WO2023284608A1 - 字符识别模型生成方法、装置、计算机设备和存储介质 - Google Patents
字符识别模型生成方法、装置、计算机设备和存储介质 Download PDFInfo
- Publication number
- WO2023284608A1 WO2023284608A1 PCT/CN2022/104107 CN2022104107W WO2023284608A1 WO 2023284608 A1 WO2023284608 A1 WO 2023284608A1 CN 2022104107 W CN2022104107 W CN 2022104107W WO 2023284608 A1 WO2023284608 A1 WO 2023284608A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- character data
- data set
- recognized
- similarity
- model
- Prior art date
Links
- 238000000034 method Methods 0.000 title claims abstract description 37
- 238000012549 training Methods 0.000 claims abstract description 162
- 238000012545 processing Methods 0.000 claims description 38
- 230000009466 transformation Effects 0.000 claims description 13
- 238000003062 neural network model Methods 0.000 claims description 9
- 238000012216 screening Methods 0.000 claims description 8
- 230000008569 process Effects 0.000 claims description 4
- 238000010586 diagram Methods 0.000 description 8
- 238000004364 calculation method Methods 0.000 description 5
- 230000000694 effects Effects 0.000 description 3
- 238000004519 manufacturing process Methods 0.000 description 3
- 238000013528 artificial neural network Methods 0.000 description 2
- 238000010276 construction Methods 0.000 description 2
- 238000013499 data model Methods 0.000 description 2
- 238000012546 transfer Methods 0.000 description 2
- 238000004458 analytical method Methods 0.000 description 1
- 238000007405 data analysis Methods 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000003287 optical effect Effects 0.000 description 1
- 238000003672 processing method Methods 0.000 description 1
- 230000003068 static effect Effects 0.000 description 1
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
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/60—Analysis of geometric attributes
- G06T7/62—Analysis of geometric attributes of area, perimeter, diameter or volume
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/60—Analysis of geometric attributes
- G06T7/64—Analysis of geometric attributes of convexity or concavity
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
- Y02T10/00—Road transport of goods or passengers
- Y02T10/10—Internal combustion engine [ICE] based vehicles
- Y02T10/40—Engine management systems
Definitions
- the present application relates to the technical field of computer recognition, in particular to a method, device, computer equipment and storage medium for generating a character recognition model.
- a method, device, computer equipment and storage medium for generating a character recognition model are provided.
- a method for generating a character recognition model comprising:
- the pre-training model is a pre-trained model for identifying the target character data set
- a character recognition model generation device comprising:
- the similarity acquisition module is used to obtain the similarity between a plurality of recognized character data sets and the character data sets to be recognized, and the recognized character data sets matched with the similarity between the character data sets to be recognized are used as target character dataset;
- a model building module configured to obtain a pre-training model corresponding to the target character data set, and construct a target training model according to the pre-training model; the pre-training model is used to identify the target character data after pre-training set model;
- a model generation module configured to generate a target training data set according to the recognized character data set and the character data set to be recognized; according to the target training data set, train the target training model to obtain the same as the described The character recognition model corresponding to the character data set to be recognized.
- a computer device comprising a memory and a processor, the memory stores computer-readable instructions, and the processor implements the following steps when executing the computer-readable instructions:
- the pre-training model is a pre-trained model for identifying the target character data set
- a computer-readable storage medium on which computer-readable instructions are stored, and when the computer-readable instructions are executed by a processor, the following steps are implemented:
- the pre-training model is a pre-trained model for identifying the target character data set
- Fig. 1 is an application environment diagram of a method for generating a character recognition model in one or more embodiments
- FIG. 2 is a schematic flow diagram of a method for generating a character recognition model in one or more embodiments
- FIG. 3 is a schematic flow diagram of the steps of obtaining the similarity between multiple recognized character data sets and character data sets to be recognized in one or more embodiments;
- FIG. 4 is a schematic flow diagram of the step of obtaining the similarity of picture parameters between the recognized character data set and the character data set to be recognized in one or more embodiments;
- Fig. 5 is a schematic flow diagram of the steps of obtaining the text outline similarity between the recognized character data set and the character data set to be recognized in one or more embodiments;
- Fig. 6 is a structural block diagram of a character recognition model generation device in one or more embodiments.
- Figure 7 is an internal block diagram of a computer device in one or more embodiments.
- the character recognition model generation method provided in this application can be applied to the application environment shown in FIG. 1 .
- the terminal 11 communicates with the server 12 through the network.
- the server 12 obtains the character data set to be recognized from the terminal 11 through the network, and obtains the similarity between a plurality of recognized character data sets and the character data set to be recognized, and the server 12 compares the similarity with the character data set to be recognized
- the matched identified character data set is used as the target character data set;
- the server 12 obtains the pre-training model corresponding to the target character data set, and the server 12 builds the target training model according to the pre-training model;
- the pre-training model is used for pre-training Identify the model of the target character data set;
- Server 12 generates the target training data set according to the recognized character data set and the character data set to be recognized;
- Server 12 trains the target training model according to the target training data set to obtain the character data to be recognized
- the terminal 11 can directly upload the character data to be recognized to the server 12, so that the target training model after the server runs the training to identify the character data and returns the recognition result to the terminal 11; the server 12 can also transfer the target training model after the training to the terminal 11, so that the terminal 11 can directly recognize the character data.
- the terminal 11 can be, but not limited to, various personal computers, notebook computers, smart phones, tablet computers and portable wearable devices
- the server 12 can be realized by an independent server or a server cluster composed of multiple servers.
- a method for generating a character recognition model is provided, and the method is applied to the server 12 in FIG. 1 as an example, including the following steps:
- Step 21 obtaining the similarity between multiple recognized character data sets and the character data sets to be recognized, and using the recognized character data set matching the similarity between the character data sets to be recognized as the target character data set.
- the recognized character data set refers to the character data set that can be recognized by the pre-trained model after the model has been trained with the character data set;
- the unrecognized character data set refers to the character data set that has not been recognized or The character dataset used for model training.
- the character data set can be composed of multiple pictures containing characters, and there are no restrictions on the file format and size of the pictures.
- the similarity refers to the overall similarity between the recognized character data set and the unrecognized character data set judged by preset conditions.
- Combination weighting determination for example, the number, size, format, and color of pictures in a character dataset can be used as the basis for determining the similarity.
- shape, outline, and color of characters in the picture can also be used as the basis for determining the similarity.
- similarity matching means that the similarity between the recognized character data set and the character data set to be recognized falls within the preset similarity interval, that is, the similarity between the two can be regarded as similar if it reaches a certain preset condition. degree match.
- the target character data set refers to the recognized character data set whose similarity with the to-be-recognized character data set satisfies the matching condition, and there may be one or more.
- the server After the server obtains the data set of characters to be recognized, it performs data analysis and processing on the data set of characters to be recognized to obtain a value that can be used for similarity judgment; the server then obtains at least one recognized character that has been used for model training
- the character data set using the same analysis and processing method, also obtains a value that can be used for similarity judgment; compares the value of the character data set to be recognized with the value of at least one recognized character data set, and obtains multiple similarities degree; select at least one degree of similarity that reaches a preset standard from the plurality of degrees of similarity, and use the recognized character data set corresponding to the degree of similarity as the target character data set.
- Step 22 obtaining a pre-training model corresponding to the target character data set, and constructing a target training model according to the pre-training model;
- the pre-training model is a pre-trained model for identifying the target character data set.
- the pre-training model is a model trained in advance through the target character data set (recognized character data set), and the weight parameters in the model have been adjusted when using the target character data set (recognized character data set) for training, Character data in the target character data set (recognized character data set) can be recognized.
- the server obtains the pre-training model corresponding to the target character data set, and constructs the target training model according to the neural network structure and weight parameters of the pre-training model; during the construction process, the size of the weight parameters and the neural network structure can be adjusted accordingly .
- Step 23 Generate a target training data set according to the recognized character data set and the unrecognized character data set; train the target training model according to the target training data set to obtain a character recognition model corresponding to the unrecognized character data set.
- the target training data set is generated according to the recognized character data set and the unrecognized character data set, for example, character data is randomly taken from the recognized character data set, and the ratio of 2:8 to all unrecognized character data sets is used to form a new
- the character data set is used as the target training data set; there are no restrictions on the specific character data acquisition method, acquisition quantity, combination ratio, etc.
- the server generates a new character data set as the target training data set according to the recognized character data set and the unrecognized character data set; uses the target training data set as the training data of the target training model to train the target training model, When the loss value of the target training model reaches the preset standard, or the recognition accuracy rate of the target training model reaches the preset standard for the character data set to be recognized, it is determined that the training of the target training model is completed, and the character recognition model corresponding to the character data set to be recognized is obtained. .
- a company obtained a large number of pictures containing character data in scene a to form character data set A, and used the character data set A to train the character recognition model A'.
- scene a is very similar to scene b, and the character recognition model A' can be directly applied to the recognition character data set B; that is, the character data set A can be used as a character data set in the recognized character data set, and the character data set B can be used as a character data set to be recognized.
- the character recognition model A' can be used to construct a new character data model B', or the character recognition model A' can be directly used as the character data model B' for training; the target training data set can be obtained according to the character data set A and the character data set B, It is also possible to only use the character data set B; that is, it is also feasible to directly use the character data set B to train the character recognition model A'.
- the above character recognition model generation method, device, computer equipment and storage medium includes: obtaining the similarity between a plurality of recognized character data sets and the character data set to be recognized, and comparing the similarity with the character data set to be recognized
- the matching recognized character data set is used as the target character data set;
- the pre-training model corresponding to the target character data set is obtained, and the target training model is constructed according to the pre-training model;
- the pre-training model is used to identify the target character data after pre-training generate a target training data set according to the recognized character data set and the character data set to be recognized; train the target training model according to the target training data set to obtain a character recognition model corresponding to the character data set to be recognized.
- this application uses the pre-trained model of the recognized character data set as the training model of the character data set to be recognized, and realizes the The similarity between them is used to learn and transfer the model training, thereby improving the generation efficiency of the character recognition model.
- both the recognized character data set and the unrecognized character data set carry picture parameter information and text outline information.
- step 21 obtain the similarity between a plurality of recognized character data sets and the character data sets to be recognized, including:
- Step 31 according to the picture parameter information carried by the recognized character data set and the picture parameter information carried by the unrecognized character data set, obtain the picture parameter similarity between the recognized character data set and the unrecognized character data set;
- Step 32 according to the text contour information carried by the recognized character data set and the text contour information carried by the unrecognized character data set, obtain the text contour similarity between the recognized character data set and the unrecognized character data set;
- Step 33 weighting the image parameter similarity and the text outline similarity, and determining the similarity between the recognized character data set and the character data set to be recognized according to the result of the weighting processing.
- the image parameter information may include the number, size, file format, pixels, color, etc. of the images in the character data set;
- the text outline information includes text outline, geometric convex hull, etc.
- the picture parameter similarity refers to the similarity of picture parameter information between the recognized character data set and the unrecognized character data set; for example, the sizes of the five pictures in the recognized character data set are all 10 ⁇ 10, and The size of the 6 pictures is 10 ⁇ 10; if only the size is considered as the similarity judgment condition, the similarity of the picture parameters of the two character data sets can be 100%.
- the text outline similarity refers to the similarity of the text outline information between the recognized character data set and the character data set to be recognized;
- the size of the shape of the character area in the 6 pictures in the character data set is 10 ⁇ 10; if only the size is considered as the similarity judgment condition, the similarity of the picture parameters of the two character data sets can be 100%.
- the server can set the weight of the picture parameter similarity to 0.6, and the text outline similarity's weight to 0.4, then the weighted similarity can be 0.6c +0.4d.
- step 31 according to the picture parameter information carried by the recognized character data set, and the picture parameter information carried by the character data set to be recognized, obtain the recognized character data set and the character data to be recognized
- the similarity of image parameters between sets including:
- Step 41 obtaining the recognized character data set and the unrecognized character data set, the mean value of the color channel and the mean value of width and height of all images;
- Step 42 according to the color channel mean value and the width and height mean value, determine the cosine distance of the color channel mean value of the recognized character data set and the character data set to be recognized, and the cosine of the width and height mean value of the recognized character data set and the character data set to be recognized distance;
- step 43 the sum of the cosine distance of the color channel mean and the cosine distance of the width and height mean is used as the similarity of the picture parameters.
- the server obtains the RGB mean value of the image pixels in the two character data sets as the color channel mean value; the RGB mean value includes the color value mean value of all pixels in the image in each color channel.
- the server obtains the average width and height of images in the two character datasets as the average width and height.
- P is the character data set to be recognized
- Qi is the recognized character data set
- Ai is the sum of the cosine distance of the mean value of the color channel and the cosine distance of the mean width and height
- rgb[P] is the mean value of the color channel of the character data set to be recognized
- rgb[Qi] is the average color channel value of the recognized character dataset
- hw[P] is the average width and height value of the character dataset to be recognized
- hw[Qi] is the average width and height value of the recognized character dataset
- cos is the cosine distance Calculation symbols.
- the image parameter similarity between the recognized character data set and the character data set to be recognized is obtained by calculating the mean value of the color channel and the mean value of width and height respectively, so as to achieve accurate calculation of the recognized character data set and the character data to be recognized
- the effect of the similarity between the sets improves the generation efficiency of the character recognition model.
- step 32 according to the text outline information carried by the recognized character data set and the text outline information carried by the unrecognized character data set, obtain the recognized character data set and the unrecognized character data Text contour similarity between sets, including:
- Step 51 identifying the contour feature information corresponding to the text information in the recognized character data set and the unrecognized character data set;
- Step 52 determine the convex hull information of the text information in the recognized character data set and the unrecognized character data set according to the contour feature information;
- Step 53 according to the convex hull information, determine the convex hull area of the text information and the convex hull overlapping area of the text information in the recognized character data set and the character data set to be recognized;
- Step 54 according to the convex hull area of the text information and the overlapping area of the convex hull of the text information, obtain the text outline similarity between the recognized character data set and the character data set to be recognized.
- the server can use the contour matching algorithm to find the text information in the region of interest as the contour feature information; the contour matching algorithm can divide the images in the recognized character data set and the character data set to be recognized into multiple channels to find the text information
- the edge is converted into a contour feature, and the contour matching is performed; wherein, the contour matching method may include a convex hull, etc., and is not specifically limited.
- the contour feature information includes the detected text contour; when the convex hull is used for calculation, each text contour in the recognized character data set and the unrecognized character data set can be translated separately, so that the circumscribed text contour in each text contour The upper left corner of the rectangle coincides with the origin of the preset coordinate system.
- convexhull[P] the minimum geometry of the text outline in the character data set P to be recognized convex hull
- Qi the minimum geometric convex hull of the text outline in the recognized character dataset Qi).
- the text contour similarity is defined as the intersection-over-union ratio of the geometric convex hulls of these two texts:
- Bi is the text outline similarity between the recognized character dataset and the character dataset to be recognized
- S[convexhull[P]] and S[convexhull[Qi]] are the character dataset to be recognized and the recognized character dataset respectively,
- the area of the convex hull of the text is included in the convex hull information
- S[inter] is the area of the overlapping area of the convex hull of the text between the character dataset to be recognized and the recognized character dataset, that is, the overlap area of the convex hull.
- the image parameter similarity between the recognized character data set and the character data set to be recognized is obtained by calculating the mean value of the color channel and the mean value of width and height respectively, so as to achieve accurate calculation of the recognized character data set and the character data to be recognized
- the effect of the similarity between the sets improves the generation efficiency of the character recognition model.
- using the recognized character data set matched with the similarity between the character data sets to be recognized as the target character data set includes: screening out at least one of the similarities between multiple recognized character data sets Recognized character data sets whose similarity is greater than or equal to the preset similarity threshold; among the similarities of at least one recognized character data set obtained after screening, the recognized character data set with the highest similarity is identified as the target character data set .
- the server can set a preset similarity threshold as a criterion for judging whether the similarity matches; for example, if the preset similarity threshold is set to 85, then the recognized character data set whose calculated similarity is greater than or equal to 85 can be used as the target character dataset.
- the preset similarity threshold can also be set to a range, for example, a similarity of 85, ⁇ 5 can be regarded as a match, that is, a similarity range of 80-90 or above can be used as a recognized character dataset.
- Target character dataset can be set to a range, for example, a similarity of 85, ⁇ 5 can be regarded as a match, that is, a similarity range of 80-90 or above can be used as a recognized character dataset.
- the generation efficiency of the character recognition model is improved by setting the preset similarity threshold as the recognition basis of the target character data set.
- constructing the target training model according to the pre-training model includes: obtaining model parameters of the pre-training model; applying the model parameters to the pre-built neural network model to obtain the target training model.
- the server obtains the model parameters of the pre-trained model, and sets the model parameters in the pre-built neural network model as the target training model.
- the pre-built neural network model to which the model parameters are applied can be trained on the basis of a certain recognition ability, which improves the generation efficiency of the character recognition model .
- the target training model before training the target training model according to the target training data set, it also includes: performing gamma transformation processing and histogram equalization processing on the images in the target training data set; The image after the equalization processing is processed in a unified image size; the image after the uniform size processing is input into the target training model for training.
- gamma transformation processing is to enhance the gray value of the darker area of the image in the target training data set through nonlinear transformation, and reduce the gray value of the area with too large gray value in the image; after gamma transformation processing to enhance the details of the image as a whole.
- Histogram equalization can increase the global contrast of the image, especially when the contrast of useful data in the image in the target training dataset is close; after histogram equalization, the brightness can be better distributed on the histogram.
- Unified size processing can unify images of different sizes in the target training data set into images of the same size, improving image processing efficiency.
- the generation efficiency of the character recognition model is improved by performing image processing on the images in the target training data set.
- a character recognition model generation device including a similarity acquisition module 61, a model construction module 62 and a model generation module 63, wherein:
- the similarity acquisition module 61 is used to acquire the similarity between a plurality of recognized character data sets and the character data sets to be recognized, and use the recognized character data sets matched with the similarity between the character data sets to be recognized as the target character data set;
- Model building module 62 is used for obtaining the pre-training model corresponding to target character data set, constructs target training model according to pre-training model;
- Pre-training model is the model that is used to identify target character data set after pre-training;
- the model generation module 63 is used to generate a target training data set according to the recognized character data set and the character data set to be recognized; according to the target training data set, the target training model is trained to obtain character recognition corresponding to the character data set to be recognized Model.
- both the recognized character data set and the unrecognized character data set carry picture parameter information and text outline information
- the similarity acquisition module 61 is also used to obtain the picture parameters between the recognized character data set and the unrecognized character data set according to the picture parameter information carried by the recognized character data set and the picture parameter information carried by the unrecognized character data set Similarity: According to the text outline information carried by the recognized character data set and the text outline information carried by the unrecognized character data set, the similarity of the text outline between the recognized character data set and the unrecognized character data set is obtained; the image parameters The similarity and the similarity of the text outline are weighted, and the similarity between the recognized character data set and the character data set to be recognized is determined according to the result of the weighted processing.
- the similarity acquisition module 61 is also used to obtain the color channel mean value and the width and height mean value of all images in the recognized character data set and the character data set to be recognized; according to the color channel mean value and the width and height mean value, determine the The cosine distance of the color channel mean value of the recognized character data set and the character data set to be recognized, and the cosine distance of the width and height mean value of the recognized character data set and the character data set to be recognized; the cosine distance of the color channel mean value and the width and height mean value of The sum of cosine distances is used as the similarity of image parameters.
- the similarity acquisition module 61 is also used to identify the contour feature information corresponding to the text information in the recognized character data set and the character data set to be recognized; determine the recognized character data set and the character data set to be recognized according to the contour feature information In the character data set, the convex hull information of the text information; according to the convex hull information, determine the convex hull area of the text information and the convex hull overlapping area of the text information in the recognized character data set and the character data set to be recognized; according to the convex hull information of the text information The envelope area and the overlapping area of the convex hull of the text information are used to obtain the text outline similarity between the recognized character data set and the character data set to be recognized.
- the similarity acquisition module 61 is also used to screen out at least one recognized character data set whose similarity is greater than or equal to a preset similarity threshold from the similarities of multiple recognized character data sets; Among the obtained similarities of at least one recognized character data set, the recognized character data set with the highest similarity is identified as the target character data set.
- the model building module 62 is also used to obtain model parameters of the pre-trained model; apply the model parameters to the pre-built neural network model to obtain the target training model.
- the model generation module 63 is also used to perform gamma transformation processing and histogram equalization processing on images in the target training data set; perform image size uniform processing on images that have undergone gamma transformation processing and histogram equalization processing ; Input the image with unified size into the target training model for training.
- Each module in the above-mentioned character recognition model generation device can be fully or partially realized by software, hardware and a combination thereof.
- the above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, and can also be stored in the memory of the computer device in the form of software, so that the processor can invoke and execute the corresponding operations of the above-mentioned modules.
- a computer device is provided.
- the computer device may be a server, and its internal structure may be as shown in FIG. 7 .
- the computer device includes a processor, memory and a network interface connected by a system bus. Wherein, the processor of the computer device is used to provide calculation and control capabilities.
- the memory of the computer device includes a non-volatile storage medium and an internal memory.
- the non-volatile storage medium stores an operating system, computer readable instructions and a database.
- the internal memory provides an environment for the execution of the operating system and computer readable instructions in the non-volatile storage medium.
- the database of the computer device is used to store the data generated by the character recognition model.
- the network interface of the computer device is used to communicate with the external terminal through the network connection.
- Figure 7 is a block diagram of a partial structure related to the solution of the present application, and does not constitute a limitation to the computer equipment on which the solution of the application is applied.
- the specific computer equipment may include There may be more or fewer components than shown in the figures, or certain components may be combined, or have different component arrangements.
- a computer device including a memory and a processor, where computer-readable instructions are stored in the memory, and the processor implements the following steps when executing the computer-readable instructions:
- the pre-training model is a pre-trained model for identifying the target character data set
- both the recognized character data set and the character data set to be recognized carry picture parameter information and text outline information; when the processor executes the computer-readable instructions, the following steps are also implemented: according to the picture carried by the recognized character data set Parameter information, and the picture parameter information carried by the character data set to be recognized, obtain the similarity of the picture parameters between the recognized character data set and the character data set to be recognized; according to the text outline information carried by the recognized character data set, and the The text contour information carried by the character data set is used to obtain the text contour similarity between the recognized character data set and the character data set to be recognized; the similarity of the image parameters and the text contour similarity are weighted, and the weighted processing results are used to determine the The similarity between the recognized character dataset and the character dataset to be recognized.
- the processor executes the computer-readable instructions, the following steps are also implemented: obtaining the recognized character data set and the character data set to be recognized, the color channel mean value and the width and height mean value of all images; according to the color channel mean value and width and height Mean value, determine the cosine distance of the color channel mean value of the recognized character data set and the character data set to be recognized, and the cosine distance of the width and height mean value of the recognized character data set and the character data set to be recognized; the cosine distance of the color channel mean value and The sum of the cosine distances of the average width and height is used as the similarity of the image parameters.
- the processor when the processor executes the computer-readable instructions, the following steps are also implemented: identifying the contour feature information corresponding to the text information in the recognized character data set and the character data set to be recognized; determining the recognized character data according to the contour feature information set and the character data set to be recognized, the convex hull information of the text information; according to the convex hull information, determine the convex hull area of the text information and the convex hull overlapping area of the text information in the recognized character data set and the character data set to be recognized; The convex hull area of the text information and the overlapping area of the convex hull of the text information are used to obtain the text outline similarity between the recognized character data set and the character data set to be recognized.
- the processor executes the computer-readable instructions, the following steps are further implemented: screening out at least one recognized character data whose similarity is greater than or equal to a preset similarity threshold from the similarities of multiple recognized character data sets set; among the similarities of at least one recognized character data set obtained after screening, the recognized character data set with the highest similarity is identified as the target character data set.
- the processor executes the computer-readable instructions, the following steps are further implemented: acquiring model parameters of the pre-trained model; applying the model parameters to the pre-built neural network model to obtain the target training model.
- the processor when the processor executes the computer-readable instructions, the following steps are also implemented: performing gamma transformation processing and histogram equalization processing on the images in the target training data set; performing gamma transformation processing and histogram equalization processing on the images Carry out uniform image size processing; input the image after size uniform processing into the target training model for training.
- a computer-readable storage medium on which computer-readable instructions are stored, and when the computer-readable instructions are executed by a processor, the following steps are implemented:
- the pre-training model is a pre-trained model for identifying the target character data set
- both the recognized character data set and the character data set to be recognized carry picture parameter information and text outline information; when the computer readable instructions are executed by the processor, the following steps are also implemented: according to the recognized character data set carried The picture parameter information, and the picture parameter information carried by the character data set to be recognized, obtain the picture parameter similarity between the recognized character data set and the character data set to be recognized; according to the text outline information carried by the recognized character data set, and Recognize the text contour information carried by the character data set, and obtain the text contour similarity between the recognized character data set and the character data set to be recognized; perform weighted processing on the image parameter similarity and text contour similarity, and determine according to the weighted processing results The similarity between the recognized character dataset and the character dataset to be recognized.
- the following steps are also implemented: obtaining the recognized character data set and the character data set to be recognized, the color channel mean value and the width and height mean value of all images; according to the color channel mean value and width High mean value, determine the cosine distance of the color channel mean value of the recognized character data set and the character data set to be recognized, and the cosine distance of the width and height mean value of the recognized character data set and the character data set to be recognized; the cosine distance of the color channel mean value The sum of the cosine distances from the average width and height is used as the similarity of the image parameters.
- the following steps are also implemented: identifying the contour feature information corresponding to the text information in the recognized character data set and the character data set to be recognized; determining the recognized character according to the contour feature information
- the convex hull information of the text information in the dataset and the character dataset to be recognized ; determine the convex hull area of the text information and the convex hull overlapping area of the text information in the recognized character dataset and the character dataset to be recognized according to the convex hull information;
- the text outline similarity between the recognized character data set and the character data set to be recognized is obtained.
- the following steps are further implemented: screening out at least one recognized character whose similarity is greater than or equal to a preset similarity threshold from the similarities of multiple recognized character data sets A data set; among the similarities of at least one recognized character data set obtained after screening, the recognized character data set with the highest similarity is identified as the target character data set.
- the following steps are further implemented: acquiring model parameters of the pre-trained model; applying the model parameters to the pre-built neural network model to obtain the target training model.
- the following steps are also implemented: performing gamma transformation processing and histogram equalization processing on the images in the target training data set; The image is processed in a unified image size; the image after the unified size processing is input into the target training model for training.
- Non-volatile memory may include read-only memory (Read-Only Memory, ROM), magnetic tape, floppy disk, flash memory or optical memory, etc.
- Volatile memory can include Random Access Memory (RAM) or external cache memory.
- RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Geometry (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- General Engineering & Computer Science (AREA)
- Character Discrimination (AREA)
Abstract
本申请涉及一种字符识别模型生成方法、装置、计算机设备和存储介质,方法包括:获取多个已识别字符数据集与待识别字符数据集之间的相似度,将与待识别字符数据集之间的相似度匹配的已识别字符数据集,作为目标字符数据集(21);获取与目标字符数据集对应的预训练模型,根据预训练模型构建目标训练模型;预训练模型为经过预先训练后的用于识别目标字符数据集的模型(22);根据已识别字符数据集与待识别字符数据集,生成目标训练数据集;根据目标训练数据集,对目标训练模型进行训练,得到与待识别字符数据集对应的字符识别模型(23)。
Description
本申请要求于2021年07月13日提交中国专利局,申请号为2021107876810,申请名称为“字符识别模型生成方法、装置、计算机设备和存储介质”的中国专利申请的优先权,其全部内容通过引用结合在本申请中。
本申请涉及计算机识别技术领域,特别是涉及一种字符识别模型生成方法、装置、计算机设备和存储介质。
随着工业的发展,越来越多的生产场景开始通过字符识别模型来识别生产设备、生产产品等上的字符信息。
但是,针对不同的生产场景往往需要从头开始训练字符识别模型,字符识别模型训练的周期较长,训练所需的数据较多,字符识别生成的效率还较低。
发明内容
根据本申请的各种实施例,提供一种字符识别模型生成方法、装置、计算机设备和存储介质。
一种字符识别模型生成方法,包括:
获取多个已识别字符数据集与待识别字符数据集之间的相似度,将与所述待识别字符数据集之间的相似度匹配的已识别字符数据集,作为目标字符数据集;
获取与所述目标字符数据集对应的预训练模型,根据所述预训练模型构建目标训练模型;所述预训练模型为经过预先训练后的用于识别所述目标字符数据集的模型;
根据所述已识别字符数据集与所述待识别字符数据集,生成目标训练数据集;根据所述目标训练数据集,对所述目标训练模型进行训练,得到与所述待识别字符数据集对应的字符识别模型。
一种字符识别模型生成装置,所述装置包括:
相似度获取模块,用于获取多个已识别字符数据集与待识别字符数据集之间的相似度,将与所述待识别字符数据集之间的相似度匹配的已识别字符数据集,作为目标字符数据集;
模型构建模块,用于获取与所述目标字符数据集对应的预训练模型,根据所述预训练模型构建目标训练模型;所述预训练模型为经过预先训练后的用于识别所述目标字符数据集的模型;
模型生成模块,用于根据所述已识别字符数据集与所述待识别字符数据集,生成目标训练数据集;根据所述目标训练数据集,对所述目标训练模型进行训练,得到与所述待识别字符数据集对应的字符识别模型。
一种计算机设备,包括存储器和处理器,所述存储器存储有计算机可读指令,所述处理器执行所述计算机可读指令时实现以下步骤:
获取多个已识别字符数据集与待识别字符数据集之间的相似度,将与所述待识别字符数据集之间的相似度匹配的已识别字符数据集,作为目标字符数据集;
获取与所述目标字符数据集对应的预训练模型,根据所述预训练模型构建目标训练模型;所述预训练模型为经过预先训练后的用于识别所述目标字符数据集的模型;
根据所述已识别字符数据集与所述待识别字符数据集,生成目标训练数据集;根据所述目标训练数据集,对所述目标训练模型进行训练,得到与所述待识别字符数据集对应的字符识别模型。
一种计算机可读存储介质,其上存储有计算机可读指令,所述计算机可读指令被处理器执行时实现以下步骤:
获取多个已识别字符数据集与待识别字符数据集之间的相似度,将与所述待识别字符数据集之间的相似度匹配的已识别字符数据集,作为目标字符数据集;
获取与所述目标字符数据集对应的预训练模型,根据所述预训练模型构建目标训练模型;所述预训练模型为经过预先训练后的用于识别所述目标字符数据集的模型;
根据所述已识别字符数据集与所述待识别字符数据集,生成目标训练数据集;根据所述目标训练数据集,对所述目标训练模型进行训练,得到与所述待识别字符数据集对应的字符识别模型。
本申请的一个或多个实施例的细节在下面的附图和描述中提出。本申请的其它特征、目的和优点将从说明书、附图以及权利要求书变得明显。
为了更清楚地说明本申请实施例中的技术方案,下面将对实施例中所需要使用的附图作简单地介绍,显而易见地,下面描述中的附图仅仅是本申请的一些实施例,对于本领域普通技术人员来讲,在不付出创造性劳动的前提下,还可以根据这些附图获得其它的附图。
图1为一个或多个实施例中字符识别模型生成方法的应用环境图;
图2为一个或多个实施例中字符识别模型生成方法的流程示意图;
图3为一个或多个实施例中获取多个已识别字符数据集与待识别字符数据集之间的相似度步骤的流程示意图;
图4为一个或多个实施例中得到已识别字符数据集与待识别字符数据集之间的图片参数相似度步骤的流程示意图;
图5为一个或多个实施例中得到已识别字符数据集与待识别字符数据集之间的文本 轮廓相似度步骤的流程示意图;
图6为一个或多个实施例中字符识别模型生成装置的结构框图;
图7为一个或多个实施例中计算机设备的内部结构图。
为了使本申请的目的、技术方案及优点更加清楚明白,以下结合附图及实施例,对本申请进行进一步详细说明。应当理解,此处描述的具体实施例仅仅用以解释本申请,并不用于限定本申请。
本申请提供的字符识别模型生成方法,可以应用于如图1所示的应用环境中。其中,终端11通过网络与服务器12进行通信。服务器12通过网络从终端11处获取待识别字符数据集,并获取多个已识别字符数据集与待识别字符数据集之间的相似度,服务器12将与待识别字符数据集之间的相似度匹配的已识别字符数据集,作为目标字符数据集;服务器12获取与目标字符数据集对应的预训练模型,服务器12根据预训练模型构建目标训练模型;预训练模型为经过预先训练后的用于识别目标字符数据集的模型;服务器12根据已识别字符数据集与待识别字符数据集,生成目标训练数据集;服务器12根据目标训练数据集,对目标训练模型进行训练,得到与待识别字符数据集对应的字符识别模型。终端11可以直接上传待识别的字符数据至服务器12,使得服务器运行训练完成后的目标训练模型对字符数据进行识别并向终端11返回识别结果;服务器12也可以将训练完成后的目标训练模型转移至终端11中,使得终端11可以直接对字符数据进行识别。
其中,终端11可以但不限于是各种个人计算机、笔记本电脑、智能手机、平板电脑和便携式可穿戴设备,服务器12可以用独立的服务器或者是多个服务器组成的服务器集群来实现。
在一个实施例中,如图2所示,提供了一种字符识别模型生成方法,以该方法应用于图1中的服务器12为例进行说明,包括以下步骤:
步骤21,获取多个已识别字符数据集与待识别字符数据集之间的相似度,将与待识别字符数据集之间的相似度匹配的已识别字符数据集,作为目标字符数据集。
其中,已识别字符数据集是指之前已经利用该字符数据集对模型进行训练后,能够被经过预先训练的模型所识别的字符数据集;待识别字符数据集是指还未被识别,或者未应用于模型训练的字符数据集。字符数据集可以由多张包含有字符的图片组成,图片的文件格式、尺寸等不做限制。
其中,相似度是指通过预设条件判断出的已识别字符数据集与待识别字符数据集之间整体的相似程度,字符数据集的相似度有多种判断方式,也可以由多种判断方式组合加权确定;例如字符数据集中图片数量、尺寸、格式、颜色等都可以作为相似度的确定依据,同理,图片中字符的形状、轮廓、颜色等同样可以作为相似度的确定依据。
其中,相似度匹配是指已识别字符数据集与待识别字符数据集之间的相似度落在预先 设置的相似度区间内,即两者的相似度达到某一预设条件即可视作相似度匹配。
其中,目标字符数据集是指与待识别字符数据集的相似度满足匹配条件的已识别字符数据集,可以是一个或多个。
具体地,服务器获取到待识别字符数据集后,对待识别字符数据集进行数据分析、处理,得到一个可以用于进行相似度判断的数值;服务器再获取至少一个之前已经用于模型训练的已识别字符数据集,利用同样的分析、处理方式,同样得到一个可以用于进行相似度判断的数值;将待识别字符数据集的数值与至少一个已识别字符数据集的数值进行比较,得到多个相似度;从多个相似度中选出至少一个相似度达到预设标准的相似度,将该相似度对应的已识别字符数据集作为目标字符数据集。
步骤22,获取与目标字符数据集对应的预训练模型,根据预训练模型构建目标训练模型;预训练模型为经过预先训练后的用于识别目标字符数据集的模型。
其中,预训练模型是预先经过目标字符数据集(已识别字符数据集)训练后的模型,该模型中的权重参数在利用目标字符数据集(已识别字符数据集)进行训练时已经调整过,能够对目标字符数据集(已识别字符数据集)中的字符数据进行识别。
具体地,服务器获取与目标字符数据集对应的预训练模型,根据该预训练模型的神经网络结构、权重参数等构建得到目标训练模型;构建过程中权重参数的大小以及神经网络结构等可以适应调整。
步骤23,根据已识别字符数据集与待识别字符数据集,生成目标训练数据集;根据目标训练数据集,对目标训练模型进行训练,得到与待识别字符数据集对应的字符识别模型。
其中,目标训练数据集是根据已识别字符数据集与待识别字符数据集生成的,例如随机从已识别字符数据集中取出字符数据,与全部待识别字符数据集以2:8的比例,组成新的字符数据集作为目标训练数据集;具体字符数据的获取方式、获取数量、组合比例等不做限制。
具体地,服务器根据已识别字符数据集与待识别字符数据集,生成目新的字符数据集作为目标训练数据集;将目标训练数据集作为目标训练模型的训练数据,对目标训练模型进行训练,当目标训练模型的损失值达到预设标准,或者目标训练模型对待识别字符数据集的识别准确率达到预设标准后,确定目标训练模型训练完成,得到与待识别字符数据集对应的字符识别模型。
举例说明,例如某公司之前为了解决a场景下字符识别的问题,获取了大量的a场景下的包含有字符数据的图片形成字符数据集A,并且利用该字符数据集A训练得到的字符识别模型A’。现在需要解决b场景下字符识别的问题,获取了大量的b场景下的包含有字符数据的图片形成字符数据集B。并且,a场景与b场景十分类似,可以直接将字符识别模型A’应用于识别字符数据集B;即字符数据集A可以作为已识别字符数据集中的一个字符数据集,字符数据集B作为待识别字符数据集;若确定字符数据集A与字符数据集B 的相似度匹配,则确认字符数据集A是目标字符数据集。可以利用字符识别模型A’构建新的字符数据模型B’,或者直接将字符识别模型A’作为字符数据模型B’进行训练;目标训练数据集可以根据字符数据集A、字符数据集B得到,也可以只利用字符数据集B;即直接利用字符数据集B训练字符识别模型A’也是可行的。
上述字符识别模型生成方法、装置、计算机设备和存储介质,方法包括:获取多个已识别字符数据集与待识别字符数据集之间的相似度,将与待识别字符数据集之间的相似度匹配的已识别字符数据集,作为目标字符数据集;获取与目标字符数据集对应的预训练模型,根据预训练模型构建目标训练模型;预训练模型为经过预先训练后的用于识别目标字符数据集的模型;根据已识别字符数据集与待识别字符数据集,生成目标训练数据集;根据目标训练数据集,对目标训练模型进行训练,得到与待识别字符数据集对应的字符识别模型。本申请通过已识别字符数据集与待识别字符数据集之间的相似度,将经过已识别字符数据集的预训练模型作为待识别字符数据集的训练模型进行训练,实现了根据字符数据集之间的相似度对模型训练进行学习迁移,从而提高了字符识别模型的生成效率。
在一个实施例中,已识别字符数据集与待识别字符数据集均携带有图片参数信息以及文本轮廓信息。如图3所示,步骤21,获取多个已识别字符数据集与待识别字符数据集之间的相似度,包括:
步骤31,根据已识别字符数据集携带的图片参数信息,以及待识别字符数据集携带的图片参数信息,得到已识别字符数据集与待识别字符数据集之间的图片参数相似度;
步骤32,根据已识别字符数据集携带的文本轮廓信息,以及待识别字符数据集携带的文本轮廓信息,得到已识别字符数据集与待识别字符数据集之间的文本轮廓相似度;
步骤33,将图片参数相似度与文本轮廓相似度进行加权处理,根据加权处理的结果确定已识别字符数据集与待识别字符数据集之间的相似度。
其中,图片参数信息可以包括字符数据集中图片的数量、尺寸、文件格式、像素、颜色等;文本轮廓信息包括文本轮廓、几何凸包等。
其中,图片参数相似度是指已识别字符数据集与待识别字符数据集之间图片参数信息的相似程度;例如已识别字符数据集中5张图片的尺寸都是10×10,待识别字符数据集中6张图片的尺寸都是10×10;若只考虑尺寸作为相似度判断条件,则两字符数据集的图片参数相似度可以为100%。
其中,文本轮廓相似度是指已识别字符数据集与待识别字符数据集之间文本轮廓信息的相似程度;例如已识别字符数据集中5张图片中包含字符区域的形状的尺寸为a,待识别字符数据集中6张图片中包含字符区域的形状的尺寸为a的尺寸都是10×10;若只考虑尺寸作为相似度判断条件,则两字符数据集的图片参数相似度可以为100%。
具体地,若图片参数相似度为c,文本轮廓相似度为d,服务器可以设置图片参数相似度的权重为0.6,文本轮廓相似度的权重为0.4,则加权处理后的相似度可以是0.6c+0.4d。
本实施例中,通过分别计算图片参数相似度与文本轮廓相似度,能够达到准确计算已 识别字符数据集与待识别字符数据集之间的相似度的效果,提高了字符识别模型的生成效率。
在一个实施例中,如图4所示,步骤31,根据已识别字符数据集携带的图片参数信息,以及待识别字符数据集携带的图片参数信息,得到已识别字符数据集与待识别字符数据集之间的图片参数相似度,包括:
步骤41,获取已识别字符数据集与待识别字符数据集中,所有图像的色彩通道均值以及宽高均值;
步骤42,根据色彩通道均值和宽高均值,确定已识别字符数据集与待识别字符数据集的色彩通道均值的余弦距离,以及已识别字符数据集与待识别字符数据集的宽高均值的余弦距离;
步骤43,将色彩通道均值的余弦距离与宽高均值的余弦距离之和,作为图片参数相似度。
具体地,服务器获取两个字符数据集中图像像素点的RGB均值作为色彩通道均值;RGB均值包括图像中所有像素点在各颜色通道的色值平均值。服务器获取两个字符数据集中图像的宽度高度均值作为宽高均值。
图片参数相似度通过以下方式获取:
Ai=cos(rgb[P],rgb[Qi])+cos(hw[P],hw[Qi]);
其中,P为待识别字符数据集,Qi为已识别字符数据集,Ai为色彩通道均值的余弦距离与宽高均值的余弦距离之和;rgb[P]为待识别字符数据集的色彩通道均值,rgb[Qi]为已识别字符数据集的色彩通道均值;hw[P]为待识别字符数据集的宽高均值,hw[Qi]为已识别字符数据集的宽高均值,cos为余弦距离计算符号。
本实施例中,通过分别计算色彩通道均值和宽高均值,得到已识别字符数据集与待识别字符数据集之间的图片参数相似度,能够达到准确计算已识别字符数据集与待识别字符数据集之间的相似度的效果,提高了字符识别模型的生成效率。
在一个实施例中,如图5所示,步骤32,根据已识别字符数据集携带的文本轮廓信息,以及待识别字符数据集携带的文本轮廓信息,得到已识别字符数据集与待识别字符数据集之间的文本轮廓相似度,包括:
步骤51,识别已识别字符数据集与待识别字符数据集中,文本信息对应的轮廓特征信息;
步骤52,根据轮廓特征信息确定出已识别字符数据集与待识别字符数据集中,文本信息的凸包信息;
步骤53,根据凸包信息,确定出已识别字符数据集与待识别字符数据集中,文本信息的凸包面积以及文本信息的凸包重叠面积;
步骤54,根据文本信息的凸包面积以及文本信息的凸包重叠面积,得到已识别字符数据集与待识别字符数据集的文本轮廓相似度。
具体地,服务器可以采用轮廓匹配算法查找感兴趣区域中与文本信息作为轮廓特征信息;轮廓匹配算法可以将已识别字符数据集与待识别字符数据集中的图像划分为多个通道,以查找文本信息边缘并转化为轮廓特征,并进行轮廓匹配;其中,轮廓匹配方法可以包括凸包等,具体不做限定。
例如,轮廓特征信息包括检测出的文本轮廓;当采用凸包进行计算时,可以首先对已识别字符数据集与待识别字符数据集中每个文本轮廓分别进行平移,使得每个文本轮廓中的外接矩形的左上角和预设坐标系原点重合。将平移后的文本轮廓绘制在同一平面上,使用opencv库中的凸包检测算法找到这些文本轮廓的最小几何凸包,记为convexhull[P](待识别字符数据集P中文本轮廓的最小几何凸包)。使用同样的方法,得到已识别字符数据集中所有文本的几何凸包,记为convexhull[Qi](已识别字符数据集Qi中文本轮廓的最小几何凸包)。
文本轮廓相似度定义为这两个文本几何凸包的交并比:
Bi=S[inter]/(S[convexhull[P]]+S[convexhull[Qi]]–S[inter])。
其中,Bi为已识别字符数据集与待识别字符数据集的文本轮廓相似度;S[convexhull[P]]和S[convexhull[Qi]]分别为待识别字符数据集与已识别字符数据集中,文本凸包的面积,包含于凸包信息中;S[inter]为待识别字符数据集与已识别字符数据集中文本凸包交叠区域的面积,即凸包重叠面积。
本实施例中,通过分别计算色彩通道均值和宽高均值,得到已识别字符数据集与待识别字符数据集之间的图片参数相似度,能够达到准确计算已识别字符数据集与待识别字符数据集之间的相似度的效果,提高了字符识别模型的生成效率。
在一个实施例中,将与待识别字符数据集之间的相似度匹配的已识别字符数据集,作为目标字符数据集,包括:从多个已识别字符数据集的相似度中筛选出至少一个相似度大于或等于预设相似度阈值的已识别字符数据集;将筛选后得到的至少一个已识别字符数据集的相似度中,相似度最高的已识别字符数据集,识别为目标字符数据集。
具体地,服务器可以设置预设相似度阈值作为判断相似度是否匹配的标准;例如设置预设相似度阈值为85,则计算出的相似度大于或等于85的可以已识别字符数据集可以作为目标字符数据集。预设相似度阈值还可以设置为一个范围,例如85,±5的相似度都可以视作匹配,即相似度范围落在80-90及90以上的相似度应的已识别字符数据集可作为目标字符数据集。
本实施例中,通过设置预设相似度阈值作为目标字符数据集的识别依据,提高了字符识别模型的生成效率。
在一个实施例中,根据预训练模型构建目标训练模型,包括:获取预训练模型的模型参数;将模型参数应用于预先构建的神经网络模型,得到目标训练模型。
具体地,服务器获取预训练模型的模型参数,将该模型参数设置在预先构建的神经网络模型中,作为目标训练模型。
本实施例中,通过将模型参数应用于预先构建的神经网络模型,使得应用了模型参数的预先构建的神经网络模型能够在具备一定识别能力的基础上进行训练,提高了字符识别模型的生成效率。
在一个实施例中,根据目标训练数据集,对目标训练模型进行训练之前,还包括:对目标训练数据集中的图像进行伽马变换处理以及直方图均衡处理;对经过伽马变换处理以及直方图均衡处理的图像进行图像尺寸统一处理;将尺寸统一处理后的图像输入目标训练模型中进行训练。
具体地,伽马变换处理就是通过非线性变换,让目标训练数据集中图像较暗的区域的灰度值得到增强,图像中灰度值过大的区域的灰度值得到降低;经过伽马变换处理,增强了图像整体的细节表现。直方图均衡处理能够增加图像的全局对比度,尤其是当目标训练数据集中图像的有用数据的对比度接近的时候;经过直方图均衡处理后,亮度可以更好地在直方图上分布。尺寸统一处理能够将目标训练数据集中不同大小的图像,统一为同样大小的图像,提高图像的处理效率。
本实施例中,通过对目标训练数据集中的图像进行图像处理,提高了字符识别模型的生成效率。
应该理解的是,虽然图2-5的流程图中的各个步骤按照箭头的指示依次显示,但是这些步骤并不是必然按照箭头指示的顺序依次执行。除非本文中有明确的说明,这些步骤的执行并没有严格的顺序限制,这些步骤可以以其它的顺序执行。而且,图2-5中的至少一部分步骤可以包括多个步骤或者多个阶段,这些步骤或者阶段并不必然是在同一时刻执行完成,而是可以在不同的时刻执行,这些步骤或者阶段的执行顺序也不必然是依次进行,而是可以与其它步骤或者其它步骤中的步骤或者阶段的至少一部分轮流或者交替地执行。
在一个实施例中,如图6所示,提供了一种字符识别模型生成装置,包括相似度获取模块61、模型构建模块62以及模型生成模块63,其中:
相似度获取模块61,用于获取多个已识别字符数据集与待识别字符数据集之间的相似度,将与待识别字符数据集之间的相似度匹配的已识别字符数据集,作为目标字符数据集;
模型构建模块62,用于获取与目标字符数据集对应的预训练模型,根据预训练模型构建目标训练模型;预训练模型为经过预先训练后的用于识别目标字符数据集的模型;
模型生成模块63,用于根据已识别字符数据集与待识别字符数据集,生成目标训练数据集;根据目标训练数据集,对目标训练模型进行训练,得到与待识别字符数据集对应的字符识别模型。
在一个实施例中,已识别字符数据集与待识别字符数据集均携带有图片参数信息以及文本轮廓信息;
相似度获取模块61,还用于根据已识别字符数据集携带的图片参数信息,以及待识别字符数据集携带的图片参数信息,得到已识别字符数据集与待识别字符数据集之间的图 片参数相似度;根据已识别字符数据集携带的文本轮廓信息,以及待识别字符数据集携带的文本轮廓信息,得到已识别字符数据集与待识别字符数据集之间的文本轮廓相似度;将图片参数相似度与文本轮廓相似度进行加权处理,根据加权处理的结果确定已识别字符数据集与待识别字符数据集之间的相似度。
在一个实施例中,相似度获取模块61,还用于获取已识别字符数据集与待识别字符数据集中,所有图像的色彩通道均值以及宽高均值;根据色彩通道均值和宽高均值,确定已识别字符数据集与待识别字符数据集的色彩通道均值的余弦距离,以及已识别字符数据集与待识别字符数据集的宽高均值的余弦距离;将色彩通道均值的余弦距离与宽高均值的余弦距离之和,作为图片参数相似度。
在一个实施例中,相似度获取模块61,还用于识别已识别字符数据集与待识别字符数据集中,文本信息对应的轮廓特征信息;根据轮廓特征信息确定出已识别字符数据集与待识别字符数据集中,文本信息的凸包信息;根据凸包信息,确定出已识别字符数据集与待识别字符数据集中,文本信息的凸包面积以及文本信息的凸包重叠面积;根据文本信息的凸包面积以及文本信息的凸包重叠面积,得到已识别字符数据集与待识别字符数据集的文本轮廓相似度。
在一个实施例中,相似度获取模块61,还用于从多个已识别字符数据集的相似度中筛选出至少一个相似度大于或等于预设相似度阈值的已识别字符数据集;将筛选后得到的至少一个已识别字符数据集的相似度中,相似度最高的已识别字符数据集,识别为目标字符数据集。
在一个实施例中,模型构建模块62,还用于获取预训练模型的模型参数;将模型参数应用于预先构建的神经网络模型,得到目标训练模型。
在一个实施例中,模型生成模块63,还用于对目标训练数据集中的图像进行伽马变换处理以及直方图均衡处理;对经过伽马变换处理以及直方图均衡处理的图像进行图像尺寸统一处理;将尺寸统一处理后的图像输入目标训练模型中进行训练。
关于字符识别模型生成装置的具体限定可以参见上文中对于字符识别模型生成方法的限定,在此不再赘述。上述字符识别模型生成装置中的各个模块可全部或部分通过软件、硬件及其组合来实现。上述各模块可以硬件形式内嵌于或独立于计算机设备中的处理器中,也可以以软件形式存储于计算机设备中的存储器中,以便于处理器调用执行以上各个模块对应的操作。
在一个实施例中,提供了一种计算机设备,该计算机设备可以是服务器,其内部结构图可以如图7所示。该计算机设备包括通过***总线连接的处理器、存储器和网络接口。其中,该计算机设备的处理器用于提供计算和控制能力。该计算机设备的存储器包括非易失性存储介质、内存储器。该非易失性存储介质存储有操作***、计算机可读指令和数据库。该内存储器为非易失性存储介质中的操作***和计算机可读指令的运行提供环境。该计算机设备的数据库用于存储字符识别模型生成数据。该计算机设备的网络接口用于与外 部的终端通过网络连接通信。该计算机可读指令被处理器执行时以实现一种字符识别模型生成法。
本领域技术人员可以理解,图7中示出的结构,是与本申请方案相关的部分结构的框图,并不构成对本申请方案所应用于其上的计算机设备的限定,具体的计算机设备可以包括比图中所示更多或更少的部件,或者组合某些部件,或者具有不同的部件布置。
在一个实施例中,提供了一种计算机设备,包括存储器和处理器,存储器中存储有计算机可读指令,该处理器执行计算机可读指令时实现以下步骤:
获取多个已识别字符数据集与待识别字符数据集之间的相似度,将与待识别字符数据集之间的相似度匹配的已识别字符数据集,作为目标字符数据集;
获取与目标字符数据集对应的预训练模型,根据预训练模型构建目标训练模型;预训练模型为经过预先训练后的用于识别目标字符数据集的模型;
根据已识别字符数据集与待识别字符数据集,生成目标训练数据集;根据目标训练数据集,对目标训练模型进行训练,得到与待识别字符数据集对应的字符识别模型。
在一个实施例中,已识别字符数据集与待识别字符数据集均携带有图片参数信息以及文本轮廓信息;处理器执行计算机可读指令时还实现以下步骤:根据已识别字符数据集携带的图片参数信息,以及待识别字符数据集携带的图片参数信息,得到已识别字符数据集与待识别字符数据集之间的图片参数相似度;根据已识别字符数据集携带的文本轮廓信息,以及待识别字符数据集携带的文本轮廓信息,得到已识别字符数据集与待识别字符数据集之间的文本轮廓相似度;将图片参数相似度与文本轮廓相似度进行加权处理,根据加权处理的结果确定已识别字符数据集与待识别字符数据集之间的相似度。
在一个实施例中,处理器执行计算机可读指令时还实现以下步骤:获取已识别字符数据集与待识别字符数据集中,所有图像的色彩通道均值以及宽高均值;根据色彩通道均值和宽高均值,确定已识别字符数据集与待识别字符数据集的色彩通道均值的余弦距离,以及已识别字符数据集与待识别字符数据集的宽高均值的余弦距离;将色彩通道均值的余弦距离与宽高均值的余弦距离之和,作为图片参数相似度。
在一个实施例中,处理器执行计算机可读指令时还实现以下步骤:识别已识别字符数据集与待识别字符数据集中,文本信息对应的轮廓特征信息;根据轮廓特征信息确定出已识别字符数据集与待识别字符数据集中,文本信息的凸包信息;根据凸包信息,确定出已识别字符数据集与待识别字符数据集中,文本信息的凸包面积以及文本信息的凸包重叠面积;根据文本信息的凸包面积以及文本信息的凸包重叠面积,得到已识别字符数据集与待识别字符数据集的文本轮廓相似度。
在一个实施例中,处理器执行计算机可读指令时还实现以下步骤:从多个已识别字符数据集的相似度中筛选出至少一个相似度大于或等于预设相似度阈值的已识别字符数据集;将筛选后得到的至少一个已识别字符数据集的相似度中,相似度最高的已识别字符数据集,识别为目标字符数据集。
在一个实施例中,处理器执行计算机可读指令时还实现以下步骤:获取预训练模型的模型参数;将模型参数应用于预先构建的神经网络模型,得到目标训练模型。
在一个实施例中,处理器执行计算机可读指令时还实现以下步骤:对目标训练数据集中的图像进行伽马变换处理以及直方图均衡处理;对经过伽马变换处理以及直方图均衡处理的图像进行图像尺寸统一处理;将尺寸统一处理后的图像输入目标训练模型中进行训练。
在一个实施例中,提供了一种计算机可读存储介质,其上存储有计算机可读指令,计算机可读指令被处理器执行时实现以下步骤:
获取多个已识别字符数据集与待识别字符数据集之间的相似度,将与待识别字符数据集之间的相似度匹配的已识别字符数据集,作为目标字符数据集;
获取与目标字符数据集对应的预训练模型,根据预训练模型构建目标训练模型;预训练模型为经过预先训练后的用于识别目标字符数据集的模型;
根据已识别字符数据集与待识别字符数据集,生成目标训练数据集;根据目标训练数据集,对目标训练模型进行训练,得到与待识别字符数据集对应的字符识别模型。
在一个实施例中,已识别字符数据集与待识别字符数据集均携带有图片参数信息以及文本轮廓信息;计算机可读指令被处理器执行时还实现以下步骤:根据已识别字符数据集携带的图片参数信息,以及待识别字符数据集携带的图片参数信息,得到已识别字符数据集与待识别字符数据集之间的图片参数相似度;根据已识别字符数据集携带的文本轮廓信息,以及待识别字符数据集携带的文本轮廓信息,得到已识别字符数据集与待识别字符数据集之间的文本轮廓相似度;将图片参数相似度与文本轮廓相似度进行加权处理,根据加权处理的结果确定已识别字符数据集与待识别字符数据集之间的相似度。
在一个实施例中,计算机可读指令被处理器执行时还实现以下步骤:获取已识别字符数据集与待识别字符数据集中,所有图像的色彩通道均值以及宽高均值;根据色彩通道均值和宽高均值,确定已识别字符数据集与待识别字符数据集的色彩通道均值的余弦距离,以及已识别字符数据集与待识别字符数据集的宽高均值的余弦距离;将色彩通道均值的余弦距离与宽高均值的余弦距离之和,作为图片参数相似度。
在一个实施例中,计算机可读指令被处理器执行时还实现以下步骤:识别已识别字符数据集与待识别字符数据集中,文本信息对应的轮廓特征信息;根据轮廓特征信息确定出已识别字符数据集与待识别字符数据集中,文本信息的凸包信息;根据凸包信息,确定出已识别字符数据集与待识别字符数据集中,文本信息的凸包面积以及文本信息的凸包重叠面积;根据文本信息的凸包面积以及文本信息的凸包重叠面积,得到已识别字符数据集与待识别字符数据集的文本轮廓相似度。
在一个实施例中,计算机可读指令被处理器执行时还实现以下步骤:从多个已识别字符数据集的相似度中筛选出至少一个相似度大于或等于预设相似度阈值的已识别字符数据集;将筛选后得到的至少一个已识别字符数据集的相似度中,相似度最高的已识别字符 数据集,识别为目标字符数据集。
在一个实施例中,计算机可读指令被处理器执行时还实现以下步骤:获取预训练模型的模型参数;将模型参数应用于预先构建的神经网络模型,得到目标训练模型。
在一个实施例中,计算机可读指令被处理器执行时还实现以下步骤:对目标训练数据集中的图像进行伽马变换处理以及直方图均衡处理;对经过伽马变换处理以及直方图均衡处理的图像进行图像尺寸统一处理;将尺寸统一处理后的图像输入目标训练模型中进行训练。
本领域普通技术人员可以理解实现上述实施例方法中的全部或部分流程,是可以通过计算机可读指令来指令相关的硬件来完成,上述的计算机可读指令可存储于一非易失性计算机可读取存储介质中,该计算机可读指令在执行时,可包括如上述各方法的实施例的流程。其中,本申请所提供的各实施例中所使用的对存储器、存储、数据库或其它介质的任何引用,均可包括非易失性和易失性存储器中的至少一种。非易失性存储器可包括只读存储器(Read-Only Memory,ROM)、磁带、软盘、闪存或光存储器等。易失性存储器可包括随机存取存储器(Random Access Memory,RAM)或外部高速缓冲存储器。作为说明而非局限,RAM可以是多种形式,比如静态随机存取存储器(Static Random Access Memory,SRAM)或动态随机存取存储器(Dynamic Random Access Memory,DRAM)等。
以上实施例的各技术特征可以进行任意的组合,为使描述简洁,未对上述实施例中的各个技术特征所有可能的组合都进行描述,然而,只要这些技术特征的组合不存在矛盾,都应当认为是本说明书记载的范围。
以上各个实施例仅表达了本申请的几种实施方式,其描述较为具体和详细,但并不能因此而理解为对本申请范围的限制。应当指出的是,对于本领域的普通技术人员来说,在不脱离本申请构思的前提下,还可以做出若干变形和改进,这些都属于本申请的保护范围。因此,本申请专利的保护范围应以所附权利要求为准。
Claims (16)
- 一种字符识别模型生成方法,包括:获取多个已识别字符数据集与待识别字符数据集之间的相似度,将与所述待识别字符数据集之间的相似度匹配的已识别字符数据集,作为目标字符数据集;获取与所述目标字符数据集对应的预训练模型,根据所述预训练模型构建目标训练模型;所述预训练模型为经过预先训练后的用于识别所述目标字符数据集的模型;根据所述已识别字符数据集与所述待识别字符数据集,生成目标训练数据集;根据所述目标训练数据集,对所述目标训练模型进行训练,得到与所述待识别字符数据集对应的字符识别模型。
- 根据权利要求1所述的方法,其特征在于,所述已识别字符数据集与待识别字符数据集均携带有图片参数信息以及文本轮廓信息;所述获取多个已识别字符数据集与待识别字符数据集之间的相似度,包括:根据所述已识别字符数据集携带的图片参数信息,以及所述待识别字符数据集携带的图片参数信息,得到所述已识别字符数据集与所述待识别字符数据集之间的图片参数相似度;根据所述已识别字符数据集携带的文本轮廓信息,以及所述待识别字符数据集携带的文本轮廓信息,得到所述已识别字符数据集与所述待识别字符数据集之间的文本轮廓相似度;将所述图片参数相似度与所述文本轮廓相似度进行加权处理,根据加权处理的结果确定所述已识别字符数据集与所述待识别字符数据集之间的相似度。
- 根据权利要求2所述的方法,其特征在于,所述根据所述已识别字符数据集携带的图片参数信息,以及所述待识别字符数据集携带的图片参数信息,得到所述已识别字符数据集与所述待识别字符数据集之间的图片参数相似度,包括:获取所述已识别字符数据集与待识别字符数据集中,所有图像的色彩通道均值以及宽高均值;根据所述色彩通道均值和所述宽高均值,确定所述已识别字符数据集与待识别字符数据集的色彩通道均值的余弦距离,以及所述已识别字符数据集与待识别字符数据集的宽高均值的余弦距离;将所述色彩通道均值的余弦距离与所述宽高均值的余弦距离之和,作为所述图片参数相似度。
- 根据权利要求2所述的方法,其特征在于,所述根据所述已识别字符数据集携带的文本轮廓信息,以及所述待识别字符数据集携带的文本轮廓信息,得到所述已识别字符数据集与所述待识别字符数据集之间的文本轮廓相似度,包括:识别所述已识别字符数据集与所述待识别字符数据集中,文本信息对应的轮廓特征信息;根据所述轮廓特征信息确定出所述已识别字符数据集与所述待识别字符数据集中,文本信息的凸包信息;根据所述凸包信息,确定出所述已识别字符数据集与所述待识别字符数据集中,文本信息的凸包面积以及文本信息的凸包重叠面积;根据所述文本信息的凸包面积以及文本信息的凸包重叠面积,得到所述已识别字符数据集与所述待识别字符数据集的文本轮廓相似度。
- 根据权利要求1所述的方法,其特征在于,所述将与所述待识别字符数据集之间的相似度匹配的已识别字符数据集,作为目标字符数据集,包括:从多个已识别字符数据集的相似度中筛选出至少一个相似度大于或等于预设相似度阈值的已识别字符数据集;将筛选后得到的至少一个已识别字符数据集的相似度中,相似度最高的已识别字符数据集,识别为所述目标字符数据集。
- 根据权利要求1至5任意一项所述的方法,其特征在于,根据所述预训练模型构建目标训练模型,包括:获取所述预训练模型的模型参数;将所述模型参数应用于预先构建的神经网络模型,得到所述目标训练模型。
- 根据权利要求1至5任意一项所述的方法,其特征在于,在根据所述目标训练数据集,对所述目标训练模型进行训练之前,还包括:对所述目标训练数据集中的图像进行伽马变换处理以及直方图均衡处理;对经过所述伽马变换处理以及直方图均衡处理的图像进行图像尺寸统一处理;将尺寸统一处理后的图像输入所述目标训练模型中进行训练。
- 一种字符识别模型生成装置,所述装置包括:相似度获取模块,用于获取多个已识别字符数据集与待识别字符数据集之间的相似度,将与所述待识别字符数据集之间的相似度匹配的已识别字符数据集,作为目标字符数据集;模型构建模块,用于获取与所述目标字符数据集对应的预训练模型,根据所述预训练模型构建目标训练模型;所述预训练模型为经过预先训练后的用于识别所述目标字符数据集的模型;模型生成模块,用于根据所述已识别字符数据集与所述待识别字符数据集,生成目标训练数据集;根据所述目标训练数据集,对所述目标训练模型进行训练,得到与所述待识别字符数据集对应的字符识别模型。
- 根据权利要求8所述的装置,其特征在于,所述已识别字符数据集与待识别字符数据集均携带有图片参数信息以及文本轮廓信息;所述相似度获取模块,还用于根据所述已识别字符数据集携带的图片参数信息,以及所述待识别字符数据集携带的图片参数信息,得到所述已识别字符数据集与所述待识别字 符数据集之间的图片参数相似度;根据所述已识别字符数据集携带的文本轮廓信息,以及所述待识别字符数据集携带的文本轮廓信息,得到所述已识别字符数据集与所述待识别字符数据集之间的文本轮廓相似度;将所述图片参数相似度与所述文本轮廓相似度进行加权处理,根据加权处理的结果确定所述已识别字符数据集与所述待识别字符数据集之间的相似度。
- 根据权利要求9所述的装置,其特征在于,所述相似度获取模块,还用于获取所述已识别字符数据集与待识别字符数据集中,所有图像的色彩通道均值以及宽高均值;根据所述色彩通道均值和所述宽高均值,确定所述已识别字符数据集与待识别字符数据集的色彩通道均值的余弦距离,以及所述已识别字符数据集与待识别字符数据集的宽高均值的余弦距离;将所述色彩通道均值的余弦距离与所述宽高均值的余弦距离之和,作为所述图片参数相似度。
- 根据权利要求9所述的装置,其特征在于,所述相似度获取模块,还用于识别所述已识别字符数据集与所述待识别字符数据集中,文本信息对应的轮廓特征信息;根据所述轮廓特征信息确定出所述已识别字符数据集与所述待识别字符数据集中,文本信息的凸包信息;根据所述凸包信息,确定出所述已识别字符数据集与所述待识别字符数据集中,文本信息的凸包面积以及文本信息的凸包重叠面积;根据所述文本信息的凸包面积以及文本信息的凸包重叠面积,得到所述已识别字符数据集与所述待识别字符数据集的文本轮廓相似度。
- 根据权利要求8所述的装置,其特征在于,所述相似度获取模块,还用于从多个已识别字符数据集的相似度中筛选出至少一个相似度大于或等于预设相似度阈值的已识别字符数据集;将筛选后得到的至少一个已识别字符数据集的相似度中,相似度最高的已识别字符数据集,识别为所述目标字符数据集。
- 根据权利要求8至12任意一项所述的装置,其特征在于,所述模型构建模块,还用于获取所述预训练模型的模型参数;将所述模型参数应用于预先构建的神经网络模型,得到所述目标训练模型。
- 根据权利要求8至12任意一项所述的装置,其特征在于,所述模型生成模块,还用于对所述目标训练数据集中的图像进行伽马变换处理以及直方图均衡处理;对经过所述伽马变换处理以及直方图均衡处理的图像进行图像尺寸统一处理;将尺寸统一处理后的图像输入所述目标训练模型中进行训练。
- 一种计算机设备,包括存储器和处理器,所述存储器存储有计算机可读指令,其特征在于,所述处理器执行所述计算机可读指令时实现权利要求1至7中任一项所述的方法的步骤。
- 一种计算机可读存储介质,其上存储有计算机可读指令,其特征在于,所述计算机可读指令被处理器执行时实现权利1至7中任一项所述的方法的步骤。
Applications Claiming Priority (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202110787681.0 | 2021-07-13 | ||
CN202110787681.0A CN113469092B (zh) | 2021-07-13 | 2021-07-13 | 字符识别模型生成方法、装置、计算机设备和存储介质 |
Publications (1)
Publication Number | Publication Date |
---|---|
WO2023284608A1 true WO2023284608A1 (zh) | 2023-01-19 |
Family
ID=77879893
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/CN2022/104107 WO2023284608A1 (zh) | 2021-07-13 | 2022-07-06 | 字符识别模型生成方法、装置、计算机设备和存储介质 |
Country Status (2)
Country | Link |
---|---|
CN (1) | CN113469092B (zh) |
WO (1) | WO2023284608A1 (zh) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN116664966A (zh) * | 2023-03-27 | 2023-08-29 | 北京鹰之眼智能健康科技有限公司 | 一种红外图像处理*** |
Families Citing this family (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN113469092B (zh) * | 2021-07-13 | 2023-09-08 | 深圳思谋信息科技有限公司 | 字符识别模型生成方法、装置、计算机设备和存储介质 |
CN113971806B (zh) * | 2021-10-26 | 2023-05-05 | 北京百度网讯科技有限公司 | 一种模型训练、字符识别方法、装置、设备及存储介质 |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108446621A (zh) * | 2018-03-14 | 2018-08-24 | 平安科技(深圳)有限公司 | 票据识别方法、服务器及计算机可读存储介质 |
US10163022B1 (en) * | 2017-06-22 | 2018-12-25 | StradVision, Inc. | Method for learning text recognition, method for recognizing text using the same, and apparatus for learning text recognition, apparatus for recognizing text using the same |
CN112307858A (zh) * | 2019-08-30 | 2021-02-02 | 北京字节跳动网络技术有限公司 | 一种图像识别及处理方法、装置、设备及存储介质 |
CN113469092A (zh) * | 2021-07-13 | 2021-10-01 | 深圳思谋信息科技有限公司 | 字符识别模型生成方法、装置、计算机设备和存储介质 |
Family Cites Families (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20190354850A1 (en) * | 2018-05-17 | 2019-11-21 | International Business Machines Corporation | Identifying transfer models for machine learning tasks |
KR102107444B1 (ko) * | 2018-07-30 | 2020-06-02 | 주식회사 한글과컴퓨터 | 원본 이미지와 생성된 이미지 간의 대조를 통한 문자 인식 장치 및 이의 동작 방법 |
CN109871847B (zh) * | 2019-03-13 | 2022-09-30 | 厦门商集网络科技有限责任公司 | 一种ocr识别方法及终端 |
CN110377587B (zh) * | 2019-07-15 | 2023-02-10 | 腾讯科技(深圳)有限公司 | 基于机器学习的迁移数据确定方法、装置、设备及介质 |
CN111461238B (zh) * | 2020-04-03 | 2024-03-05 | 讯飞智元信息科技有限公司 | 模型训练方法、字符识别方法、装置、设备及存储介质 |
CN111860670B (zh) * | 2020-07-28 | 2022-05-17 | 平安科技(深圳)有限公司 | 域自适应模型训练、图像检测方法、装置、设备及介质 |
CN111738269B (zh) * | 2020-08-25 | 2020-11-20 | 北京易真学思教育科技有限公司 | 模型训练方法、图像处理方法及装置、设备、存储介质 |
CN112465012A (zh) * | 2020-11-25 | 2021-03-09 | 创新奇智(南京)科技有限公司 | 机器学习建模方法、装置、电子设备和可读存储介质 |
-
2021
- 2021-07-13 CN CN202110787681.0A patent/CN113469092B/zh active Active
-
2022
- 2022-07-06 WO PCT/CN2022/104107 patent/WO2023284608A1/zh active Application Filing
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US10163022B1 (en) * | 2017-06-22 | 2018-12-25 | StradVision, Inc. | Method for learning text recognition, method for recognizing text using the same, and apparatus for learning text recognition, apparatus for recognizing text using the same |
CN108446621A (zh) * | 2018-03-14 | 2018-08-24 | 平安科技(深圳)有限公司 | 票据识别方法、服务器及计算机可读存储介质 |
CN112307858A (zh) * | 2019-08-30 | 2021-02-02 | 北京字节跳动网络技术有限公司 | 一种图像识别及处理方法、装置、设备及存储介质 |
CN113469092A (zh) * | 2021-07-13 | 2021-10-01 | 深圳思谋信息科技有限公司 | 字符识别模型生成方法、装置、计算机设备和存储介质 |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN116664966A (zh) * | 2023-03-27 | 2023-08-29 | 北京鹰之眼智能健康科技有限公司 | 一种红外图像处理*** |
CN116664966B (zh) * | 2023-03-27 | 2024-02-20 | 北京鹰之眼智能健康科技有限公司 | 一种红外图像处理*** |
Also Published As
Publication number | Publication date |
---|---|
CN113469092A (zh) | 2021-10-01 |
CN113469092B (zh) | 2023-09-08 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
WO2023284608A1 (zh) | 字符识别模型生成方法、装置、计算机设备和存储介质 | |
CN109492643B (zh) | 基于ocr的证件识别方法、装置、计算机设备及存储介质 | |
WO2020232872A1 (zh) | 表格识别方法、装置、计算机设备和存储介质 | |
WO2022033095A1 (zh) | 一种文本区域的定位方法及装置 | |
CN108229296B (zh) | 人脸皮肤属性识别方法和装置、电子设备、存储介质 | |
WO2019100724A1 (zh) | 训练多标签分类模型的方法和装置 | |
WO2022042123A1 (zh) | 图像识别模型生成方法、装置、计算机设备和存储介质 | |
EP2806374B1 (en) | Method and system for automatic selection of one or more image processing algorithm | |
EP3852003A1 (en) | Feature point locating method, storage medium and computer device | |
US20230021661A1 (en) | Forgery detection of face image | |
US10410354B1 (en) | Method and apparatus for multi-model primitive fitting based on deep geometric boundary and instance aware segmentation | |
CN110930296B (zh) | 图像处理方法、装置、设备及存储介质 | |
CN111310800B (zh) | 图像分类模型生成方法、装置、计算机设备和存储介质 | |
WO2022048151A1 (zh) | 语义分割模型训练方法及装置、图像语义分割方法及装置 | |
US20180253852A1 (en) | Method and device for locating image edge in natural background | |
US9129152B2 (en) | Exemplar-based feature weighting | |
US20190311183A1 (en) | Feature matching with a subspace spanned by multiple representative feature vectors | |
US10373022B1 (en) | Text image processing using stroke-aware max-min pooling for OCR system employing artificial neural network | |
US11893773B2 (en) | Finger vein comparison method, computer equipment, and storage medium | |
CN114677565B (zh) | 特征提取网络的训练方法和图像处理方法、装置 | |
US20230045715A1 (en) | Text detection method, text recognition method and apparatus | |
CN111986117A (zh) | 一种算术作业批改***及方法 | |
US20220207917A1 (en) | Facial expression image processing method and apparatus, and electronic device | |
CN112836682B (zh) | 视频中对象的识别方法、装置、计算机设备和存储介质 | |
CN115713769A (zh) | 文本检测模型的训练方法、装置、计算机设备和存储介质 |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
NENP | Non-entry into the national phase |
Ref country code: DE |
|
122 | Ep: pct application non-entry in european phase |
Ref document number: 22841245 Country of ref document: EP Kind code of ref document: A1 |