CN107729936B - Automatic error correction review method and system - Google Patents
Automatic error correction review method and system Download PDFInfo
- Publication number
- CN107729936B CN107729936B CN201710947823.9A CN201710947823A CN107729936B CN 107729936 B CN107729936 B CN 107729936B CN 201710947823 A CN201710947823 A CN 201710947823A CN 107729936 B CN107729936 B CN 107729936B
- Authority
- CN
- China
- Prior art keywords
- image
- answer
- point
- character
- handwritten character
- 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.)
- Active
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V30/00—Character recognition; Recognising digital ink; Document-oriented image-based pattern recognition
- G06V30/10—Character recognition
- G06V30/14—Image acquisition
- G06V30/148—Segmentation of character regions
- G06V30/153—Segmentation of character regions using recognition of characters or words
Abstract
The embodiment of the invention provides a method and a device for automatically evaluating a corrected error question, belonging to the field of image processing. The method comprises the following steps: acquiring a single-point answer image corresponding to each written answer on a medium to be read; carrying out image recognition on each acquired single-point answer image by using a preset image detection method and/or based on a preset recognition model to obtain a recognition result; and obtaining corresponding evaluation characteristics according to the identification information, and evaluating each written answer according to the evaluation characteristics based on a preset evaluation model. The method of the embodiment can overcome the technical problem that the answer image is inaccurate in identification due to the complexity of the wrong answer mode, and therefore the method has the beneficial effect of quickly and accurately evaluating the wrong answer.
Description
Technical Field
The invention relates to the technical field of image information processing, in particular to a method and a system for automatically evaluating error correction questions.
Background
In the current language examination, the (short text) correction problem belongs to the objective examination of subjective examination problems. The basic knowledge of all aspects of the answerer is examined, and the capability of the answerer for comprehensively utilizing the language is also examined. From the aspect of setting questions, the questions examine the ability of the testee to comprehensively use the corresponding languages in the language passages from three aspects of morphology, syntax and literary logic; temporal, morphological, lexical, predicate verbs, nouns, articles, adjectives, conjunctions, pronouns, and conjunctions of sentences are examined in detail. The answerer answers the test question in a manner of adding, deleting or modifying characters in the test question in a handwriting manner.
In recent years, with the rapid development of computer technology and information technology, especially the rapid advance of artificial intelligence technology, the use of artificial intelligence to replace traditional human being has become a hot spot direction in all industries. The examination paper reading mode in the educational field to be tested is also realized by traditional purely manual evaluation, and gradually changes into the mode that manual intelligence replaces partial manual work to realize automatic evaluation. The existing mainstream test paper can be automatically evaluated, and besides practical objective questions can be realized at first, some products with advanced technical development stations in the front of the industry can be automatically evaluated, and composition and blank filling questions can be automatically evaluated.
However, the problem of error correction is limited by the complicated and diversified handwriting in answering by the answerers (for example, the 'increased' handwriting mark 'Λ' can generate various handwriting patterns according to different answerers), so the error correction still needs manual review by the answerers. Particularly, in large-scale examinations such as high and middle level examinations and language level examinations, the manual review mode consumes time and labor, and brings great working pressure to the reviewers; in addition, the manual review is also easily affected by the subjectivity of the review personnel, so that the review result is unfair and positive.
In view of the foregoing, it is desirable to provide an automatic review scheme for correcting the error problems in the prior art.
Disclosure of Invention
In order to solve the above problems, embodiments of the present invention provide a method and an apparatus for automatically reviewing a wrong question that overcome the above problems or at least partially solve the above problems.
According to a first aspect of the embodiments of the present invention, there is provided a method for automatically reviewing a problem, the method including:
acquiring a single-point answer image corresponding to each written answer on a medium to be read;
carrying out image recognition on each acquired single-point answer image by using a preset image detection method and/or based on a preset recognition model to obtain a recognition result;
and obtaining corresponding evaluation characteristics according to the identification information, and evaluating each written answer according to the evaluation characteristics based on a preset evaluation model.
According to another aspect of the embodiments of the present invention, there is provided an automatic review device for error correction questions, the device including:
the acquisition module is used for acquiring single-point answer images corresponding to all written answers on a medium to be answered;
the identification module is used for carrying out image identification on each acquired single-point answer image by using a preset image detection method and/or based on a preset identification model to obtain an identification result;
and the evaluation module is used for obtaining corresponding evaluation characteristics according to the identification information and evaluating each written answer according to the evaluation characteristics based on a preset evaluation model.
According to a third aspect of the embodiments of the present invention, there is provided an electronic device for error correction automatic review, including:
at least one processor; and
at least one memory communicatively coupled to the processor, wherein:
the memory stores program instructions executable by the processor, and the processor calls the program instructions to execute the automatic error correction review method provided by any one of the various possible implementations of the first aspect.
According to a fourth aspect of the present invention, there is provided a non-transitory computer readable storage medium storing computer instructions for causing a computer to execute the automatic error correction review method provided in any one of the various possible implementations of the first aspect.
The embodiment of the invention provides a method and a device for automatically evaluating wrong-answer questions, which can overcome the technical problem that the identification of the image information of the wrong-answer questions is inaccurate due to the complexity of the wrong-answer mode, thereby having the beneficial effect of automatically, quickly and accurately evaluating the wrong-answer questions.
Drawings
FIG. 1 is a schematic flow chart illustrating an automatic review method for error correction questions according to an embodiment of the present invention;
FIG. 2 is a schematic diagram illustrating processing of an image of a to-be-reviewed wrong question in an automatic wrong question reviewing method according to an embodiment of the present invention;
fig. 3 is a schematic structural diagram of a specific pre-constructed restricted network in an automatic error correction review method according to an embodiment of the present invention;
FIG. 4 is a block diagram of an automatic review device for error correction questions according to an embodiment of the present invention;
fig. 5 is a block diagram of an electronic device for automatic error correction review according to an embodiment of the present invention.
Detailed Description
The following detailed description of embodiments of the present invention is provided in connection with the accompanying drawings and examples. The following examples are intended to illustrate the invention but are not intended to limit the scope of the invention.
With the development of image recognition technology, the application field of the image recognition technology is wider and wider. Currently, image recognition technology has been applied in the field of education. In the prior art, methods for evaluating partial test question scores by using image recognition technology exist. However, in the language type examination, the (short text) wrong answer is limited by the diversity and individuation of various handwriting of the answerer due to the complexity of the answer rules (including addition, deletion, modification, etc.), and the prior art does not have the technical scheme of automatic examination paper for the question type.
In view of the above situation, the embodiment of the present invention provides an automatic review method for error correction. The method is suitable for evaluating the language type error (short text error) question types; the method can be applied to an intelligent device or system with an image acquisition and recognition function, and the embodiment of the invention is not particularly limited in this respect. For convenience of description, the embodiment of the present invention takes an execution subject as an example of an intelligent device. As shown in fig. 1, the method includes: s1, acquiring single-point answer images corresponding to the written answers on the medium to be answered; s2, performing image recognition on each acquired single-point answer image by using a preset image detection method and/or based on a preset recognition model to obtain a recognition result; s3, obtaining corresponding review features according to the identification information, and reviewing each written answer based on a preset review model according to the review features; the identification information includes the identification result and identification intermediate process information.
In the foregoing embodiment, the single-point answer image corresponding to each written answer on the medium to be answered in S1 is an image including a handwritten answer of each answerer, and the single-point answer image may be obtained by taking a picture, scanning, or directly importing an image file that is taken in advance, which is not limited in this embodiment of the present invention.
Further, when the correct answer is deleted in S2, it is possible to further determine whether the single-point answer image has a corresponding deletion answer by performing image detection on the single-point answer image; when the correct answer is modified or added, the whole character string of each single-point answer image can be identified based on a pre-generated neural network identification model, and an identification result is obtained; or the single handwritten character can be segmented for each single-point answer image, and each sub-image containing each single handwritten character after segmentation is identified to obtain an identification result.
Further, the S3 obtains the review result of each written answer based on the neural network review model trained in advance according to the recognition result obtained in S2.
On the basis of any of the above embodiments of the present invention, there is provided an automatic wrong answer review method, where the step of obtaining a single-point answer image corresponding to each written answer on a medium to be reviewed includes: acquiring a single-point answer image corresponding to each written answer on a medium to be read according to an answer image of the medium to be read and an answer medium template corresponding to the medium to be read; the answer medium template comprises position information of each character in a test question to be answered and at least one single-point answer area preset according to a correct answer, wherein the single-point answer image is obtained according to the corresponding single-point answer area.
In the above specific embodiment, the obtaining of the answer image of the medium to be answered may be performed by taking a picture, scanning, or directly importing a pre-shot image file, which is not specifically limited in this embodiment of the present invention. Further, the answer image of the medium to be read is obtained by segmenting the layout of the answer template (which can be the answer sheet of the medium to be read, or other answer media containing the question characters, such as the answer paper of the medium to be read, the electronic answer text file, and the like) of the medium to be read. The corrected wrong-question image may include the whole answer image of the medium to be answered, or may include a part of the answer image of the medium to be answered, and the embodiment of the present invention is not specifically limited herein.
Further, in this embodiment of the present invention, since the answer image of the medium to be reviewed inevitably includes both the question printed character image in the wrong question to be reviewed and the handwritten answer information image in which the answerer answers. Therefore, in the embodiment, the template of the test questions to be evaluated is generated in advance, and when the template is set, corresponding position regions are selected correspondingly based on preset rules according to different answering modes of wrong answers, so that the template can include the selected position regions of each correct answer (namely each single-point answer image) in the questions. In this embodiment, each single-point answer image is obtained from the answer image of the medium to be answered by using the selected position region in the template.
Further, as shown in fig. 2, an answer form is preset for a correction type of a correction question, for example, a college entrance examination english correction, and the correction type includes modification, deletion, and addition. Wherein, the modification answering mode is that a horizontal line is marked below the word to be modified and corresponding correct answers are written (as shown in the figure by printing characters parent, crowed, on, hooks, where and telling and a and corresponding answering images); the deleted answer mode is to mark lines on the words to be deleted to indicate deletion (as shown by a printed font try and a corresponding answer image in the figure); the answer mode added is that a hand-written mark 'Λ' is arranged below the position of the two character strings to be added and the added answer is written (as shown in the figure, the printed fonts between the perfect and the her and the corresponding answer images thereof).
Next, the process of generating the template of the examination questions to be evaluated, which is generated in advance in this embodiment, will be further described. In fig. 2, firstly, horizontal projection is performed on the error correction test question image without answering, so as to implement horizontal segmentation of the test question image; and then, vertically projecting each line of test question image to realize the segmentation of single characters in the test questions, and obtaining the position information of each character string in the error-corrected test questions.
Further, the single-point answering area is as follows: when the correct answer is deletion, the area occupied by the character string to be deleted in the test paper is increased by the width of the preset proportion or the preset length, and the area surrounded by the height of the preset proportion or the preset length is increased. And when the correct answer is modification, a preset area below the printing character string to be modified in the test paper. And when the correct answer is addition, a preset area below the middle of the two printed character strings is to be added in the test paper.
Preferably, the selected position area of each answer image in the embodiment is specifically: and selecting a position area of the answer image for deleting the answer type, wherein the area occupied by the printing character string to be deleted in the question is an area surrounded by left and right expansion 1/n width and up and down expansion 1/m height, n is more than or equal to 1, and m is more than or equal to 1. Preferably, the selected position region of the answer image for modifying the answer type is a region surrounded by word answers left and right expanded 1/2 width and line spacing below the character string to be modified in the question. Preferably, the answer image selection position region to which the answer type is added is a region surrounded by line spaces, wherein the width of the region is the width of the widest character obtained based on statistics below two printing character strings to be added in the question.
On the basis of any of the above embodiments of the present invention, an automatic wrong answer review method is provided, where image recognition is performed on each acquired image of a single-point answer by using a preset image detection method and/or based on a preset recognition model to obtain a recognition result, where the method includes:
when the correct answer is deleted, whether at least one deletion mark exists in the corresponding single-point answer image or not is detected, and an identification result of whether a deletion answer exists or not is obtained;
when the correct answer is modified or added, performing single handwritten character segmentation on the corresponding single-point answer image based on a pre-constructed limited network corresponding to the correct answer to obtain each sub-image of the single-point answer image; and identifying each sub-image based on a pre-constructed error correction recognition model to obtain the recognition result of a single handwritten character contained in each sub-image.
In the above embodiment, for the deleted answer mode, it may be determined whether the answer point is answered correctly by detecting whether at least one deletion mark exists on the printed characters in the single-point answer image. The specific and simple scheme is to detect the coincidence degree of all pixel points in each single-point answer image after the target printing character string in the test question in the pre-generated template of the test question to be evaluated and the answer of the test question are deleted if the number or the proportion of the non-coincident pixel points is greater than a set threshold value. Further preferably, regarding the detection of the deleted answer, other horizontal line/oblique line detection methods existing in the prior art may also be adopted, which are not described herein again.
Further, when the single-point answer image with the correct answer being modified or added is detected, the single-point answer image is required to be used as input, and the target single-point answer image is identified based on a pre-trained wrong answer identification model to obtain a corresponding identification result. In this embodiment of the present invention, specifically, in order to more accurately recognize the target answer image, the whole character string in the target single-point answer image needs to be segmented into the single handwritten character image. In this embodiment, a limited network established for the handwritten character strings in the correct answers of each answer sample is used to traverse the target single-point answer image, and obtain sub-images of each single handwritten character in the target single-point answer image. And taking the handwritten character subimages of the single-point characters in the target single-point answer image as input, identifying the target single-point answer image based on a wrong-answer-modifying recognition model obtained by pre-training, and obtaining the probability value that each single handwritten character in the target single-point answer image is each character, wherein the corresponding character with the largest probability value is the final recognition result.
On the basis of any of the above embodiments of the present invention, there is provided an automatic review method for error correction, wherein the pre-constructed restricted network is pre-constructed according to the correct answer corresponding thereto, and comprises a starting layer, an ending layer, each single handwritten character subunit layer, and an absorption layer existing between the foregoing layers:
each single handwritten character subunit layer is constructed according to the handwritten sample image of each single character in the corresponding correct answer;
and front jump arcs and back jump arcs exist from each absorption layer to each single handwritten character subunit layer and each single handwritten character subunit layer.
In the above specific embodiment, each level in the restricted network has a score value and a jump score value belonging to the layer when jumping to each level (generally, there is a penalty score belonging to the absorption layer, and there is a matching score belonging to the character subunit). The absorption layer includes fil (noise absorption) and sil (silent sound absorption) nodes.
Further, since the correct answer of the test paper to be reviewed is known, that is, it is known in advance which character string in the question needs to be modified correspondingly to which character string and which character string needs to be added between which two character strings, the target single-point answer image can be processed by using the known correct answer information, that is, the target single-point answer image can be subjected to the segmentation of a single handwritten character through a pre-constructed restricted network, so that each sub-image containing the single handwritten character is obtained.
Specifically, the construction of the pre-constructed restricted network is divided into the following steps. The restricted network requires a restricted network to be constructed for each corresponding answer point character string. As shown in fig. 3, the limiting network includes a start layer S, an end layer E, each single handwritten character subunit layer ai in the character, and an absorbing layer FS between all the layers, and each absorbing layer has a jump arc (i.e. there is a forward jump and a backward jump) to all the single handwritten character subunits and end layers, and has a score value and a jump score value belonging to the layer when jumping to each layer (generally, there is a penalty score belonging to the absorbing layer, and there is a matching score belonging to the character subunit). And each single handwritten character subunit layer is constructed according to the handwritten sample image of each single character in the correct answer of the corresponding answer point.
On the basis of any of the above embodiments of the present invention, an automatic wrong answer review method is provided, where the method includes performing single handwritten character segmentation on a corresponding single-point answer image based on a pre-constructed restricted network corresponding to a correct answer to obtain each sub-image of the single-point answer image, and includes:
traversing the single-point answer image by using a window with preset pixel width and single-point answer image height along the x-axis direction, and inputting pixels obtained by traversing into a pre-constructed limited network corresponding to correct answers to obtain each sub-image of the single-point answer image.
In the above-mentioned embodiment, based on the network shown in fig. 3 (the image recognition model in which a1, a2, a3, and a4 in fig. 3 correspond to "that" four characters respectively), it is assumed that the target answer image "that" is divided into individual characters: traversing the target single-point answer image 'that' by using a window with preset pixel width (for example, 5 pixel width) and the height of the target single-point answer image along the x-axis direction, and inputting pixels obtained by traversing into a pre-constructed limited network corresponding to the target single-point answer image 'that'; decoding the currently received pixels in a left-to-right manner in fig. 3 using the single handwritten character subunit layers of a1 to a4, respectively, and if the first single handwritten character is confirmed to be "t", then it is recognized by a 1; if the second single handwritten character "h" is successfully recognized by a2, there is no jump between a1 and a 2. Assuming that the target single point is "tthat", when the first single handwritten character is indeed recognized by a1, a1 performs self-skipping because the second single handwritten character is still "t". Assuming that the target single point is "txhat", when the first single handwritten character is identified by a1, a1 jumps ahead to an absorption layer FS before a1 because the second single handwritten character is not any one of the four characters in "that"; further assuming that the target single point is "tat", when the first single handwritten character is indeed recognized by a1, a1 jumps back to a3 because the second single handwritten character is "a".
With the above pre-constructed constrained network, the following conclusions can be drawn: the proportion of the character subunits of the forward skip or the self skip in the total number of the character subunits in the single-point answer image is in direct proportion to the number of newly added characters of correct answers; the proportion of the character subunits of the jump after occurrence to the total number of the character subunits in the single-point answer image is in direct proportion to the number of characters missed by the correct answer; the number of pixels absorbed by the absorption layer of pixels obtained by traversing the target single-point answer image by using a window with a preset pixel width along the x-axis direction accounts for the proportion of the total number of the pixels of the single-point answer image, and the proportion is also in direct proportion to the number of wrong characters in the target single-point answer image.
On the basis of any one of the above embodiments of the present invention, an automatic review method for error correction is provided, wherein the error correction recognition model is obtained by training through the following steps:
carrying out single handwritten character segmentation on the sample single-point answer image to obtain each sample sub-image of the sample single-point answer image;
labeling the recognition result of each sample sub-image;
and training to obtain the error correction recognition model by utilizing a predetermined neural network topological structure and a training algorithm according to each sample subimage.
In the above specific embodiment, the method for specifically constructing the pre-constructed error correction problem identification model is as follows: firstly, collecting modified and added answer image data, and labeling the answer image data (including segmentation of a handwritten character image of an answer sheet and labeling of a recognition result of each handwritten character subunit); then determining a topological structure of the error-correcting recognition model (only by adopting the existing common neural network, such as DNN, CNN and the like); and finally, training the error-correcting recognition model by adopting the existing training algorithm (such as BP algorithm) based on the training data and the determined topological structure.
On the basis of any one of the above embodiments of the present invention, there is provided an automatic review method for revising wrong questions, where the method includes obtaining corresponding review features according to identification information, and reviewing each written answer based on a preset review model according to the review features, and includes: when the correct answer to be identified is deleted, acquiring a corresponding evaluation result according to whether the single-point answer image has the identification result of deleting the answer;
when the correct answer to be identified is modified or added, obtaining at least one item of review characteristics of the single-point answer image based on a preset review method according to the skip information of the single handwritten character segmentation of each sub-image of the corresponding single-point answer image and the probability that each sub-image contains a single handwritten character as a corresponding character; and obtaining the evaluation result of the target single-point answer image based on a pre-constructed evaluation model according to the evaluation characteristics.
In the above embodiment, the entire wrong-answer review result is given based on the review of each character sub-sample image in all the target single-point answer images, such as the average value, linear weighting, and the like of the review of all the target single-point answer images.
Further, obtaining the evaluation result of the target single-point answer image based on a pre-constructed evaluation model according to the evaluation characteristics comprises: and taking the evaluation characteristics as the input of the pre-constructed evaluation model to obtain evaluation results, such as point values and the like, of the target single-point answer image.
On the basis of any one of the above embodiments of the present invention, there is provided an automatic review method for revising wrong questions, where the method includes obtaining corresponding review features according to identification information, and reviewing each written answer based on a preset review model according to the review features, and includes: when the correct answer to be identified is deleted, acquiring a corresponding evaluation result according to whether the single-point answer image has the identification result of deleting the answer;
when the correct answer to be identified is modified or added, obtaining at least one item of review characteristics of the single-point answer image based on a preset review method according to the skip information of the single handwritten character segmentation of each sub-image of the corresponding single-point answer image and the probability that each sub-image contains a single handwritten character as a corresponding character; and obtaining the evaluation result of the target single-point answer image based on a pre-constructed evaluation model according to the evaluation characteristics.
On the basis of any one of the above embodiments of the present invention, an automatic wrong-answer-correction review method is provided, where the review features at least include an answer recognition posterior probability feature and a coverage feature of a character segmentation decoding path, and may further include at least one of an increased character proportion feature, a missed character proportion feature, and an absorbed pixel proportion feature; wherein:
the answer recognition posterior probability characteristic is an average value of probabilities that a single handwritten character contained in the single-point answer image is a corresponding character;
the coverage characteristic of the character segmentation decoding path is the ratio of the number of traversed single handwritten characters to the number of handwritten character subunit layers in the limited network when the single-point answer image is subjected to single handwritten character segmentation;
the character increasing character proportion characteristic is the proportion of the character subunits of the forward jumping or the self jumping in the single-point answer image to the total number of the character subunits in the single-point answer image when the single-point answer image is subjected to single handwritten character cutting;
the character missing ratio characteristic is the ratio of the back skip character subunit to the total number of the character subunits in the single-point answer image when the single-point answer image is subjected to single handwritten character segmentation;
the absorbed pixel proportion characteristic is the proportion of the number of pixels absorbed by the absorbing layer of pixels obtained by traversing the target single-point answer image by using a window with a preset pixel width along the x-axis direction when the single-point answer image is subjected to single handwritten character segmentation, and the pixel proportion accounts for the total number of the pixel points of the single-point answer image.
As shown in fig. 4, on the basis of any of the above embodiments of the present invention, there is provided an automatic review device for error correction, including:
the obtaining module A1 is used for obtaining single-point answer images corresponding to all written answers on the medium to be answered; the identification module A2 is used for carrying out image identification on each acquired single-point answer image by using a preset image detection method and/or based on a preset identification model to obtain an identification result; the evaluation module A3 is used for obtaining corresponding evaluation characteristics according to the identification information and evaluating each written answer according to the evaluation characteristics based on a preset evaluation model; the identification information includes the identification result and identification intermediate process information.
On the basis of any one of the above embodiments of the present invention, there is provided an automatic review device for error correction, wherein the obtaining module is further configured to:
acquiring a single-point answer image corresponding to each written answer on a medium to be read according to an answer image of the medium to be read and an answer medium template corresponding to the medium to be read; the answer medium template comprises position information of each character in a test question to be answered and at least one single-point answer area preset according to a correct answer, wherein the single-point answer image is obtained according to the corresponding single-point answer area.
On the basis of any one of the above embodiments of the present invention, an automatic review device for error correction is provided, where the identification module is further configured to:
when the correct answer is deleted, whether at least one deletion mark exists in the corresponding single-point answer image or not is detected, and an identification result of whether a deletion answer exists or not is obtained;
when the correct answer is modified or added, performing single handwritten character segmentation on the corresponding single-point answer image based on a pre-constructed limited network corresponding to the correct answer to obtain each sub-image of the single-point answer image; and identifying each sub-image based on a pre-constructed error correction recognition model to obtain the recognition result of a single handwritten character contained in each sub-image.
On the basis of any one of the above embodiments of the present invention, there is provided an automatic review device for error correction, wherein the pre-constructed restricted network is pre-constructed according to the correct answer corresponding thereto, and comprises a starting layer, an ending layer, each single handwritten character subunit layer, and an absorption layer existing between the foregoing layers:
each single handwritten character subunit layer is constructed according to the handwritten sample image of each single character in the corresponding correct answer;
and front jump arcs and back jump arcs exist from each absorption layer to each single handwritten character subunit layer and each single handwritten character subunit layer.
On the basis of any one of the above embodiments of the present invention, an automatic review device for error correction is provided, where the identification module is further configured to:
traversing the single-point answer image by using a window with preset pixel width and single-point answer image height along the x-axis direction, and inputting pixels obtained by traversing into a pre-constructed limited network corresponding to correct answers to obtain each sub-image of the single-point answer image.
On the basis of any one of the above embodiments of the present invention, an automatic review device for error correction questions is provided, wherein the error correction question recognition model is obtained by training through the following steps:
carrying out single handwritten character segmentation on the sample single-point answer image to obtain each sample sub-image of the sample single-point answer image;
labeling the recognition result of each sample sub-image;
and training to obtain the error correction recognition model by utilizing a predetermined neural network topological structure and a training algorithm according to each sample subimage.
On the basis of any one of the above embodiments of the present invention, an automatic review device for error correction is provided, wherein the review module is further configured to:
when the correct answer to be identified is deleted, acquiring a corresponding evaluation result according to whether the single-point answer image has the identification result of deleting the answer;
when the correct answer to be identified is modified or added, obtaining at least one item of review characteristics of the single-point answer image based on a preset review method according to the skip information of the single handwritten character segmentation of each sub-image of the corresponding single-point answer image and the probability that each sub-image contains a single handwritten character as a corresponding character; and obtaining the evaluation result of the target single-point answer image based on a pre-constructed evaluation model according to the evaluation characteristics.
On the basis of any one of the above embodiments of the present invention, an automatic wrong answer review device is provided, where the review features at least include an answer recognition posterior probability feature and a coverage feature of a character segmentation decoding path, and may further include at least one of an increased character proportion feature, a missed character proportion feature, and an absorbed pixel proportion feature; wherein:
the answer recognition posterior probability characteristic is an average value of probabilities that a single handwritten character contained in the single-point answer image is a corresponding character;
the coverage characteristic of the character segmentation decoding path is the ratio of the number of traversed single handwritten characters to the number of handwritten character subunit layers in the limited network when the single-point answer image is subjected to single handwritten character segmentation;
the character increasing character proportion characteristic is the proportion of the character subunits of the forward jumping or the self jumping in the single-point answer image to the total number of the character subunits in the single-point answer image when the single-point answer image is subjected to single handwritten character cutting;
the character missing ratio characteristic is the ratio of the back skip character subunit to the total number of the character subunits in the single-point answer image when the single-point answer image is subjected to single handwritten character segmentation;
the absorbed pixel proportion characteristic is the proportion of the number of pixels absorbed by the absorbing layer of pixels obtained by traversing the target single-point answer image by using a window with a preset pixel width along the x-axis direction when the single-point answer image is subjected to single handwritten character segmentation, and the pixel proportion accounts for the total number of the pixel points of the single-point answer image.
On the basis of any one of the above embodiments of the present invention, an electronic device for automatic review of error problems is provided. Referring to fig. 5, the error correction question automatic review electronic device includes: a processor (processor)501, a memory (memory)502, and a bus 503;
the processor 501 and the memory 502 respectively complete communication with each other through a bus 503;
the processor 501 is used for calling the program instructions in the memory 502 to execute the error correction question automatic review method provided by the above embodiment, for example, including: acquiring a single-point answer image corresponding to each written answer on a medium to be read; carrying out image recognition on each acquired single-point answer image by using a preset image detection method and/or based on a preset recognition model to obtain a recognition result; and obtaining corresponding evaluation characteristics according to the identification information, and evaluating each written answer according to the evaluation characteristics based on a preset evaluation model.
An embodiment of the present invention provides a non-transitory computer-readable storage medium, where the non-transitory computer-readable storage medium stores computer instructions, and the computer instructions enable a computer to execute the error correction question automatic review method provided in the foregoing embodiment, for example, the method includes: acquiring a single-point answer image corresponding to each written answer on a medium to be read; carrying out image recognition on each acquired single-point answer image by using a preset image detection method and/or based on a preset recognition model to obtain a recognition result; and obtaining corresponding evaluation characteristics according to the identification information, and evaluating each written answer according to the evaluation characteristics based on a preset evaluation model.
Those of ordinary skill in the art will understand that: all or part of the steps for implementing the method embodiments may be implemented by hardware related to program instructions, and the program may be stored in a computer readable storage medium, and when executed, the program performs the steps including the method embodiments; and the aforementioned storage medium includes: various media that can store program codes, such as ROM, RAM, magnetic or optical disks.
The above-described embodiments of the information interaction device and the like are merely illustrative, where units illustrated as separate components may or may not be physically separate, and components displayed as units may or may not be physical units, may be located in one place, or may also be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of the present embodiment. One of ordinary skill in the art can understand and implement it without inventive effort.
Through the above description of the embodiments, those skilled in the art will clearly understand that each embodiment can be implemented by software plus a necessary general hardware platform, and certainly can also be implemented by hardware. With this understanding in mind, the above technical solutions may be embodied in the form of a software product, which can be stored in a computer-readable storage medium, such as ROM/RAM, magnetic disk, optical disk, etc., and includes instructions for causing a computer device (which may be a personal computer, a server, or a network device) to execute the various embodiments or some parts of the methods of the embodiments.
Finally, the method of the present application is only a preferred embodiment and is not intended to limit the scope of the embodiments of the present invention. Any modification, equivalent replacement, improvement and the like made within the spirit and principle of the embodiments of the present invention should be included in the protection scope of the embodiments of the present invention.
Claims (14)
1. An automatic review method for error problems is characterized by comprising the following steps:
acquiring a single-point answer image corresponding to each written answer on a medium to be read;
carrying out image recognition on each acquired single-point answer image by using a preset image detection method and/or based on a preset recognition model to obtain a recognition result;
obtaining corresponding evaluation characteristics according to the identification information, and evaluating each written answer based on a preset evaluation model according to the evaluation characteristics; the identification information comprises the identification result and identification intermediate process information;
the wrong answer type comprises at least one of deletion, modification and addition;
the image recognition of the acquired single-point answer images by using a preset image detection method and/or based on a preset recognition model to obtain a recognition result specifically comprises the following steps:
when the correct answer is deleted, judging whether the single-point answer image has a corresponding deletion answer or not by carrying out image detection on the single-point answer image;
when the correct answer is modified or added, firstly segmenting each single-point answer image by using a single handwritten character, and identifying each sub-image which contains each single handwritten character after segmentation to obtain an identification result;
the method comprises the following steps of firstly segmenting each single-point answer image by using a single handwritten character, identifying each subimage containing each single handwritten character after segmentation, and obtaining an identification result, wherein the identification result comprises the following steps:
performing single handwritten character segmentation on a corresponding single-point answer image based on a pre-constructed restricted network corresponding to a correct answer to obtain each sub-image of the single-point answer image; identifying each sub-image based on a pre-constructed error correction recognition model to obtain a recognition result of a single handwritten character contained in each sub-image; the limited network is established aiming at the handwritten character strings in the correct answers of all answer samples;
the step of obtaining the single-point answer image corresponding to each written answer on the medium to be answered comprises the following steps:
acquiring a single-point answer image corresponding to each written answer on a medium to be read according to an answer image of the medium to be read and an answer medium template corresponding to the medium to be read; the answer medium template comprises position information of each character in a test question to be answered and at least one single-point answer area preset according to a correct answer, wherein the single-point answer image is obtained according to the corresponding single-point answer area.
2. The method of claim 1, wherein the pre-constructed restricted network is pre-constructed according to its corresponding correct answer, and comprises a starting layer, an ending layer, each single handwritten character subunit layer, and an absorbing layer existing between the preceding layers:
each single handwritten character subunit layer is constructed according to the handwritten sample image of each single character in the corresponding correct answer;
and front jump arcs and back jump arcs exist from each absorption layer to each single handwritten character subunit layer and each single handwritten character subunit layer.
3. The method according to claim 2, wherein the obtaining of each sub-image of the single-point answer image by performing single handwritten character segmentation on the corresponding single-point answer image based on a pre-constructed restricted network corresponding to the correct answer comprises:
traversing the single-point answer image by using a window with preset pixel width and single-point answer image height along the x-axis direction, and inputting pixels obtained by traversing into a pre-constructed limited network corresponding to correct answers to obtain each sub-image of the single-point answer image.
4. The method of claim 1, wherein the error correction problem recognition model is obtained by training through the following steps:
carrying out single handwritten character segmentation on the sample single-point answer image to obtain each sample sub-image of the sample single-point answer image;
labeling the recognition result of each sample sub-image;
and training to obtain the error correction recognition model by utilizing a predetermined neural network topological structure and a training algorithm according to each sample subimage.
5. The method of claim 1, wherein obtaining corresponding review features according to the identification information, and reviewing each written answer according to the review features based on a preset review model comprises: when the correct answer to be identified is deleted, acquiring a corresponding evaluation result according to whether the single-point answer image has the identification result of deleting the answer;
when the correct answer to be identified is modified or added, obtaining at least one item of review characteristics of the single-point answer image based on a preset review method according to the skip information of the single handwritten character segmentation of each sub-image of the corresponding single-point answer image and the probability that each sub-image contains a single handwritten character as a corresponding character; and obtaining the evaluation result of the single-point answer image based on a pre-constructed evaluation model according to the evaluation characteristics.
6. The method of claim 2, wherein the review features include at least answer recognition posterior probability features and coverage features of character segmentation decoding paths, and further include at least one of raised character scale features, missed character scale features, and absorbed pixel scale features; wherein:
the answer recognition posterior probability characteristic is an average value of probabilities that a single handwritten character contained in the single-point answer image is a corresponding character;
the coverage characteristic of the character segmentation decoding path is the ratio of the number of traversed single handwritten characters to the number of handwritten character subunit layers in the limited network when the single-point answer image is subjected to single handwritten character segmentation;
the character increasing character proportion characteristic is the proportion of the character subunits of the forward jumping or the self jumping in the single-point answer image to the total number of the character subunits in the single-point answer image when the single-point answer image is subjected to single handwritten character cutting;
the character missing ratio characteristic is the ratio of the back skip character subunit to the total number of the character subunits in the single-point answer image when the single-point answer image is subjected to single handwritten character segmentation;
the absorbed pixel proportion characteristic is the proportion of the number of pixels absorbed by the absorbing layer of pixels obtained by traversing the single-point answer image by using a window with a preset pixel width along the x-axis direction when the single-point answer image is subjected to single handwritten character segmentation, and the pixel proportion accounts for the total number of the pixel points of the single-point answer image.
7. The utility model provides an automatic device of reviewing of wrong examination questions which characterized in that includes:
the acquisition module is used for acquiring single-point answer images corresponding to all written answers on a medium to be answered;
the identification module is used for carrying out image identification on each acquired single-point answer image by using a preset image detection method and/or based on a preset identification model to obtain an identification result;
the evaluation module is used for obtaining corresponding evaluation characteristics according to the identification information and evaluating each written answer according to the evaluation characteristics based on a preset evaluation model; the identification information comprises the identification result and identification intermediate process information;
the wrong answer type comprises at least one of deletion, modification and addition;
the identification module is specifically configured to:
when the correct answer is deleted, judging whether the single-point answer image has a corresponding deletion answer or not by carrying out image detection on the single-point answer image;
when the correct answer is modified or added, firstly segmenting each single-point answer image by using a single handwritten character, and identifying each sub-image which contains each single handwritten character after segmentation to obtain an identification result;
the identification module is further configured to:
when the correct answer is modified or added, performing single handwritten character segmentation on the corresponding single-point answer image based on a pre-constructed limited network corresponding to the correct answer to obtain each sub-image of the single-point answer image; identifying each sub-image based on a pre-constructed error correction recognition model to obtain a recognition result of a single handwritten character contained in each sub-image; the limited network is established aiming at the handwritten character strings in the correct answers of all answer samples;
the acquisition module is further configured to:
acquiring a single-point answer image corresponding to each written answer on a medium to be read according to an answer image of the medium to be read and an answer medium template corresponding to the medium to be read; the answer medium template comprises position information of each character in a test question to be answered and at least one single-point answer area preset according to a correct answer, wherein the single-point answer image is obtained according to the corresponding single-point answer area.
8. The apparatus of claim 7, wherein the pre-constructed restricted network is pre-constructed based on its corresponding correct answers, including a start layer, an end layer, each single handwritten character subunit layer, and an absorbing layer existing between the preceding layers:
each single handwritten character subunit layer is constructed according to the handwritten sample image of each single character in the corresponding correct answer;
and front jump arcs and back jump arcs exist from each absorption layer to each single handwritten character subunit layer and each single handwritten character subunit layer.
9. The apparatus of claim 8, wherein the identification module is further configured to:
traversing the single-point answer image by using a window with preset pixel width and single-point answer image height along the x-axis direction, and inputting pixels obtained by traversing into a pre-constructed limited network corresponding to correct answers to obtain each sub-image of the single-point answer image.
10. The apparatus of claim 7, wherein the error correction problem recognition model is obtained by training through the following steps:
carrying out single handwritten character segmentation on the sample single-point answer image to obtain each sample sub-image of the sample single-point answer image;
labeling the recognition result of each sample sub-image;
and training to obtain the error correction recognition model by utilizing a predetermined neural network topological structure and a training algorithm according to each sample subimage.
11. The apparatus of claim 7, wherein the review module is further configured to:
when the correct answer to be identified is deleted, acquiring a corresponding evaluation result according to whether the single-point answer image has the identification result of deleting the answer;
when the correct answer to be identified is modified or added, obtaining at least one item of review characteristics of the single-point answer image based on a preset review method according to the skip information of the single handwritten character segmentation of each sub-image of the corresponding single-point answer image and the probability that each sub-image contains a single handwritten character as a corresponding character; and obtaining the evaluation result of the single-point answer image based on a pre-constructed evaluation model according to the evaluation characteristics.
12. The apparatus of claim 8, wherein the review features comprise at least answer recognition posterior probability features and coverage features of character segmentation decoding paths, and further comprise at least one of raised character scale features, missed character scale features, and absorbed pixel scale features; wherein:
the answer recognition posterior probability characteristic is an average value of probabilities that a single handwritten character contained in the single-point answer image is a corresponding character;
the coverage characteristic of the character segmentation decoding path is the ratio of the number of traversed single handwritten characters to the number of handwritten character subunit layers in the limited network when the single-point answer image is subjected to single handwritten character segmentation;
the character increasing character proportion characteristic is the proportion of the character subunits of the forward jumping or the self jumping in the single-point answer image to the total number of the character subunits in the single-point answer image when the single-point answer image is subjected to single handwritten character cutting;
the character missing ratio characteristic is the ratio of the back skip character subunit to the total number of the character subunits in the single-point answer image when the single-point answer image is subjected to single handwritten character segmentation;
the absorbed pixel proportion characteristic is the proportion of the number of pixels absorbed by the absorbing layer of pixels obtained by traversing the single-point answer image by using a window with a preset pixel width along the x-axis direction when the single-point answer image is subjected to single handwritten character segmentation, and the pixel proportion accounts for the total number of the pixel points of the single-point answer image.
13. The utility model provides an automatic electronic equipment that reviews of wrong questions which characterized in that includes:
at least one processor; and
at least one memory communicatively coupled to the processor, wherein:
the memory stores program instructions executable by the processor, the processor invoking the program instructions to perform the method of any of claims 1 to 6.
14. A non-transitory computer-readable storage medium storing computer instructions for causing a computer to perform the method according to any one of claims 1 to 6.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710947823.9A CN107729936B (en) | 2017-10-12 | 2017-10-12 | Automatic error correction review method and system |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710947823.9A CN107729936B (en) | 2017-10-12 | 2017-10-12 | Automatic error correction review method and system |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN107729936A CN107729936A (en) | 2018-02-23 |
| CN107729936B true CN107729936B (en) | 2020-12-08 |
Family
ID=61211192
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710947823.9A Active CN107729936B (en) | 2017-10-12 | 2017-10-12 | Automatic error correction review method and system |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107729936B (en) |
Families Citing this family (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108388895B (en) * | 2018-03-04 | 2022-02-18 | 南京理工大学 | Machine learning-based automatic processing method for test paper answer sheet |
| CN110969052A (en) * | 2018-09-29 | 2020-04-07 | 杭州萤石软件有限公司 | Operation correction method and equipment |
| CN109815955B (en) * | 2019-03-04 | 2021-09-28 | 杭州大拿科技股份有限公司 | Question assisting method and system |
| CN111008594B (en) * | 2019-12-04 | 2024-04-19 | 科大讯飞股份有限公司 | Error-correction question review method, related device and readable storage medium |
| CN111104881B (en) * | 2019-12-09 | 2023-12-01 | 科大讯飞股份有限公司 | Image processing method and related device |
| CN112507879A (en) * | 2020-12-08 | 2021-03-16 | 科大讯飞股份有限公司 | Evaluation method, evaluation device, electronic equipment and storage medium |
| CN112257703B (en) * | 2020-12-24 | 2021-03-16 | 北京世纪好未来教育科技有限公司 | Image recognition method, device, equipment and readable storage medium |
| CN112926587B (en) * | 2021-02-19 | 2024-03-29 | 北京大米未来科技有限公司 | Text recognition method and device, readable storage medium and electronic equipment |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102568475A (en) * | 2011-12-31 | 2012-07-11 | 安徽科大讯飞信息科技股份有限公司 | System and method for assessing proficiency in Putonghua |
| CN104268603A (en) * | 2014-09-16 | 2015-01-07 | 科大讯飞股份有限公司 | Intelligent marking method and system for text objective questions |
| CN105160342A (en) * | 2015-08-11 | 2015-12-16 | 成都数联铭品科技有限公司 | HMM-GMM-based automatic word picture splitting method and system |
| CN105183712A (en) * | 2015-08-27 | 2015-12-23 | 北京时代焦点国际教育咨询有限责任公司 | Method and apparatus for scoring English composition |
Family Cites Families (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| KR101616909B1 (en) * | 2012-10-31 | 2016-04-29 | 에스케이텔레콤 주식회사 | Automatic scoring apparatus and method |
| CN103903491A (en) * | 2014-02-14 | 2014-07-02 | 广东小天才科技有限公司 | Method and device for achieving writing inspection |
| CN106781784A (en) * | 2017-01-04 | 2017-05-31 | 王骁乾 | A kind of intelligence correction system |
| CN106846203B (en) * | 2017-03-09 | 2020-12-18 | 常州工学院 | Intelligent marking method and system |
-
2017
- 2017-10-12 CN CN201710947823.9A patent/CN107729936B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102568475A (en) * | 2011-12-31 | 2012-07-11 | 安徽科大讯飞信息科技股份有限公司 | System and method for assessing proficiency in Putonghua |
| CN104268603A (en) * | 2014-09-16 | 2015-01-07 | 科大讯飞股份有限公司 | Intelligent marking method and system for text objective questions |
| CN105160342A (en) * | 2015-08-11 | 2015-12-16 | 成都数联铭品科技有限公司 | HMM-GMM-based automatic word picture splitting method and system |
| CN105183712A (en) * | 2015-08-27 | 2015-12-23 | 北京时代焦点国际教育咨询有限责任公司 | Method and apparatus for scoring English composition |
Non-Patent Citations (1)
| Title |
|---|
| An Automatic English Composition scoring model based on neural network algorithm;Ya Zhou et al.;《 2014 IEEE/ACIS 13th International Conference on Computer and Information Science (ICIS)》;20140929;全文 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN107729936A (en) | 2018-02-23 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107729936B (en) | Automatic error correction review method and system | |
| CN110110585B (en) | Intelligent paper reading implementation method and system based on deep learning and computer program | |
| CN106534548B (en) | Voice error correction method and device | |
| CN111597908A (en) | Test paper correcting method and test paper correcting device | |
| CN111753767A (en) | Method and device for automatically correcting operation, electronic equipment and storage medium | |
| JP5043892B2 (en) | Automatic evaluation of excessive repeated word usage in essays | |
| CN110363194A (en) | Intelligently reading method, apparatus, equipment and storage medium based on NLP | |
| KR101813683B1 (en) | Method for automatic correction of errors in annotated corpus using kernel Ripple-Down Rules | |
| CN110188654B (en) | Video behavior identification method based on mobile uncut network | |
| KR102034346B1 (en) | Method and Device for Detecting Slang Based on Learning | |
| CN105468468A (en) | Data error correction method and apparatus facing question answering system | |
| Pacha et al. | Towards self-learning optical music recognition | |
| CN112446259A (en) | Image processing method, device, terminal and computer readable storage medium | |
| CN114841176B (en) | Neural machine translation robustness enhancement method, equipment and storage medium | |
| CN112347997A (en) | Test question detection and identification method and device, electronic equipment and medium | |
| CN111008594A (en) | Error correction evaluation method, related equipment and readable storage medium | |
| CN117011686B (en) | Multi-mode generation type large model training method and device and computer equipment | |
| CN113205046A (en) | Method, system, device and medium for identifying question book | |
| CN110889412B (en) | Medical long text positioning and classifying method and device in physical examination report | |
| CN112818693A (en) | Automatic extraction method and system for electronic component model words | |
| CN112686263A (en) | Character recognition method and device, electronic equipment and storage medium | |
| JP7293658B2 (en) | Information processing device, information processing method and program | |
| Zhu et al. | Interactive reinforcement learning for object grounding via self-talking | |
| CN113111869B (en) | Method and system for extracting text picture and description thereof | |
| CN115601768A (en) | Method, device and equipment for judging written characters and storage medium |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |