A paper score recognition method and system based on reverse drip algorithm

Abstract

The present invention discloses a test paper score recognition method and system based on a reverse drip algorithm. Obtain a binary image containing handwritten scores at the position of the total score and each question score box in the RGB image of the test paper to be processed; use a reverse drip segmentation algorithm to segment multiple handwritten numbers into single numbers; use an indirect centroid method to process the single number image into a fixed image; a classifier predicts the number to form a score. The present invention proposes a reverse drip algorithm based on the traditional character segmentation method. Since the appearance of adhesions is mainly due to connected strokes written by humans, and connected strokes often appear from the lower left to the upper right, it is easier to find the depression of dripping. The reverse drip algorithm can solve the dilemma of the general character segmentation algorithm, and compared with the forward drip algorithm, it can greatly improve the prediction accuracy of the single number after cutting.

Description

Technical Field

[0001] The present invention belongs to the field of image recognition technology, and in particular relates to a test paper score recognition method and system based on a reverse drip algorithm. Background Art

[0002] With online grading not widely available and not widely used in daily exams, exams are primarily paper-based. Teachers have to grade, compile, and record scores in Excel spreadsheets. This creates a significant workload for teachers with a large student population and frequent exams. This system, based on this background, has developed an algorithm that can directly export student scores to Excel spreadsheets after capturing test paper images, significantly reducing teacher workload and improving teaching quality.

[0003] After scanning all students' exam papers using an automated scanner, the system batch-processes each student's paper, generates scores, and stores them in an Excel spreadsheet. By designing exam papers with specific score and student ID boxes, the student ID information and teacher's scoring information are precisely cut out and sorted. The cutouts are then processed sequentially, with reverse drip-cutting of individual digits performed in a loop. The indirect centroid method is then used to fill in the numbers and predict them. The predicted individual digits are combined to obtain the student ID and score information, which are then imported into an Excel spreadsheet, achieving automated grading. Summary of the Invention

[0004] The first purpose of the present invention is to address the deficiencies in the prior art and propose a test paper score recognition method based on a reverse drip algorithm.

[0005] A method for identifying a test paper score based on a reverse drip algorithm includes a test paper score extraction part; the method is characterized in that the test paper score extraction part specifically includes the following steps:

[0006] Step S21: Obtain the images of the total score and the score boxes of each question in the RGB image of the test paper to be processed, and convert them into HSV images; perform specific color threshold segmentation on the above HSV images to obtain the handwritten score images in the required specific color space, and convert them into binary images; perform an expansion operation on the handwritten scores in the binary image to obtain a clearer binary image containing the handwritten scores;

[0007] Step S22: Determine whether the number in the binary image containing the fraction in step S21 is a single number. If so, do not process it and jump to step S24; if not, proceed to step S23;

[0008] Step S23: Segment multiple handwritten digits into single digits

[0009] Perform a cutting operation on the multi-digit binary image using a reverse dripping algorithm to obtain multiple single digital images, perform position encoding on the single digital images, and then jump to step S24;

[0010] S24 indirect centroid method to fill pictures

[0011] The indirect centroid method was used to process a single digit image into an image of size 28 × 28 pixels, with the digit at the center and the interval between the digit boundary and the image boundary being a1 pixels on the left and a2 pixels on the top.

[0012] Step S25: Single digit prediction and combination

[0013] The classifier receives the single digital image after processing in step S24 and outputs the corresponding number; if the original image in step S22 is judged to be a single number, it is directly used as the final score; if the original image in step S22 is judged to be a multi-digit number, the classifier output results are combined according to the position code of the multi-digit number to form the corresponding score.

[0014] Preferably, the method further includes a student ID extraction part; specifically comprising the following steps:

[0015] Step S11: obtaining a picture of the middle school number frame in the RGB image of the test paper to be processed;

[0016] Step S12: further segmenting the image of the student ID frame to obtain a student ID image;

[0017] Step S13: grayscale, binarization, and cutting operations are performed on the student ID image to obtain a single digital image;

[0018] Step S14: Number the cut single digital pictures in sequence, use the indirect centroid method to repair the pictures in sequence, predict each single number and then combine them to obtain the student ID information.

[0019] The second object of the present invention is to provide an intelligent statistical system for detailed test paper scores based on machine learning, comprising: a memory, a processor, and a computer program stored in the memory and runnable on the processor, wherein the computer program, when executed by the processor, implements the steps of the above-mentioned intelligent statistical method for detailed test paper scores based on machine learning.

[0020] The beneficial effects of the present invention are as follows:

[0021] (1) Based on the traditional character segmentation method, the present invention proposes a reverse dripping algorithm. Since the appearance of contiguous bodies is mainly caused by handwritten connected strokes, and connected strokes often appear from the lower left to the upper right, it is easier to find the dripping depression. The reverse dripping algorithm can solve the difficulties of the general character segmentation algorithm and, compared with the forward dripping algorithm, can greatly improve the prediction accuracy of individual digits after segmentation.

[0022] (2) The present invention proposes an indirect centroid method to pre-process the cut images, which can greatly improve the prediction accuracy.

[0023] (3) The present invention has the function of automatic review and proofreading. After obtaining all the score information of the entire test paper, it will automatically check whether there is any error in the result of adding the predicted total score to the predicted score of each question. Test papers with doubts about proofreading will be automatically marked to indicate possible errors, so as to facilitate subsequent manual calibration. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 This is a flow chart of an intelligent statistical method for detailed test scores based on machine learning according to the present invention.

[0025] Figure 2 Extract the schematic diagram for the student ID.

[0026] Figure 3 (1) is a schematic diagram of fractional patch extraction, where (a) is the total score frame, (b) is the threshold segmentation, and (c) is the converted binary image; (2) is a schematic diagram of the image optimized by the dilation operation.

[0027] Figure 4 Schematic diagram of number segmentation using the reverse drip algorithm. (a) shows the result of segmenting the handwritten number 86, and (b) shows the result of segmenting the handwritten number 89.

[0028] Figure 5 Schematic diagram of filling pictures using the indirect centroid method, where (a) is before filling and (b) is after filling.

[0029] Figure 6 Schematic diagram for multi-digit score prediction.

[0030] Figure 7 These are the image sets of digital cutting results of the dripping algorithm and the reverse dripping algorithm in test case 1, where (1) is the image set of the dripping algorithm, and all black represents unsuccessful cutting; (2) is the image set of the reverse dripping algorithm.

[0031] Figure 8 The results of direct prediction (left) and indirect centroid method (right) for the deviation image in test case 2 and the deviation image set used for testing.

[0032] Figure 9 For the review result.

[0033] Figure 10 is the starting image of the moving point P in the binary image. DETAILED DESCRIPTION

[0034] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0035] The present invention provides a method for identifying test scores based on a reverse drip algorithm. Figure 1 As shown, it includes the student number extraction part and the paper score extraction part;

[0036] A paper score recognition method based on reverse drip algorithm includes a student ID extraction part and a paper score extraction part;

[0037] The student number extraction part specifically includes the following steps:

[0038] Step S11: obtaining a picture of the middle school number frame in the RGB image of the test paper to be processed;

[0039] Step S12: further segmenting the image of the student ID frame to obtain a student ID image;

[0040] Step S13: grayscale, binarization, and cutting operations are performed on the student ID image to obtain a single digital image;

[0041] Since the student ID number is located at a fixed position on the test paper, you can simply perform regular cutting operations on the student ID frame at the fixed position on the test paper.

[0042] Step S14: Number the cut single digit pictures in sequence, use the indirect centroid method to repair the pictures in sequence, predict each single digit and combine them to get the student number information, see Figure 2 .

[0043] The paper score extraction part specifically includes the following steps:

[0044] Step S21: Since the total score and the score frame of each question are fixed, conventional technology can be used to obtain the image of the total score and the score frame of each question in the RGB image of the test paper to be processed, and convert it into an HSV image; perform specific color threshold segmentation on the above HSV image to obtain the handwritten score image in the required specific color space, and convert it into a binary image; perform an expansion operation on the handwritten score in the binary image to obtain a clearer binary image containing the handwritten score, see Figure 3 .

[0045] The conversion into HSV images is achieved through the Python image processing module OpenCV.

[0046] The HSV component ranges of the specific color space are 0-10, 156-180, 43-255, and 46-255.

[0047] Step S22: Determine whether the number in the binary image containing the fraction in step S21 is a single number. If so, do not process it and jump to step S24; if not, proceed to step S23;

[0048] The judgment step S21 is to determine whether the number in the binary image containing the handwritten score is a single digit. Specifically, the operation is to confirm the width W and height H of the largest box of the character in the binary image containing the score by traversing the rows and columns, and calculate the aspect ratio W / H. If W / H>a, it is considered to be a multi-digit number. If W / H<a, it is considered to be a single digit. a represents a threshold value, which can be 0.85 in the present invention.

[0049] Step S23: Segment multiple handwritten digits into single digits

[0050] Perform a cutting operation on the multi-digit binary image using a reverse dripping algorithm to obtain multiple single digital images, perform position encoding on the single digital images, and then jump to step S24;

[0051] Since the appearance of adhesions is mainly due to connected strokes of human handwriting, and connected strokes often appear from the lower left to the upper right, it is easier to find the depression caused by dripping. Based on the dripping algorithm, the present invention proposes a reverse dripping algorithm, which can solve the dilemma of the general character segmentation algorithm. Compared with the general forward dripping algorithm, the reverse dripping algorithm can greatly improve the prediction accuracy of single digits after cutting.

[0052] The reverse dripping algorithm specifically includes the following steps: Figure 4 :

[0053] ① Normalize the binary image of multiple digits;

[0054] ②See Figure 10, after normalization, a moving point P is introduced into the image, and the five pixels at the upper left, upper right, upper left, and right of point P are marked as P1, P2, P3, P4, and P5 respectively; P0 is taken as the starting point, where the coordinates of P0 are (roundup(L×0.55+R×0.45), H); roundup represents the upward rounding function, L and R represent the left and right horizontal coordinates of the maximum frame of the multi-digit number in the image, and H is the height of the image; the moving point P is given the following behavior mode: if P4 and P5 are non-handwriting pixels, the moving point P moves up one pixel; if P1, P2, and P4 are handwriting pixels, P 3. If P5 is a non-handwriting pixel, the moving point P moves 1 pixel to the right and 1 pixel upwards; if P1, P2, P3, and P4 are handwriting pixels and P5 is a non-handwriting pixel, the moving point P moves 2 pixels to the right and 1 pixel upwards; if P1, P2, P3, and P5 are handwriting pixels and P4 is a non-handwriting pixel, the moving point P moves 2 pixels to the left and 1 pixel upwards; except for the above cases, the moving point P moves 1 pixel to the left and 1 pixel upwards; the moving point P moves from bottom to top according to the above behavior pattern, and records the position of the moving point P before each movement until it reaches the first row of pixels; specifically:

[0055] The initial coordinate is P0(X0, Y0), and the coordinate after one judgment operation is P ’ 0(X ’ 0,Y ’ 0).

[0056]

[0057] ③ Record the motion trajectory of the moving point P in step ② above, and segment the original image into individual digital images according to the motion trajectory of the moving point P.

[0058] S24 indirect centroid method to fill pictures

[0059] Using the indirect centroid method, a single digital image is processed into an image with a size of 28×28 pixels, with the number located in the center, the number boundary and the image boundary separated by a1 pixels on the left and a2 pixels on the top; in the present invention, a1 can be 5 and a2 can be 3;

[0060] The indirect center of gravity method specifically includes the following steps: Figure 5 :

[0061] ① Get the center of gravity of a single digital image (the center of gravity can be obtained by obtaining the distance between the upper, lower, left, and right boundaries of the entire digital character and the upper, lower, left, and right edges of the image); specifically:

[0062] Traverse the rows and columns of the image and obtain the distance L1 between the left boundary of the handwritten digit and the left edge of the image. Similarly, obtain the distances R1, U1, and D1 between the right, upper, and lower boundaries and the right, upper, and lower edges of the image, respectively.

[0063] ② Adjust the single digital image according to the horizontal coordinate of the center of gravity. If the center of gravity is biased to the left, add a black area to the left side of the image or cut off the black area on the right side; if the center of gravity is biased to the right, add a black area to the right side of the image or cut off the black area on the left side;

[0064] ③ According to the vertical coordinate of the center of gravity of the digital image, adjust the image after step ②. If the center of gravity is upward, add a black area to the upper side of the image or cut off the black area on the lower side; if the center of gravity is downward, add a black area to the lower side of the image or cut off the black area on the upper side;

[0065] ④ Adjust the digital image processed in step ③ to a fixed size;

[0066] The specific operation process of cutting off or adding black areas:

[0067] Determine whether L1 is equal to a1. If so, do nothing. If not, compare L1 with a1. If it is greater than a1, remove pixel X1. Otherwise, add pixel X1.

[0068] Determine whether R1 is equal to a1. If so, do nothing. If not, compare R1 with a1. If it is greater than a1, remove pixel X2. Otherwise, add pixel X2.

[0069] Determine whether U1 is equal to a2. If so, do nothing. If not, compare U1 with a2. If it is greater than a2, remove pixel X3. Otherwise, add pixel X3.

[0070] Determine whether D1 is equal to a2. If so, do nothing. If not, compare D1 with a2. If greater than a2, remove pixel X4. Otherwise, add pixel X4.

[0071]

[0072] Where i = 1, 2; roundup (Xi) represents X1 and X2 respectively; Yi represents Y1 and Y2, corresponding to L1 and R1 respectively;

[0073]

[0074] Where j = 3, 4; roundup (Xj) represents X3 and X4 respectively; Yj represents Y3 and Y4, corresponding to U1 and D1 respectively;

[0075] Example: Assuming L = 8, then you need to truncate pixel X. After truncation, the distance between the digital character and the left border is (8-X). Calculate the ratio of (8-X) to (28-X). When , X is rounded to 4, that is, 4 pixels are cut off on the left and 4 pixels are added on the right.

[0076] Step S25: Single digit prediction and combination

[0077] The classifier receives the single digit image processed in step S24 and outputs the corresponding digit. If the original image in step S22 is judged to be a single digit, it is directly used as the final score. If the original image in step S22 is judged to be a multi-digit digit, the classifier output results are combined according to the position code of the multi-digit digit to form the corresponding score. Figure 6 .

[0078] Import the score of step S25 into the corresponding question score position in Excel to obtain the score summary of the test paper; after obtaining the scores of all questions and the total score, the system checks whether the sum of the question scores is the same as the total score. If so, the score of the test paper is directly output; if not, the test paper is marked for subsequent manual calibration. The predicted score of each question is recorded as Q1, Q2, Q3... The total score obtained by the prediction is recorded as Q. If Q=Q1+Q2+Q3+..., the prediction is accurate and marked as correct; if Q≠Q1+Q2+Q3+..., it is marked as wrong. Enter the student ID information of step S14 into the student ID directory in the Excel table. In this way, traverse each test paper scan and each set of student test papers to obtain the score information of all students. Figure 9 For automatic review of the results, if the predicted total score is the same as the sum of the predicted scores of each question, it is considered correct.

[0079] In step S21, the HSI color model is based on the human visual system, using H for hue, S for saturation, and I for intensity to describe color. Generally, effective processing of images in color space is performed in the HSV space, and then a strict range needs to be given to the corresponding HSV components in the basic colors. The fuzzy range of the red part (0-10, 156-180, 43-255, 46-255) is obtained through experimental calculation, and color is taken. After the score of the specific color space is extracted, it is converted into a binary image containing the score to obtain an image of the total score of the paper and an image of the score of each question.

[0080] Currently, there are no established and widely adopted algorithms specifically designed for handwritten digit recognition in exams. While traditional recognition algorithms like KNN have been effective for handwritten digit prediction, drawbacks such as low recognition rates and slow speeds have led to the emergence of new recognition technologies. In recent years, convolutional neural networks (CNNs) have been widely used for handwritten digit recognition due to their excellent generalization, memory, nonlinearity, self-learning capabilities, and high-speed computing power. These include models like CNN and LENET-5.

[0081] The MNIST handwritten digit dataset is a collection of handwritten digits from the National Institute of Standards and Technology (NIS). It contains 60,000 training samples and 10,000 test samples. The grayscale values ​​of each digit image are normalized to a 28x28 pixel size. The present invention uses this handwritten digit dataset as both the training and test sets, using the final accuracy to select the final handwritten digit predictor. However, the prediction accuracy is very low for unprocessed or irregular handwritten digit images. Therefore, the present invention develops an indirect centroid method to preprocess the cut images, which can significantly improve prediction accuracy.

[0082] Test 1: The prediction accuracy of a single digit image cut by the reverse drip algorithm is compared with the accuracy of the (forward) drip algorithm. The results are as follows Figure 7 As shown in the figure, the accuracy of the dripping algorithm is only 47%, while the accuracy of the improved reverse dripping algorithm is increased to about 90%.

[0083] Test 2: For the cut digital images, if they are predicted directly without any processing, the accuracy will be very low. 50 images were selected and predicted directly and using the indirect centroid method to pre-process the cut images before prediction. The results are as follows: Figure 8 As shown in Figure 2, the accuracy of direct prediction is only 16%, while all 50 deviation images are accurately predicted after processing using the indirect centroid method.

Claims

1. A paper score recognition method based on the reverse drip algorithm, including a paper score extraction part; characterized in that The paper score extraction part specifically includes the following steps: Step S21: Obtain the images of the total score and the score boxes of each question in the RGB image of the test paper to be processed, and convert them into HSV images; perform color threshold segmentation on the HSV images to obtain the handwritten score images in the required color space, and convert them into binary images; perform an expansion operation on the handwritten scores in the binary images to obtain a clearer binary image containing the handwritten scores; Step S22: Determine whether the number in the binary image containing the handwriting score in step S21 is a single digit. If so, do not process it and jump to step S24; if not, proceed to step S23; Step S23: Segment multiple handwritten digits into single digits Perform a cutting operation on the multi-digit binary image using a reverse dripping algorithm to obtain multiple single digital images, perform position encoding on the single digital images, and then jump to step S24; The reverse dripping algorithm specifically includes the following steps: ① Normalize the binary image of multiple digits; ② After normalization, a moving point P is introduced into the image. The five pixels at the upper left, upper right, upper left, and right of point P are marked as P1, P2, P3, P4, and P5 respectively. P0 is taken as the starting point, where the coordinates of P0 are (roundup(L×0.55+R×0.45), H). Roundup represents the upward rounding function, L and R represent the left and right horizontal coordinates of the maximum frame of the multi-digit number in the image, and H is the height of the image. The moving point P is given the following behavior mode: if P4 and P5 are non-handwriting pixels, the moving point P moves up one pixel; if P1, P2, and P4 are handwriting pixels, the moving point P moves up one pixel. , P3 and P5 are non-handwriting pixels, then the moving point P moves 1 pixel to the right and 1 pixel upwards; if P1, P2, P3, and P4 are handwriting pixels and P5 is a non-handwriting pixel, then the moving point P moves 2 pixels to the right and 1 pixel upwards; if P1, P2, P3, and P5 are handwriting pixels and P4 is a non-handwriting pixel, then the moving point P moves 2 pixels to the left and 1 pixel upwards; except for the above cases, the moving point P moves 1 pixel to the left and 1 pixel upwards; the moving point P moves from bottom to top according to the above behavior pattern, and records the position of the moving point P before each movement until it reaches the first row of pixels; ③ Record the motion trajectory of the moving point P in step ② above, and segment the original image into individual digital images based on the motion trajectory of the moving point P; Step S24: Filling the image using the indirect centroid method The indirect centroid method was used to process a single digit image into an image of size 28 × 28 pixels, with the digit at the center and the interval between the digit boundary and the image boundary being a1 pixels on the left and a2 pixels on the top. Step S25: Single digit prediction and combination The classifier receives the single digital image after processing in step S24 and outputs the corresponding number; if the original image in step S22 is judged to be a single number, it is directly used as the final score; if the original image in step S22 is judged to be a multi-digit number, the classifier output results are combined according to the position code of the multi-digit number to form the corresponding score.

2. The method according to claim 1, characterized in that The method further comprises a student number extraction part; The student number extraction part specifically includes the following steps: Step S11: obtaining a picture of the middle school number frame in the RGB image of the test paper to be processed; Step S12: further segmenting the image of the student ID frame to obtain a student ID image; Step S13: grayscale, binarization, and cutting operations are performed on the student ID image to obtain a single digital image; Step S14: Number the cut single digital pictures in sequence, use the indirect centroid method to repair the pictures in sequence, predict each single number and then combine them to obtain the student ID information.

3. The method according to claim 1, characterized in that The conversion to HSV image described in step 21 is implemented through the Python image processing module OpenCV.

4. The method according to claim 1, characterized in that The HSV component ranges of the color space in step 21 are 0-10, 156-180, 43-255, and 46-255.

5. The method according to claim 1, characterized in that The operation of determining whether the number in the binary image containing the handwriting score in step S21 is a single number in step 22 is specifically as follows: confirming the width W and height H of the largest box of the character in the binary image containing the handwriting score by traversing the rows and columns, calculating the aspect ratio W / H, if W / H>a, it is considered to be a multi-digit number, if W / H<a, it is considered to be a single number; a represents the threshold value.

6. The method according to claim 1, characterized in that The behavior pattern of the moving point P in step ② of the reverse dripping algorithm in step S23 is specifically: The initial coordinate is P0(X0, Y0), and the coordinate after one judgment operation is P ’ 0(X ’ 0,Y ’ 0); 7. The method according to claim 1, characterized in that The indirect centroid method in step S24 specifically includes the following steps: ① Get the center of gravity of a single digital image; specifically: Traverse the rows and columns of the image and obtain the distances L1, R1, U1, and D1 between the left, right, top, and bottom boundaries of the handwritten numbers in the image and the left, right, top, and bottom edges of the image respectively; ② Adjust the single digital image according to the horizontal coordinate of the center of gravity. If the center of gravity is biased to the left, add a black area to the left side of the image or cut off the black area on the right side; if the center of gravity is biased to the right, add a black area to the right side of the image or cut off the black area on the left side; ③ According to the vertical coordinate of the center of gravity of the digital image, adjust the image after step ②. If the center of gravity is upward, add a black area to the upper side of the image or cut off the black area on the lower side; if the center of gravity is downward, add a black area to the lower side of the image or cut off the black area on the upper side; ④ Adjust the digital image processed in step ③ to a fixed size.

8. The method according to claim 7, characterized in that The specific operation process of cutting off or adding black areas: Determine whether L1 is equal to a1. If so, do nothing. If not, compare L1 with a1. If it is greater than a1, remove pixel X1. Otherwise, add pixel X1. Determine whether R1 is equal to a1. If so, do nothing. If not, compare R1 with a1. If it is greater than a1, remove pixel X2. Otherwise, add pixel X2. Determine whether U1 is equal to a2. If so, do nothing. If not, compare U1 with a2. If it is greater than a2, remove pixel X3. Otherwise, add pixel X3. Determine whether D1 is equal to a2. If so, do nothing. If not, compare D1 with a2. If greater than a2, remove pixel X4. Otherwise, add pixel X4. Where i = 1, 2; roundup (Xi) represents X1 and X2 respectively; Yi represents Y1 and Y2, corresponding to L1 and R1 respectively; Where j = 3, 4; roundup(Xj) represents X3 and X4 respectively; Yj represents Y3 and Y4 respectively, corresponding to U1 and D1 respectively.

9. The method according to claim 1, characterized in that Import the scores of step S25 into the corresponding question score position in Excel to obtain the score summary of the test paper; after obtaining the scores of all questions and the total score, the system checks whether the sum of the question scores is the same as the total score. If so, the score of the test paper is directly output; if not, the test paper is marked for subsequent manual calibration.

10. An intelligent statistical system for detailed test scores based on machine learning, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program implements the steps of the method according to any one of claims 1 to 9 when executed by the processor.

Images