CN105095892A - Student document management system based on image processing - Google Patents
Student document management system based on image processing Download PDFInfo
- Publication number
- CN105095892A CN105095892A CN201410208709.0A CN201410208709A CN105095892A CN 105095892 A CN105095892 A CN 105095892A CN 201410208709 A CN201410208709 A CN 201410208709A CN 105095892 A CN105095892 A CN 105095892A
- Authority
- CN
- China
- Prior art keywords
- color
- student
- document
- present
- picture
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Landscapes
- Image Analysis (AREA)
Abstract
The invention discloses a student document management system based on image processing, and concretely provides a student document management system based on image processing under paper-based document conditions, relating to the image processing technology. The system comprises: scanning student documents to obtain images and input the images into the system; denoising the document images; marking out the position of each subject in the document images; performing color clarification on the document images, screening red pixels, and furthermore finding out symbols annotated on the documents by teachers; finding out and identifying symbols in each subject area; and storing images of each subject and corresponding teacher feedback information to a database. The system has high accuracy of error identification, which is a key index. Advanced reading and writing equipment, and advanced and improved software make data classification statistics easy and accurate, allow examination and homework to be completed on equipment, and provide supporting technology for the achievement of information digitlization.
Description
Technical field
The present invention relates to technical field of image processing, specifically refer to a kind of student's document file management system based on image procossing under paper document condition.
Background technology
Middle school student, in daily study, can do a large amount of exercises and paper usually, and these paper materials accumulate with accumulating over a long period, and finally will become a very useful material for review.Student by looking back the exercise done in the past, can carry out leakage detection and filling a vacancy, especially can make up the weakness of oneself.But students, in the process of management legacy papery learning materials, can run into a maximum problem: the paper material quantity of long term accumulation is very huge, be not easy to very much to consult targetedly, retrieve, review.Certainly, there are some conventional means to improve and consult efficiency, as data undertaken classifying by subject, time and sequence etc., more having student to safeguard a wrong answer list specially, the exercise question done wrong is copied, but more student can only search in the ocean of operation and paper blanket type.And in order to safeguard a wrong answer list, student also needs to expend how extra energy usually, and this causes very large burden to them undoubtedly.
For teacher, how to make full use of good student's ongoing operations and this kind of resource of examination paper, in fact also Shortcomings.Because the content in papery data is difficult to the statistics of carrying out robotization, therefore teacher is difficult to obtain the quantitative information about Students ' Learning quality.Teacher, only according to borrowing oneself memory, is difficult to advantage and the weakness of accurately understanding each student, more difficultly finds out the learning state in student's each stage etc.Solve a lot of problem in study, many times will suit the remedy to the case, if the operation testing situations of student can be followed the tracks of well, huge help must be produced to teaching.
The existing digital management technology for teaching, mainly concentrates on and gos over examination papers and criticize volume.For example use answer sheet to correct multiple-choice question, this scheme to a certain degree achieves digitizing.But also clearly, it cannot be applied to the exercise question that question and answer of filling a vacancy etc. need to write to the limitation of this technology.For example carry out electric marking by scanning again, facilitate the statistics of total marks of the examination, even can realize cooperation between school and correct.But for student, this cover system does not directly help them, they still by traditional means management paper document, can only inscribe as arranged mistake, sorting out document etc.And such set of system being directly used in ongoing operations also and improper, teacher needs computer to read and make comments every day, bears heavier.Therefore, in the field of paper document digital management, not yet there is ripe system to can be used for academics and students to use.
Summary of the invention
The object of the invention is to the disappearance and the deficiency that overcome prior art existence, propose a kind of student's document file management system based on image procossing, scan by the paper of the operation marked, examination is concentrated, then the mistake topic in automatic discrimination paper is carried out by image recognition, finally automatically collected by computer and sort out wrong topic, thus realizing the classification of data digitalization and intelligence.
Based on student's document file management system of image procossing, workflow (step):
The first, scan student's document and obtain image, input system.
The second, file and picture denoising is carried out.
Three, from file and picture, mark off per pass topic destination locations.
Four, color classification is carried out to document image pixels, filter out red pixel point, and then find out the symbol that teacher annotates and comments on document.
Five, find the symbol in per pass title field and identify.
Six, by the image of per pass exercise question, and teacher's feedback information of correspondence, stored in database.
Due in reality, the typesetting of paper, the mode of teacher comment paper vary, may far beyond the soluble category of general pattern treatment technology, therefore need to increase some restrictions to application link, appropriate definition, make it can describe most of occasion in practical application, and succinct and graceful solution can be found under this condition.
Accordingly, be defined as follows:
1. paper with the size of A4 paper for one page.In reality, most papers is all that every face prints one page or two pages, and for the latter, it can be split into two single pages by the present invention after scanning automatically, thus can be absorbed in the process of single page paper.
2. the exercise question on paper arranges from top to bottom, and character calligraph order from left to right.Have relatively abundant interval between upper and lower twice exercise question or stay white, student is at exercise question underlying empty place answer exercise question, and the usual enough students of blank parts write answer.Subfield is not carried out in page inside.
3. paper adopts blackness handwriting, hand-written or printing.The writing color of student is black or blueness, and the result of reading and making comments of teacher is write by red pen.
4. per pass exercise question, teacher adopts and makes hook or make fork to pass judgment on.
5. paper changes into digital image format by color scanner batch.Paper allows suitable inclination when scanning, but angle of inclination is generally no more than 5 degree.
In this task, input is the paper of all students of same one-stop operation or examination.The present invention obtains digitized paper image by scanner.Desired as a result, system automation ground by each the problem object region recognition on paper out, and detects that teacher's corrects mark in this examination question region, then carries out pattern-recognition to correcting mark, thus knows correcting errors of this topic.
Present system is summarized as follows:
System, adds external unit, six workflows (as shown in Figure 1) mainly comprised:
First carry out paper scanning, next will carry out file and picture denoising, exercise question location recognition, color classification, Symbol recognition four core procedures (workflow) successively.
In order to show system of the present invention more intuitively, the paper after scanning really with below as an example, shows the effect that the present invention can reach.
First the present invention carries out file and picture denoising, obtains a more clean image.Then the present invention attempts from paper, mark off each problem destination locations, and the region of per pass exercise question is taken out separately by such the present invention, just as the least unit of filing management afterwards.Next, the present invention carries out color classification to the pixel on paper, leaves red pixel, then they is merged into complete symbol, and describes its position and size with rectangle on image.Next, the present invention finds the symbol in per pass title field and identifies, thus knows that it hooks or fork, namely to still wrong.Finally, present invention obtains the image of all exercise questions, and the positive false information of every problem, just they can be deposited to database, so that wrong topic is classified and every statistics in the future.
File and picture denoising
The digital document images that paper obtains by scanning, and between the true picture that human eye is seen, be discrepant in fact, wherein except the difference of the globality such as brightness, tone, also have some trickle distorted appearance, make run into some problems when processing digital images.The performance of this distortion, needs the noise paid close attention to exactly.
Most typical noise in scanning process is the coarse dark line caused due to paper.Coarse paper can form diffuse reflection, there will be the change of light and shade under the scanning of light: the place of projection can seem brighter, and sunk place then seems darker.Time this makes the present invention observe scanning result, many tiny dim spots can be seen.Some dim spots like this, can have influence on the accuracy of follow-up work.
In order to remove these dim spots, present invention employs the mode of gaussian filtering.As is generally known, gaussian filtering is exactly be weighted average process to the local of file and picture, namely on each pixel, the pixel value that its pixel value and surrounding close on is weighted on average by the present invention, calculate a new value out, and the new file and picture that this mode calculates be exactly filtering after file and picture.After point dark in the present invention and comparatively light ground weighted mean, closer to the color of background, therefore can not see obvious difference in new file and picture.
In OpenCV storehouse, itself has implemented this function, therefore only need choose suitable parameter and call corresponding function.Parameter is window size, and implication is that the pixel in the how many distance of each pixel and surrounding is weighted on average.Window shape is rectangle, therefore will arrange length and width two numerical value.Because the size of input picture own is comparatively large, therefore window is also larger, and through experiment, arranging length and width is 9 effects that can reach best.
Finally need to carry one, the setting of this parameter, be with the quality of scanner and file and picture size closely-related.For different scanners, the best of this parameter is arranged also may not be the same.When reality uses system, this parameter needs to debug in advance together with scanner, and guarantee subsequent step stably runs.
Exercise question location recognition
Identify each problem destination locations on paper to solve, the present invention adopts statistical method.
Because the angle of inclination in scanning paper limit will within the specific limits, and pixel in the row that exercise question is occupied on file and picture is usually more, and therefore the quantity of black picture element in each pixel column of statistic document image, can draw a histogram.In figure, left one side of something is in all papers, and the pixel of black is at the histogram of every a line, and right figure is original paper.In histogram, the present invention observes some significant peak values, and namely the sum of black picture element that has of these row, approximately occupies laterally more than 70% of black picture element quantity at most.These peak values, the exercise question position in fact just corresponding paper.This paper has 5 road exercise questions, and wherein the exercise question of the first topic has 6 row, and top has the formula of black matrix and spacing comparatively large, and the second topic has 5 row, and other exercise questions all only have 1 row.By observing paper, the present invention can find corresponding relation wherein easily, and as can be seen here, statistical method has key effect to exercise question location recognition.
The present invention gives an algorithm to find each problem destination locations in figure as following table:
In design, man-machine interaction flow process can be added in this step, to ensure the robustness of net result.
Color classification
By the picture that color scanner obtains, contain comparatively accurate colouring information.In a computer, the mode that color stores has a variety of, and this depends on any colour model of employing.Colour model is the abstract mathematical model that employing one class value (being generally three to four) represents color, and such as primaries pattern (RGB) and printing four color separation patterns (CMYK) are all typical colour models.
Problem difficult point and the choice of technology
Color classification problem mainly contains two difficult points.First, it is a problem of combining closely with the subjective feeling of people, there is not a set of objective standard and judges that the color represented by a RGB tlv triple is any color actually.Therefore, the present invention will find in essence, is fuzzy, an acceptable sorter intuitively.The benefit that the subjectivity of color classification is brought need not seek accurately, only need to reach a rational limit, but the difficulty that it brings also to be obvious, is exactly judge that color exists randomness.The vision of people is very easy to the impact being subject to object of reference, and the object of a such as grey can seem bright under the background of darkness, then can seem dim under light ground.In this problem, along with the change of data set, as adopted the different red pen of color depth, or adopt the scanner that quality is different, the border of classification also has slight change, especially red white between intermediate color.How making the result of color classification " rationally " more, is the first problem that the present invention needs to solve.
Second difficult point of color classification, it is irregular for being embodied in the form of a kind of color in color space.For rgb color space, the present invention is using red, green, blue as solid axes, and so the set of all colours is exactly a cube, and a color is by the value of its three components, can correspond to a point in cube uniquely.If the present invention is aware of each point whether belong to red, just all can describe in cube by belonging to red point, these set of putting just have certain shape, should be a continuous and airtight spatial form usually.But in reality, this shape is difficult to describe in the extreme, to such an extent as to the present invention is difficult to classify by some simple rules.Even if in the HSV space having more Color Expression ability, also there is Similar Problems.Although form and aspect (Hue) can distinguish color to a certain extent, saturation degree (Saturation) and lightness (Value) still play sizable effect in differentiation color.
In view of above two factors, the present invention, in the choice of technology, first needs people to get involved color mark, thus allows the result of color classification on data set, present better visual effect.Secondly, on the model of color classification, roughly thinking of the present invention adopts the method finding neighbour: when attempting the rgb value that classification one did not mark, and the present invention is the color of inquiry with it corresponding to the immediate rgb value marked just.
On existing color classification algorithm, although there are the machine learning methods such as such as support vector machine (SVM) also can be competent at, in its final and the present invention, the result of look-up method there is no essential distinction.Because color space is also little, thus algorithm of the present invention is more succinct and efficient, is therefore the desirable selection of color classification.
Finally, because color classification needs to carry out each pixel of original document image, the efficient of inquiry therefore to be guaranteed.Present invention employs the mode of tabling look-up, whether namely pre-service goes out all possible rgb value is red, stores in the table, only need find respective items in table during inquiry.
In color classification, present invention employs the simplest RGB pattern to represent file and picture color, namely each pixel represents the size of its red, green, blue component respectively by three values, and this value is taken at 0 to 2
ninteger between-1.Like this, the size of rgb color space is (2
6)
3=2
18.The present invention found through experiments, and adopt such color space to be enough to reasonably carry out color classification, and color space is also unlikely to excessive, and available form stores, and therefore can be regarded as the equilibrium point between precision and space expense.
Whether constructed form, storing rgb value to this color is a red mapping.Build in the flow process of form, be based on the data of artificial mark at first.First the present invention allows people to get involved the process of color classification: some colors of first sampling out in color space, allows people to judge that whether it is as red.Next, infer whether the color of those the unknowns remaining is red by computing machine.The strategy of inferring is as follows: if in the color adjacent with certain non-fundamental color, redness occupies majority, and so this color is likely red, otherwise this color is not probably red.
By so simple means, by artificial mark, the present invention first can determine whether a part of rgb value is red, next only need run above-mentioned strategy iteratively, just the point of redness can be marked.Concrete algorithm is as shown in the table:
Generate this look-up table, the present invention just can judge whether each pixel is red rapidly.In addition, the present invention adopts and repeatedly manually marks, or the mode of multi-person labeling, effectively can reduce the error of artificial mark, make the result of color classification more accurate.
The mark interface (as shown in Figure 2) that the present invention writes, first this program extracts the color of all pixels in picture, and the maximum color of appearance is presented in the window of upper figure.Next, utilize the mouse of OpenCV window to click feedback mechanism, mark person clicks him and thinks red color lump, and program just records This move, and on this window, in corresponding grid, just draws rim expressive notation for red.Annotation window has 40 row 20 and arranges, and namely once marks 800 color lumps, thus generates labeled data required for the present invention.After mark terminates, obtain the annotation results of these color lumps, so the present invention generates color lookup table by algorithm 1, and preserve hereof, for present system is used.
Symbol recognition
In this part, solve two problems, first problem how to obtain the symbol that will identify, Second Problem is, judges that the symbol obtained hooks or fork.
By the color classification of previous step, the present invention can know in image, which pixel is red, and which pixel is non-redness.The connection block that all red pixels form by the present invention takes out, and can obtain symbol.But in true picture, the red writing in paper can with the exercise question of black, or the word that student writes has overlap and does not demonstrate redness, and this causes hook that image is beaten and fork to be incomplete, and some areas can disconnect because of blackness handwriting.Simple find the method being communicated with block if adopted, be to obtain a complete symbol, naturally also cannot accurately identify.
Solution of the present invention is, is identifying on the basis being communicated with block, then calculates the bee-line between any two connection blocks, then by the connection merged block of close together.Because the interval between different exercise question is larger, the threshold value being therefore easy to set out suitable distance carries out connection merged block, so the present invention just intactly can obtain symbol.
Identify that a symbol hooks or fork, the present invention adopts the method for machine learning, builds a sorter, and it typically includes following several step: data mark, feature extraction and training classifier.
Data mark
The essence of machine learning acquires rule from data.Therefore, first the present invention manually will mark data, allow computing machine acquire useful information from the data marked.
In problem of the present invention, the data that needs carry out classifying are the pictures of the symbol that the present invention finds from figure, and target of the present invention it is categorized into hook and fork.
The hook of random selecting and the data (as shown in Figure 3) of fork in data of the present invention:
In accompanying drawing, 9 symbols in left side are for hooking, and nine symbols on right side are fork.Data can be described with clear, this also embody before color classification result stable.In addition, in the data mark stage, mark data as much as possible normally helpful for the process of training pattern afterwards, because increasing along with data sample, the diversity of data set improves, and so the rule acquired of machine is also more reliable.
In the mark work of one embodiment of the present of invention, nearly 400 data of manual sort, wherein hook and pitch and respectively account for about half.Because the present invention tests in true paper used, the ratio of hook is much larger than fork, and the person of correcting of this part of paper only has two people, comparatively single, therefore asked again some people to supplement data set for the present invention in test, supplied the quantity of fork, made positive example and counter-example comparatively average.The thickness of person's handwriting is also the variable needing to pay close attention to, if data centralization only comprises thick person's handwriting or carefully takes down notes, so under the unmatched state of handwriting thickness, recognition accuracy may also can be given a discount.Therefore, the present invention has also used the pen of different thicknesses to acquire data.Adopt so relatively various data set to carry out training pattern, unknown use occasion can be adapted to better.
Feature extraction
About the classification problem of image, usually all be unable to do without feature extraction.At present at image domains, the algorithm of feature extraction is very various, as SIFT, HOG, SURF, LBP etc.But these feature extraction algorithms are used in task of the present invention, all have some limitations.And in work of the present invention, propose a kind of succinct feature extraction algorithm efficiently based on pixel, i.e. Pixel or the average algorithm of simple region, reach extraordinary classifying quality.
The picture of stored symbols of the present invention, is first normalized the black and white two-value picture becoming 144 × 144.The picture with symbol intercepted from former figure is normally rectangular.The present invention first passes through in the left and right of picture or increases certain white space up and down, is become foursquare picture, then by convergent-divergent, the invention of its Setup Cost is identified standard size used.Adopt supplement and non-stretching mode is become square, be the shape of the stretching meeting reindexing considering image, likely affect the identification of picture.
The characteristic of two-value picture itself has chosen significant impact to feature.First, the feature extraction mode that it makes with histograms of oriented gradients HOG is representative ideally can not play effect.This is the design due to histograms of oriented gradients, and exactly in order to catch the statistical information of all directions gradient in one piece of region, therefore in the picture with more thin portion texture, it can grab more information.But in two-value picture, first the gradient at edge just only has 8 directions, abundant not, and also have the region of a lot of solid color in picture, in such a region, HOG feature cannot obtain effective gradient information.Secondly, two-value picture also makes another kind of feature extraction mode, and the effect of LBP feature is had a greatly reduced quality.LBP feature is equally also a kind of feature extraction mode paying close attention to texture, and its core is exactly extract 8 pixels closed on around a pixel, by relatively around 8 pixels magnitude relationship with center pixel value, thus gives one 0 ~ 2 to this pixel
8coding between-1, obtains a histogram as feature eventually through each coding frequency of occurrences of statistics.The reason that it cannot play a role best on two-value picture, in fact very similar with HOG feature, the region that textural characteristics is not enriched it cannot extract enough information.
In the implementation, basically, file and picture itself also can use as a kind of feature, by each pixel of file and picture as each element of proper vector.But this feature is original, crude, and the cost therefore acquiring useful information from this feature is also very large, often needs a large amount of training datas.Meanwhile, the dimension of this feature is very high, makes its computing cost very high.But the present invention can find, this feature is for quite suitable two-value picture, because it can not loss of information significantly, even if for the region of monochrome, it also can reflect the information of color faithfully.Secondly, the present invention makes certain improvements in this feature, does not directly adopt pixel, but Iamge Segmentation is become a lot of lattice, calculates the average of the pixel value in each grid, as the element of proper vector.In this way, the present invention effectively can reduce the dimension of feature, and still retains enough effective informations, and this feature extraction algorithm very easily realizes, and is the very desirable feature extraction mode for two-value picture.
The present invention finally decides to adopt which kind of feature extraction mode by the mode of experiment, also draws the best results of which kind of parameter combinations under often kind of mode simultaneously.Present invention achieves the feature extraction algorithm that HOG and LBP these two kinds is ready-made, also achieve the feature extraction algorithm (Pixel or the average algorithm of simple region) based on pixel, contrast the effect of three, the self-designed algorithm of final the present invention achieves best effect and wins.
Training classifier
In machine learning, have much different classifier algorithms to complete classification problem, wherein common are based on Probability & Statistics Nave Bayesian Classifier (
bayesian), rule-based decision tree (DecisionTree), based on the artificial neural network (ArtificialNeuralNetwork) of feedback learning, also has support vector machine (SupportVectorMachine) split based on higher dimensional space etc.In work of the present invention, present invention employs the sorter of support vector machine, this has the features such as fast, the applicable high dimensional data classification of speed due to it.
In system, some third parties are adopted to increase income storehouse.
First be a ripe C++ image procossing storehouse OpenCV, utilize this storehouse, the present invention can carry out the reading of common format picture, display, amendment and preservation and operate.OpenCV provides much general image processing function, the function that for example the present invention has used the image smoothing that it carries carries out image denoising, also use the function of dimension of picture adjustment and so on, to simplify the workload that the present invention realizes, but remaining image processing function is that oneself realizes completely.
In addition, invention also uses the C++ storehouse LibSVM of a support vector machine (SVM), it is the derivation algorithm that the invention provides multiple support vector machine, coordinates image characteristics extraction algorithm, and it may be used for the picture hooking and pitch of classifying.
Accompanying drawing explanation
Fig. 1 is the student's document file management system workflow block diagram that the present invention is based on image procossing;
Fig. 2 is the color mark interface of the student's document file management system that the present invention is based on image procossing;
Fig. 3 is the hook of the student's document file management system that the present invention is based on image procossing and the data sample of fork;
Fig. 4 is the HOG feature calculation example of the student's document file management system that the present invention is based on image procossing;
Fig. 5 is the LBP feature calculation example of the student's document file management system that the present invention is based on image procossing.
Embodiment
Below in conjunction with drawings and Examples, the invention will be further described
A kind of student's document file management system (as shown in Figure 1) based on image procossing, workflow (step):
The first, scan student's document and obtain image, input system.
The second, file and picture denoising.
Three, from file and picture, mark off per pass topic destination locations.
Four, color classification is carried out to document image pixels, filter out red pixel point, and then find out the symbol that teacher annotates and comments on document.
Five, find the symbol in per pass title field and identify.
Six, by the image of per pass exercise question, and teacher's feedback information of correspondence, stored in database.
Experiment and interpretation of result
Test data of experiment of the present invention derives from the quiz of a Computer Systems Organization course of Shanghai university, and paper size is A4, one side, has 5 road exercise questions, is all the form of question-and-answer problem, and last topic needs to draw.80, paper is collected in current test altogether.The printer model that the present invention adopts is HPLaserJet200ColorMFPM276PCL6, and this printer has the function of batch scanning, therefore can complete the process of paper electronization easily.It can also set the sharpness exporting picture.The present invention adopts most high definition (600dpi) to reach best recognition result, and under this sharpness, the resolution of image is up to 4960 × 7014.During process high definition picture, travelling speed has certain loss, but in reality test of the present invention, the processing time of on average often opening paper is still less than 1 second, and will lack more than the scanning time used, therefore speed is not bottleneck in the system of the present invention.
Pay close attention to the most in evaluation and test of the present invention, the accuracy rate of is-symbol identification module, and what have the greatest impact to it is exactly choosing of feature extraction algorithm, and the setting of SVM parameter.
As previously mentioned, the present invention collects and has marked 386 symbols (as shown in Figure 3) hooking and pitch, and wherein hooking has 196, and fork has 190.Evaluation and test adopts the mode of 5-fold cross validation.Experiment of the present invention compares various features extraction algorithm, has HOG feature, LBP feature, and the self-designed feature of the present invention.The parameter of Linear SVM only has a C value, can regulate the punishment cost of the data point of classification error in training data.For often kind of feature extraction algorithm, there are again some parameters separately.The present invention through attempting the parameter combinations of many group SVM and feature extraction algorithm, and selects one group of parameter that effect is best in 5-fold cross validation as the best result of this feature.
The simple extracting mode of one of HOG feature is as follows: the grid first picture segmentation being become C × C, and each grid comprises K × K pixel (as shown in Figure 4).For the every bit in figure, the present invention can calculate its gradient, namely towards and intensity size.And for each grid, the present invention calculates the gradient histogram in 8 directions of K × K pixel of its inside.Therefore, in each grid, the present invention just can be used as its proper vector by 8 numbers.
The calculating of LBP feature also and uncomplicated.First, the present invention needs grid picture being divided into C × C equally, has K × K pixel in each grid.For each pixel, the present invention calculates the local binary patterns of its position, is coupled together by the pixel value in the surrounding of this pixel 8 grid exactly according to clockwise direction, (as shown in Figure 5), its result is certain integer between [0,255].Note, these are slightly different from the calculating of general significance LBP feature.This is because the singularity of bianry image causes.Next, for single grid, the present invention calculates the frequency that each integer occurs, calculates a histogram thus, as the proper vector of this grid inside.And whole picture comprises C × C grid, the dimension of feature should be just 256 × C × C.
For LBP feature, also having the optimization that one conventional, when losing feature representation ability hardly, intrinsic dimensionality can be reduced.First, definition " equivalent formulations " (UniformPattern), namely at the binary number that each pixel calculates, only comprises the saltus step at the most between twice 01.Statistics shows, the local binary patterns of a secondary picture overwhelming majority is all equivalent formulations, therefore in histogram, gives up all non-equivalence patterns, effectively can reduce intrinsic dimensionality, and drop-out hardly.Achieve this technology in experiment of the present invention, thus the dimension of feature is down to 59 × C × C.
For the feature of the present invention's design, only having a parameter, is exactly the size of grid.Picture is divided into the grid of C × C by the present invention, carries out sue for peace (or being averaging) in each grid, and final output dimension is the feature of C × C.
Different parameter combinations has been attempted in experiment of the present invention, wherein adopts the optimal result of three kinds of features as shown in the table:
| Algorithm | HOG | LBP | Feature of the present invention |
| Grid scale | 3×3 | 3×3 | 16×16 |
| Grid is containing pixel | 48×48 | 48×48 | 9×9 |
| Characteristic length | 3×3×8＝72 | 3×3×59＝531 | 16×16＝256 |
| SVM type | RBF | RBF | Linear |
| SVM C | 10000 | 1000 | 1000 |
| SVM Gamma | 0.1 | 0.1 | Nothing |
| Misclassification data | 2 | 1 | 0 |
| Accuracy rate | 99.7％ | 99.5％ | 100％ |
As seen from the above table, these three kinds of feature extracting methods, it is all very accurate to identify.Wherein adopt HOG misclassification 1 data, LBP misclassification 2 data, and the method that the present invention proposes can accomplish perfect classification.As can be seen here, the feature extraction algorithm of the present invention's design, is highly suitable for this task.
Under accuracy rate high like this, the exercise question that the present invention can automatically help student to do by Accurate classification of correcting errors, this also demonstrates the high reliability of present system.
The present invention, through exploring how to help student-directed paper document by digitizing technique, carries out mistake topic and arranges, and helps the aspects such as the more detailed student's answering information of teacher's acquisition to have following contribution:
The desk study helps of information digitalization means to education and instruction.
Devise the paper digitized mine of simple and effective, and the mistake topic recognizer of intelligence.
For the function that student provides wrong topic to sort out, the leakage detection facilitating them to review the stage is filled a vacancy.
For the distribution of teacher's programming count operation mistake topic, the functions such as students' work follow-up of quality are also provided, thus accurately understand the study situation of each student in real time.
Technology used by system of the present invention is all existing comparatively proven technique mostly.And the present invention is in the process realizing this system, further investigate the technology of various image procossing aspect, understand their characteristic, and in conjunction with the characteristic of problem of the present invention itself, choose technology the most suitable, and achieve them well, therefore at the key index of system---on wrong topic recognition accuracy, reach excellent effect.This also absolutely proves, the present invention has preferably can practicality, if in addition perfect in function and interface, can serve school completely.
The present invention in the future can supporting more advanced reading and apparatus for writing, and examination, operation can be allowed to carry out on equipment completely, and has and supporting correct system, thus whole process realizes information digitalization.More advanced, sophisticated software can be had based on the present invention, allow the statistic of classification of these data become easily accurately, also have abundanter course and exercise resource, expand the approach of student's Gains resources.These are all that the present invention can for educating the help brought.
Claims (3)
1., based on student's document file management system of image procossing, comprise step:
A. scan student's document and obtain image, input system;
B. file and picture denoising;
C. from file and picture, mark off per pass topic destination locations;
D. color classification is carried out to document image pixels, filter out red pixel point, and then find out the symbol that teacher annotates and comments on document;
E. find the symbol in per pass title field according to color classification result and identify;
F. by the image of per pass exercise question, and teacher's feedback information of correspondence, stored in database;
It is characterized in that,
Described a. scans student's document and obtains image, comprises setting:
A.1 file and picture every page is A4 paper size, and what print one page or two pages to every face splits into two single pages after scanning automatically;
A.2. the exercise question on paper arranges from top to bottom, character calligraph order from left to right, have between upper and lower twice exercise question for answer interval or stay white, subfield is not carried out in page inside;
A.3. paper adopts blackness handwriting that is hand-written or printing, and the writing color of student is black or blueness, and teacher reads and makes comments as red stroke handwriting;
A.4. teacher makes hook, fork, passes judgment on per pass exercise question;
A.5. paper changes into digital document images form by color scanner;
Described b. adopts the mode of gaussian filtering, carries out file and picture denoising;
Described c. adopts statistical method, carries out exercise question location recognition.
2., as claimed in claim 1 based on student's document file management system of image procossing, it is characterized in that,
Described a.5. paper allows suitable inclination when scanning, but angle of inclination is no more than 5 degree;
Described b. adopts the mode of gaussian filtering, carries out file and picture denoising, different scanners is optimized, adjusted to the parameter of gaussian filtering in advance;
Described c. adopts statistical method, carries out exercise question location recognition, comprising:
C.1. according to the quantity of black picture element in each pixel column of statistic document image;
C.2. make histogram, and find out wherein peak value, highly close peak value is merged, be considered as same problem object examination question region;
C.3., after merging peak value, find the position of first low ebb in every problem front, be considered as the position that this topic starts;
C.4. hand inspection, amendment title field segmentation result is allowed;
Described d. carries out color classification to the pixel of file and picture on paper, filters out red pixel:
D.1. adopt RGB model, add up in the pixel of all scanning papers, the number of times that each rgb value occurs, get K the rgb value that wherein frequency of occurrences is the highest, repeatedly, whether many people manually mark be red, and all the other are all set to and do not mark, following flow process iteration, until whole color space has been marked;
D.2. to each color be not marked, add up in all adjacent colors, the quantity of red and non-redness;
If d.3. adjacent redness is many, this color is labeled as redness, if adjacent non-redness is many, this color is labeled as non-redness, if both as many, this color is still set to and does not mark;
Described e. finds the symbol in per pass title field and identifies, the method for wherein said employing machine learning builds a sorter.
3. the student's document file management system based on image procossing as described in claim 1,2, is characterized in that,
Described e. finds the symbol in per pass title field and identifies, described construction sorter, comprises following a few step:
E.1. manually produce and mark training data, for the feature extraction of training data, training classifier;
E.2. described feature extraction, adopts the algorithm that bianry image pixel region is average;
E.3. described training classifier, adopts LibSVM.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410208709.0A CN105095892A (en) | 2014-05-16 | 2014-05-16 | Student document management system based on image processing |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410208709.0A CN105095892A (en) | 2014-05-16 | 2014-05-16 | Student document management system based on image processing |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN105095892A true CN105095892A (en) | 2015-11-25 |
Family
ID=54576280
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201410208709.0A Pending CN105095892A (en) | 2014-05-16 | 2014-05-16 | Student document management system based on image processing |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN105095892A (en) |
Cited By (23)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106856006A (en) * | 2016-10-12 | 2017-06-16 | 国家海洋局第海洋研究所 | The digitizing solution of two-dimentional equidistantly perspective view |
| CN107025456A (en) * | 2017-03-20 | 2017-08-08 | 安徽七天教育科技有限公司 | A kind of method of teacher comment vestige automatic identification |
| CN107038913A (en) * | 2017-05-27 | 2017-08-11 | 江苏学正教育科技有限公司 | A kind of autonomous operational method by the intelligent sub title collection of page |
| CN107443927A (en) * | 2017-08-21 | 2017-12-08 | 顾东海 | A kind of wrong topic identification sorter and its method of work |
| CN107506762A (en) * | 2017-09-01 | 2017-12-22 | 昆山中骏博研互联网科技有限公司 | A kind of achievement method for automatically inputting based on graphical analysis |
| CN107798321A (en) * | 2017-12-04 | 2018-03-13 | 海南云江科技有限公司 | A kind of examination paper analysis method and computing device |
| CN107862230A (en) * | 2017-11-14 | 2018-03-30 | 安徽虚境科技有限公司 | A kind of grade determines and input method |
| CN107908752A (en) * | 2017-11-18 | 2018-04-13 | 曲阜师范大学 | A kind of paper achievement intelligent acquisition and analysis system and method |
| CN108881708A (en) * | 2017-12-18 | 2018-11-23 | 南通使爱智能科技有限公司 | A kind of intelligent image processing unit |
| CN110060269A (en) * | 2019-04-22 | 2019-07-26 | 杭州电子科技大学 | A kind of processing flow management method of file and picture |
| CN110059184A (en) * | 2019-03-28 | 2019-07-26 | 莆田学院 | A kind of operation mistake topic collection and analysis method and system |
| CN110135339A (en) * | 2019-05-15 | 2019-08-16 | 宜春宜联科技有限公司 | Operation formulating method and device |
| CN110175616A (en) * | 2019-05-15 | 2019-08-27 | 南京庆峰和智能科技有限公司 | A kind of paper image answer extraction method based on color |
| CN110210309A (en) * | 2019-04-30 | 2019-09-06 | 宜春宜联科技有限公司 | The recognition methods of mistake topic, system, readable storage medium storing program for executing and equipment |
| CN110222678A (en) * | 2019-04-30 | 2019-09-10 | 宜春宜联科技有限公司 | A kind of item analysis method, system, readable storage medium storing program for executing and electronic equipment |
| CN110414537A (en) * | 2019-01-08 | 2019-11-05 | 电子科技大学 | A kind of recognition methods of Highland Barley In Xizang lesion |
| WO2020135472A1 (en) * | 2018-12-26 | 2020-07-02 | 杭州大拿科技股份有限公司 | Method and device for generating collection of incorrectly-answered questions |
| CN111563510A (en) * | 2020-04-30 | 2020-08-21 | 广东小天才科技有限公司 | Image processing method and system |
| CN111768820A (en) * | 2020-06-04 | 2020-10-13 | 上海森亿医疗科技有限公司 | Paper medical record digitization and target detection model training method, device and storage medium |
| CN112308807A (en) * | 2020-10-19 | 2021-02-02 | 泰州锐比特智能科技有限公司 | Subject type field detection system |
| CN113066017A (en) * | 2021-02-27 | 2021-07-02 | 华为技术有限公司 | Image enhancement method, model training method and equipment |
| CN114550191A (en) * | 2022-02-10 | 2022-05-27 | 珠海读书郎软件科技有限公司 | Intelligent correction method and intelligent desk lamp |
| CN115620332A (en) * | 2022-11-17 | 2023-01-17 | 北京十六进制科技有限公司 | Automatic reading and amending method and equipment based on paper operation |
-
2014
- 2014-05-16 CN CN201410208709.0A patent/CN105095892A/en active Pending
Cited By (29)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106856006A (en) * | 2016-10-12 | 2017-06-16 | 国家海洋局第海洋研究所 | The digitizing solution of two-dimentional equidistantly perspective view |
| CN106856006B (en) * | 2016-10-12 | 2019-10-18 | 自然资源部第一海洋研究所 | The digitizing solution of two-dimentional equidistant perspective view |
| CN107025456A (en) * | 2017-03-20 | 2017-08-08 | 安徽七天教育科技有限公司 | A kind of method of teacher comment vestige automatic identification |
| CN107038913A (en) * | 2017-05-27 | 2017-08-11 | 江苏学正教育科技有限公司 | A kind of autonomous operational method by the intelligent sub title collection of page |
| CN107443927A (en) * | 2017-08-21 | 2017-12-08 | 顾东海 | A kind of wrong topic identification sorter and its method of work |
| CN107506762A (en) * | 2017-09-01 | 2017-12-22 | 昆山中骏博研互联网科技有限公司 | A kind of achievement method for automatically inputting based on graphical analysis |
| CN107506762B (en) * | 2017-09-01 | 2020-03-20 | 昆山中骏博研互联网科技有限公司 | Score automatic input method based on image analysis |
| CN107862230A (en) * | 2017-11-14 | 2018-03-30 | 安徽虚境科技有限公司 | A kind of grade determines and input method |
| CN107908752A (en) * | 2017-11-18 | 2018-04-13 | 曲阜师范大学 | A kind of paper achievement intelligent acquisition and analysis system and method |
| CN107798321A (en) * | 2017-12-04 | 2018-03-13 | 海南云江科技有限公司 | A kind of examination paper analysis method and computing device |
| CN108881708A (en) * | 2017-12-18 | 2018-11-23 | 南通使爱智能科技有限公司 | A kind of intelligent image processing unit |
| US11410407B2 (en) | 2018-12-26 | 2022-08-09 | Hangzhou Dana Technology Inc. | Method and device for generating collection of incorrectly-answered questions |
| WO2020135472A1 (en) * | 2018-12-26 | 2020-07-02 | 杭州大拿科技股份有限公司 | Method and device for generating collection of incorrectly-answered questions |
| CN110414537A (en) * | 2019-01-08 | 2019-11-05 | 电子科技大学 | A kind of recognition methods of Highland Barley In Xizang lesion |
| CN110059184B (en) * | 2019-03-28 | 2022-03-08 | 莆田学院 | Operation error collection and analysis method and system |
| CN110059184A (en) * | 2019-03-28 | 2019-07-26 | 莆田学院 | A kind of operation mistake topic collection and analysis method and system |
| CN110060269A (en) * | 2019-04-22 | 2019-07-26 | 杭州电子科技大学 | A kind of processing flow management method of file and picture |
| CN110222678A (en) * | 2019-04-30 | 2019-09-10 | 宜春宜联科技有限公司 | A kind of item analysis method, system, readable storage medium storing program for executing and electronic equipment |
| CN110210309A (en) * | 2019-04-30 | 2019-09-06 | 宜春宜联科技有限公司 | The recognition methods of mistake topic, system, readable storage medium storing program for executing and equipment |
| CN110222678B (en) * | 2019-04-30 | 2022-02-01 | 宜春宜联科技有限公司 | Topic analysis method, system, readable storage medium and electronic device |
| CN110175616A (en) * | 2019-05-15 | 2019-08-27 | 南京庆峰和智能科技有限公司 | A kind of paper image answer extraction method based on color |
| CN110135339A (en) * | 2019-05-15 | 2019-08-16 | 宜春宜联科技有限公司 | Operation formulating method and device |
| CN111563510A (en) * | 2020-04-30 | 2020-08-21 | 广东小天才科技有限公司 | Image processing method and system |
| CN111563510B (en) * | 2020-04-30 | 2023-04-18 | 广东小天才科技有限公司 | Image processing method and system |
| CN111768820A (en) * | 2020-06-04 | 2020-10-13 | 上海森亿医疗科技有限公司 | Paper medical record digitization and target detection model training method, device and storage medium |
| CN112308807A (en) * | 2020-10-19 | 2021-02-02 | 泰州锐比特智能科技有限公司 | Subject type field detection system |
| CN113066017A (en) * | 2021-02-27 | 2021-07-02 | 华为技术有限公司 | Image enhancement method, model training method and equipment |
| CN114550191A (en) * | 2022-02-10 | 2022-05-27 | 珠海读书郎软件科技有限公司 | Intelligent correction method and intelligent desk lamp |
| CN115620332A (en) * | 2022-11-17 | 2023-01-17 | 北京十六进制科技有限公司 | Automatic reading and amending method and equipment based on paper operation |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN105095892A (en) | Student document management system based on image processing | |
| CN103971126B (en) | A kind of traffic sign recognition method and device | |
| CN103098074B (en) | Document page segmentation in optical character recognition | |
| TWI467515B (en) | Multi-color dropout for scanned document | |
| CN106096610B (en) | A kind of file and picture binary coding method based on support vector machines | |
| CN102855495B (en) | Method for implementing electronic edition standard answer, and application system thereof | |
| CN102592126B (en) | For the method for binaryzation scanning document image | |
| Mitianoudis et al. | Document image binarization using local features and Gaussian mixture modeling | |
| CN104408449B (en) | Intelligent mobile terminal scene literal processing method | |
| CN110210413A (en) | A kind of multidisciplinary paper content detection based on deep learning and identifying system and method | |
| CN103310211B (en) | A kind ofly fill in mark recognition method based on image procossing | |
| CN110597806A (en) | Wrong question set generation and answer statistics system and method based on reading and amending identification | |
| CN101599125A (en) | The binarization method that the complex background hypograph is handled | |
| CN104077577A (en) | Trademark detection method based on convolutional neural network | |
| CN109255344A (en) | A kind of digital display instrument positioning and Recognition of Reading method based on machine vision | |
| CN109740572A (en) | A kind of human face in-vivo detection method based on partial color textural characteristics | |
| CN104680130A (en) | Chinese character recognition method for identification cards | |
| CN104463138B (en) | The text positioning method and system of view-based access control model structure attribute | |
| US7564587B2 (en) | Method of scoring a printed form having targets to be marked | |
| CN106934386A (en) | A kind of natural scene character detecting method and system based on from heuristic strategies | |
| CN101615244A (en) | Handwritten plate blank numbers automatic identifying method and recognition device | |
| CN104794479A (en) | Method for detecting text in natural scene picture based on local width change of strokes | |
| CN112434699A (en) | Automatic extraction and intelligent scoring system for handwritten Chinese characters or components and strokes | |
| CN112446259A (en) | Image processing method, device, terminal and computer readable storage medium | |
| CN109360179A (en) | A kind of image interfusion method, device and readable storage medium storing program for executing |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| WD01 | Invention patent application deemed withdrawn after publication | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20151125 |