of machine vision to analyze the filled test papers, determine the validity of the test papers, and quickly screen out test papers that do not meet the standards. For anonymous, non-directional, and large-scale employment questionnaire surveys, the validity screening of the questionnaires takes a lot of time. The method of the present invention can automatically filter out test papers that do not meet the standards, with a high recognition accuracy rate, which greatly reduces the workload of analysts, thereby saving a lot of time for later investigation and analysis, and improving the efficiency of questionnaire survey analysis.

[0029] 2. The invention of the present invention is to solve the problem of color difference in scanned data, reduce the impact caused by deviation of scanning equipment, and improve the recognition accuracy.

[0030] 3. The invention of the present invention is to propose a method for extracting the apparent features of the questionnaire. Some apparent features, such as wrinkles, may reflect the respondent's attention to the questionnaire and have a certain correlation with whether the test paper meets the standards. Through this method, the wrinkles are further strengthened, and thus added to the recognition data source as a supplementary feature, so that the discriminant model can more effectively discriminate the standard degree of the test paper.

[0031] 4. The inventive point of the present invention is to propose a neural network model based on the maximum convolution layer to recognize input data and output standardized recognition results. By improving the network structure and calculation method, the robustness of recognition is improved and the recognition accuracy is improved. DETAILED DESCRIPTION

[0032] Step 1 Scan the paper questionnaire, convert it into image data, and perform necessary processing during the conversion process to ensure the reliability of the image data.

[0033] Collect a scanned image of a blank questionnaire, denoted as B.

[0034] The scanned image of the answered questionnaire to be identified is denoted by S.

[0035] Since the scanning device cannot guarantee uniform brightness, contrast, etc. for each scan, the overall color of different scanned images will deviate. In order to reduce the impact of the scanning device deviation, an image filtering method is proposed as follows.

[0036] Take the scanned image of a blank questionnaire as an example, B(x, y) represents a pixel in the image, and its position is determined by x and y. Filtering is performed on the image as follows:

[0037]

[0038] Where exp represents the natural exponential function, σ is a parameter related to the filter window size, and u and v are offsets relative to the center of the filter window.

[0039] According to the above formula, different filtering results can be obtained by selecting different window sizes σ. For example, if σ = σ 0 ,2σ 0 , 4σ 0 , three sets of filtering results can be obtained

[0040] Similarly, for a scanned image S of a questionnaire answer, three sets of filtering results can be obtained:

[0041] For the questionnaire, some surface features such as wrinkles may reflect the respondent's attention to the questionnaire and have a certain correlation with whether the test paper meets the standards, so the scanned image of the questionnaire is further processed as follows.

[0042] Let the matrix

[0043]

[0044] is an n×n matrix, n≥3 and n is an odd number, R 1 (a, b) is an element in the matrix, a and b are the subscripts of the matrix elements and a=0, b=0 are located at the center of the matrix. And it satisfies:

[0045]

[0046] R 1 As a benchmark, calculate:

[0047]

[0048] and:

[0049]

[0050] Further, calculate:

[0051]

[0052] in

[0053]

[0054] The above k=1, 2, ..., 7 can be obtained according to formula 3 including R 1 There are 8 matrices in total, namely R 1 , R 2 , …, R 8 .

[0055] The above 8 matrices defined and calculated according to formula 1-3 form a filter group, and further filtering is performed on the filtering result of the scanned image. The calculation method is as follows:

[0056]

[0057] Results Corresponding to the filtering result of the previous step Three sets of filtering results are also obtained

[0058] The filtering effect of the above formulas 1-3 is to further strengthen the folds, thereby adding them to the recognition data source as a supplementary feature, so that the discrimination model can more effectively discriminate the standard level of the test paper.

[0059] calculate:

[0060]

[0061] in

[0062]

[0063] Vertical bars indicate absolute values.

[0064] According to 4 and 5, each answered questionnaire has a filtered "difference" image with the blank questionnaire, that is, Serves as input source for identification in the next step.

[0065] Step 2: Recognition of the scanned results of the paper questionnaire, based on the input data obtained in step 1 Determine whether the corresponding questionnaire meets the survey standards.

[0066] First, Divide into several sub-areas of equal size: D s (x, y), where

[0067]

[0068] A neural network model is established so that its input is the above sub-region, and a hidden layer is further established as follows.

[0069]

[0070] in, represents the maximum value of the corresponding position in the equal-sized sub-image, α 1c is the linear convolution kernel, β 0c is the corresponding bias, c represents the number of independent convolution kernels, and c = 64 is preferred in this example. λ is the activation function and is defined as:

[0071]

[0072] Compared with the classic activation function, the above 7 equations are more robust to questionnaire data recognition.

[0073] Further, define:

[0074]

[0075] in, Represents the maximum value corresponding to the response of different convolution kernels in the first layer, α 2c is the linear convolution kernel, β 1c is the corresponding bias, c represents the number of independent convolution kernels, and c = 64 is preferred in this example. The definition of λ is shown in 7.

[0076] The maximum convolution layer defined by the above equations 6 and 8 has better recognition effect than the classic convolutional hidden layer.

[0077] Further, define:

[0078]

[0079] where α 3 represents the full connection coefficient, β 2 is the corresponding bias, and λ is defined as in 7. The output of the third hidden layer is a 64-dimensional vector, i.e., 1≤t≤64.

[0080] Furthermore, the output layer is defined:

[0081]

[0082] The above 6-10 constitute a neural network model. When training it, it is agreed that the true output value of the positive sample (the questionnaire meets the standard) is The output value of negative samples (questionnaires do not meet the standards) is The cost function during training can be

[0083]

[0084] According to the above model output recognition results, when the model output ∈>0, it means that the questionnaire corresponding to the input meets the standard, otherwise, it means that it does not meet the standard.

[0085] The present invention proposes an automatic scanning and recognition method for graduate employment questionnaires. The experimental results (Table 1) show that the recognition effect of the method of the present invention is good, and the efficiency of the survey is improved for anonymous, non-directional, and large-scale employment questionnaire surveys.

[0086] Table 1

[0087]

[0088] The present invention also includes an automatic scanning and recognition system for graduate employment questionnaires, which includes a camera and a processor. The camera is used to collect questionnaire images, and the processor is used to execute the above method. The camera and the processor are connected via network communication.

[0089] The present invention can also be embedded in an existing questionnaire scanning and processing system. It includes a standard discrimination module, which is used to execute the above method. The module is embedded in the existing questionnaire processing system and connected to the questionnaire processing statistical analysis module.

Claims

1. A scanning and recognition method for graduate employment questionnaires, Features: Collect a scanned image of a blank questionnaire, denoted as B; collect a scanned image of the answered questionnaire to be identified, denoted as S; The B and S images are filtered by selecting different window sizes σ; where σ is selected as σ 0 , 2σ 0 , 4σ 0 ; Three groups of B image filtering results can be obtained respectively And three sets of S image filtering results Compute the matrix: We get 8 matrices, among which: The above three groups of S image filtering results are filtered again using the above 8 matrices to obtain three groups of filtering results. calculate: in Will Divide into several sub-areas of equal size: D s (x, y); and feed it into the first layer of the neural network as input: in, represents the maximum value of the corresponding position in the equal-sized sub-image, α 1c is the linear convolution kernel, β 0c is the corresponding bias, c represents the number of independent convolution kernels, and λ is the activation function; Neural network is used to judge whether the questionnaire meets the standards.

2. A method for scanning and identifying a graduate employment questionnaire as claimed in claim 1, Features: c＝64。 3. A method for scanning and identifying a graduate employment questionnaire as claimed in claim 1, Features: The above filtering is Gaussian filtering.

4. A method for scanning and identifying a graduate employment questionnaire as claimed in claim 1, Features: When training a neural network, the cost function is According to the above model output recognition results, when the model output ∈>0, it means that the questionnaire corresponding to the input meets the standard, otherwise, it means that it does not meet the standard.

5. An automatic scanning and recognition system for graduate employment questionnaires, Features: It includes a camera and a processor, and the processor is used to execute the scanning and recognition method of the graduate employment questionnaire described in any one of claims 1-4.

6. The automatic scanning and recognition system for graduate employment questionnaire according to claim 5, Features: The camera is used to capture questionnaire images.

7. The automatic scanning and recognition system for graduate employment questionnaire according to claim 5, Features: The camera and the processor are connected via network communication.

8. A scanning and processing system for graduate employment questionnaires, Features: It includes a standard identification module, which is used to execute the scanning and identification method of the graduate employment questionnaire described in any one of claims 1-4.

9. A scanning and processing system for graduate employment questionnaires as claimed in claim 8, Features: This module is embedded in the existing questionnaire processing system.

10. A scanning and processing system for graduate employment questionnaires as claimed in claim 9, Features: This module is connected with the questionnaire processing and statistical analysis module.

Images