WO2021128578A1 - 图像处理方法及装置、电子设备和存储介质 - Google Patents

图像处理方法及装置、电子设备和存储介质 Download PDF

Info

Publication number
WO2021128578A1
WO2021128578A1 PCT/CN2020/077247 CN2020077247W WO2021128578A1 WO 2021128578 A1 WO2021128578 A1 WO 2021128578A1 CN 2020077247 W CN2020077247 W CN 2020077247W WO 2021128578 A1 WO2021128578 A1 WO 2021128578A1
Authority
WO
WIPO (PCT)
Prior art keywords
feature
target
text
extracted
relative position
Prior art date
Application number
PCT/CN2020/077247
Other languages
English (en)
French (fr)
Inventor
孙红斌
岳晓宇
旷章辉
蔺琛皓
张伟
Original Assignee
深圳市商汤科技有限公司
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by 深圳市商汤科技有限公司 filed Critical 深圳市商汤科技有限公司
Priority to KR1020217020203A priority Critical patent/KR20210113192A/ko
Priority to JP2021538344A priority patent/JP7097513B2/ja
Publication of WO2021128578A1 publication Critical patent/WO2021128578A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/20—Image preprocessing
    • G06V10/25—Determination of region of interest [ROI] or a volume of interest [VOI]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/0464—Convolutional networks [CNN, ConvNet]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G06V10/46—Descriptors for shape, contour or point-related descriptors, e.g. scale invariant feature transform [SIFT] or bags of words [BoW]; Salient regional features
    • G06V10/462—Salient features, e.g. scale invariant feature transforms [SIFT]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/764—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using classification, e.g. of video objects
    • 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
    • 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
    • G06V30/40—Document-oriented image-based pattern recognition
    • G06V30/42—Document-oriented image-based pattern recognition based on the type of document

Definitions

  • the present disclosure relates to the field of computer technology, and in particular to an image processing method and device, electronic equipment, and storage medium.
  • the extraction of key text information from images plays a very important role in automated office and other scenarios. For example, by extracting key text information in images, functions such as receipt information extraction, invoice information extraction, and identity information extraction can be realized.
  • the recognized text When extracting the text in the image, the recognized text will be mapped to different fields for subsequent operations such as structured storage and display of the text. For example, if the recognized text is "19.88 yuan”, it is necessary to determine whether “19.88 yuan” corresponds to the field “total price” or the corresponding field “unit price”, so that "19.88 yuan” is subsequently stored as the value of a certain field.
  • a template is defined in advance according to the arrangement rules of the text in the image, and the corresponding relationship between the text at a certain position and the field is defined in the template, so that the field corresponding to the recognized text at a certain position can be determined.
  • the field corresponding to the recognized text at a certain position can be determined. For example, predefine the field corresponding to the text in the lower right corner of the image as “Total Price”, so that it can be determined that the field corresponding to "19.88 Yuan” identified in the lower right corner of the image is “Total Price”.
  • the present disclosure proposes a technical solution for image processing.
  • an image processing method including: recognizing an image, determining a plurality of target regions in the image, where the target region is the region where the text to be extracted is located; determining each of the images in the image The relative position feature between the target areas; determine the target feature of each target area, the target feature includes the feature of the text to be extracted; through the graph convolutional neural network, the relative position feature and the target feature Perform feature extraction to obtain the extracted feature; according to the extracted feature, determine the field corresponding to the text to be extracted.
  • a graph convolutional neural network can be used to determine the field corresponding to the text to be extracted in the image based on the relative position feature between the target regions and the feature of the text to be extracted.
  • the text extraction can be performed without relying on a fixed template. Compared with the method of text extraction based on a template, the accuracy of text extraction is higher when the text is extracted from an image without a suitable template.
  • feature extraction is performed on the relative position feature and the target feature through a graph convolutional neural network to obtain the extracted features, including: taking each target feature as a node of the graph, Use each of the relative position features as the edges connecting two nodes to construct a connected graph; through the graph convolutional neural network, the connected graph is iteratively updated, and the connected graph that meets the convergence condition after the iterative update is used as the extracted feature .
  • the constructed connected graph includes not only the target features in the image, but also the relative position features between the target features in the image, which can characterize the characteristics of the text in the image as a whole, and therefore can improve the key information The accuracy of the extraction results.
  • graph convolutional neural networks can represent images in the form of connected graphs and extract features.
  • Connected graph is composed of several nodes (Node) and edges (Edge) connecting two nodes. Edges are used to describe the relationship between different nodes. Therefore, the features extracted by the graph convolutional neural network can accurately characterize the relative position between the target regions and the features of the text to be extracted, so as to improve the accuracy of subsequent text extraction.
  • determining the field corresponding to the text to be extracted according to the extracted features includes: according to a plurality of pre-defined preset categories, the nodes in the connected graph output by the graph convolutional neural network The classification is performed to obtain the category of the node.
  • the preset category includes: the category of the characterization text belonging to the identifier of the preset field, and the category of the field value of the characterization text belonging to the preset field; according to the category of the node, the text to be extracted is determined The identifier or field value corresponding to the preset field.
  • the identifier or field value of the text to be extracted corresponding to the preset field can be obtained, which improves The accuracy of text extraction is improved.
  • determining the relative position characteristics between the target areas in the image includes: determining the relative position parameters of the first target area and the second target area in the image; Perform characterization processing to obtain the relative position characteristics of the first target area and the second target area.
  • the relative position parameter includes at least one of the following: the lateral distance and the longitudinal distance of the first target area relative to the second target area; the aspect ratio of the first target area; The aspect ratio of the second target area; the relative size relationship between the first target area and the second target area.
  • the relative position parameter includes the horizontal distance and the vertical distance, the aspect ratio of the first target area, and the relative size relationship between the first target area and the second target area, so that The extraction result of key information is more accurate.
  • performing characterization processing on the relative position parameter to obtain the relative position characteristics of the first target area and the second target area includes: mapping the relative position parameter to a sine-cosine transformation matrix A D-dimensional space is used to obtain a D-dimensional eigenvector, where D is a positive integer; the D-dimensional eigenvector is converted into a 1-dimensional weight value by a preset weight matrix; the weight value is calculated by a preset activation function Perform processing to obtain relative position characteristics.
  • the relative position parameter can be converted into the data format required by the edge of the graph convolutional neural network through the feature processing, which is convenient for subsequent feature extraction through the graph convolutional neural network.
  • determining the target feature of each target area includes: determining pixel data in the target area, performing feature extraction on the pixel data to obtain visual features; determining text characters in the target area, Perform feature extraction on the text characters to obtain character features; and determine the target features of the target area according to the extracted visual features and character features.
  • determining the target feature of the target area according to the extracted visual features and character features includes: assigning different weights to the visual features and character features; and assigning weights to the visual features It merges with the character feature to obtain the target feature of the target area.
  • the method is implemented by a pre-built classification network, and the training steps of the classification network are as follows: the sample image is input into the classification network for processing, and the first part of the text to be extracted in the sample image is obtained. Prediction category, and the corresponding relationship between each category in the first prediction category; training the classification network according to the first prediction category and the label category of the sample image, the label category includes: characterization text The category of the identifier belonging to the preset field, and the category of the field value of the characterizing text belonging to the preset field; training the classification network according to the corresponding relationship and the corresponding relationship between the labeled texts to be extracted.
  • the classification network can be trained more accurately by labeling the category of the sample image and the corresponding relationship between each category.
  • the trained classification network performs text extraction on images without a suitable template. When the time, the accuracy is higher.
  • the image includes at least one of the following: a receipt image, an invoice image, and a business card image.
  • an image processing device including: a recognition module for recognizing an image and determining a plurality of target regions in the image, where the target region is the region where the text to be extracted is located;
  • the location feature determination module is used to determine the relative location feature between each target area in the image;
  • the target feature determination module is used to determine the target feature of each target area, the target feature includes the text to be extracted Features;
  • graph convolution module used to extract features from the relative position feature and the target feature through the graph convolutional neural network, to obtain the extracted features; field determination module, used to determine according to the extracted features The field corresponding to the text to be extracted.
  • a graph convolutional neural network can be used to determine the field corresponding to the text to be extracted in the image based on the relative position feature between the target regions and the feature of the text to be extracted.
  • the text extraction can be performed without relying on a fixed template. Compared with the method of text extraction based on a template, the accuracy of text extraction is higher when the text is extracted from an image without a suitable template.
  • the graph convolution module includes: a first graph convolution sub-module and a second graph convolution sub-module, where the first graph convolution sub-module is configured to take each of the target features as The nodes of the graph use each of the relative position features as the edges connecting the two nodes to construct a connected graph; the second graph convolution submodule is used to iteratively update the connected graph through the graph convolutional neural network, and After the iterative update, the connected graph that meets the convergence condition is used as the extracted feature.
  • the constructed connected graph includes not only the target features in the image, but also the relative position features between the target features in the image, which can characterize the characteristics of the text in the image as a whole, and therefore can improve the key information The accuracy of the extraction results.
  • graph convolutional neural networks can represent images in the form of connected graphs and extract features.
  • a connected graph is composed of several nodes and edges connecting two nodes. The edges are used to describe the relationship between different nodes. Therefore, the features extracted by the graph convolutional neural network can accurately characterize the relative position between the target regions and the features of the text to be extracted, so as to improve the accuracy of subsequent text extraction.
  • the field determination module includes: a first field determination sub-module and a second field determination sub-module, wherein the first field determination sub-module is configured to perform the The nodes in the connected graph output by the graph convolutional neural network are classified to obtain the category of the node.
  • the preset category includes: the category of the identifier representing the text belonging to the preset field, and the category of the field value representing the text belonging to the preset field ;
  • the second field determination sub-module is used to determine the identifier or field value of the preset field corresponding to the text to be extracted according to the category of the node.
  • the identifier or field value of the text to be extracted corresponding to the preset field can be obtained, which improves The accuracy of text extraction is improved.
  • the relative position feature determination module includes: a first relative position feature determination sub-module and a second relative position feature determination sub-module, wherein the first relative position feature determination sub-module is used to determine The relative position parameters of the first target area and the second target area; the second relative position feature determination sub-module is used to characterize the relative position parameters to obtain the relative positions of the first target area and the second target area feature.
  • the relative position parameter includes at least one of the following: the lateral distance and the longitudinal distance of the first target area relative to the second target area; the aspect ratio of the first target area; The aspect ratio of the second target area; the relative size relationship between the first target area and the second target area.
  • the relative position parameter includes the horizontal distance and the vertical distance, the aspect ratio of the first target area, and the relative size relationship between the first target area and the second target area, so that The extraction result of key information is more accurate.
  • the second relative position feature determination submodule is used to map the relative position parameter to a D-dimensional space through a sine-cosine transformation matrix to obtain a D-dimensional eigenvector, where D is a positive integer Transform the D-dimensional feature vector into a 1-dimensional weight value through a preset weight matrix; process the weight value through a preset activation function to obtain a relative position feature.
  • the relative position parameter can be converted into the data format required by the edge of the graph convolutional neural network through the feature processing, which is convenient for subsequent feature extraction through the graph convolutional neural network.
  • the target feature determination module includes a first target feature determination sub-module, a second target feature determination sub-module, and a third target feature determination sub-module, wherein the first target feature determination sub-module uses To determine the pixel data in the target area, perform feature extraction on the pixel data to obtain visual features; the second target feature determination submodule is used to determine the text characters in the target area, and perform feature extraction on the text characters to obtain Character feature; the third target feature determination sub-module is used to determine the target feature of the target area according to the extracted visual features and character features.
  • the third target feature determination submodule is used to assign different weights to the visual features and character features; to fuse the weighted visual features and character features to obtain the target area Target characteristics.
  • the device is implemented by a pre-built classification network, and the device further includes: a first training module for inputting sample images into the classification network for processing to obtain sample images to be extracted The first prediction category of the text, and the correspondence between each category in the first prediction category; the second training module is used to train the first prediction category and the label category of the sample image
  • the classification network includes: the characterization text belongs to the identification category of the preset field, and the characterization text belongs to the field value category of the preset field; the third training module is used for according to the corresponding relationship and the label to be marked. The correspondence between the texts is extracted, and the classification network is trained.
  • the classification network can be trained more accurately by labeling the classification of the sample image and the corresponding relationship between each classification.
  • the trained classification network performs text extraction on the image without a suitable template. When the time, the accuracy is higher.
  • the image includes at least one of the following: a receipt image, an invoice image, and a business card image.
  • an electronic device including: a processor; a memory for storing executable instructions of the processor; wherein the processor is configured to call the instructions stored in the memory to execute the foregoing method.
  • a computer-readable storage medium having computer program instructions stored thereon, and the computer program instructions implement the above-mentioned method when executed by a processor.
  • a computer program including computer readable code, and when the computer readable code runs in an electronic device, a processor in the electronic device executes instructions for implementing the above method .
  • a graph convolutional neural network can be used to determine the field corresponding to the text to be extracted in the image based on the relative position feature between the target regions and the feature of the text to be extracted. It does not rely on a fixed template for text extraction. Compared with the method of text extraction based on a template, the accuracy of text extraction for images without a suitable template is higher.
  • Fig. 1 shows a flowchart of an image processing method according to an embodiment of the present disclosure
  • Fig. 2 shows a schematic structural diagram of a connected graph according to an embodiment of the present disclosure
  • Fig. 3 shows a schematic structural diagram of a classification network according to an embodiment of the present disclosure
  • Fig. 4 shows a block diagram of an image processing device according to an embodiment of the present disclosure
  • Fig. 5 shows a block diagram of an electronic device according to an embodiment of the present disclosure
  • Fig. 6 shows a block diagram of an electronic device according to an embodiment of the present disclosure.
  • the technology of extracting key information based on images has made great progress.
  • the text in the image can be recognized.
  • the structure of the recognized text will be determined.
  • Information that is, to determine which field in the structured data corresponds to a certain recognized text, so as to facilitate subsequent operations such as structured storage and display of the recognized data.
  • the embodiments of the present disclosure provide an image processing method, which can determine the image in the image based on the relative position features between the target regions and the features of the text to be extracted through the graph convolutional neural network.
  • the field corresponding to the text to be extracted This method does not rely on a fixed template for text extraction.
  • the accuracy is higher when extracting text information from an image without a suitable template.
  • the image processing method provided by the embodiments of the present disclosure can be applied to the extraction of key information in the image, can realize functions such as receipt information extraction, invoice information extraction, and identity information extraction, and has high application value.
  • Fig. 1 shows a flowchart of an image processing method according to an embodiment of the present disclosure. As shown in Fig. 1, the image processing method includes:
  • Step S11 the image is recognized, and multiple target regions in the image are determined.
  • the target area is the area where the text to be extracted is located.
  • the distribution of the text to be extracted on the image is often relatively scattered, for example, there is a certain interval between the text "total price" and "19.88 yuan". Therefore, when determining the target area, you can determine the target area according to the text on the image.
  • the distribution relationship is based on the interval between the texts, and the image is divided to obtain multiple target regions.
  • the target area may also be divided according to other methods, and the specific division method may depend on the specific application scenarios of the present disclosure, which is not limited in the present disclosure.
  • the area where the text that constitutes a word, a sentence, or expresses a certain meaning can be determined as a target area, for example, the area where the text "total price" is to be extracted is a target Area, the area where "19.88 yuan" is located is a target area.
  • the present disclosure does not limit this.
  • Step S12 Determine the relative position characteristics between the target regions in the image.
  • the relative position feature can characterize the relative position relationship between the target areas.
  • the specific relative position feature can be determined according to the center points of the two target areas, or it can be determined according to a vertex of the two target areas. limit.
  • the relative position feature in the present disclosure can also be determined according to some other parameters, which will be specifically discussed in the possible implementation manners disclosed in the following text, and will not be repeated here.
  • Step S13 Determine the target feature of each target area.
  • the target feature includes the feature of the text to be extracted.
  • the feature of the text to be extracted is the feature of the text to be extracted.
  • the feature may include the visual feature of the text to be extracted as a whole, the feature of the text character of the text to be extracted, or one of the above two features.
  • Step S14 Perform feature extraction on the relative position feature and the target feature through the graph convolutional neural network to obtain the extracted feature.
  • the relative position feature and the target feature are input into the graph convolutional neural network, and feature extraction is performed to obtain the extracted features.
  • graph convolutional neural networks can represent images in the form of connected graphs and extract features.
  • Connected graph is composed of several nodes (Node) and edges (Edge) connecting two nodes. Edges are used to describe the relationship between different nodes.
  • the features extracted by the graph convolutional neural network can accurately characterize the relative position between the target regions and the features of the text to be extracted, so as to improve the accuracy of subsequent text extraction.
  • Step S15 Determine the field corresponding to the text to be extracted according to the extracted features.
  • the network can classify the text to be extracted based on the extracted features.
  • the classification category is used to characterize the text corresponding to the text to be extracted.
  • Field After the category of the text to be extracted is determined according to the extracted features, the field corresponding to the text to be extracted is determined.
  • a graph convolutional neural network can be used to determine the field corresponding to the text to be extracted in the image based on the relative position feature between the target regions and the feature of the text to be extracted.
  • the text extraction can be performed without relying on a fixed template. Compared with the method of text extraction based on a template, the accuracy of text extraction is higher when the text is extracted from an image without a suitable template.
  • determining the relative position characteristics between the target areas in the image includes: determining the relative position parameters of the first target area and the second target area in the image; and characterizing the relative position parameters , Get the relative position characteristics of the first target area and the second target area.
  • the first target area and the second target area are any two target areas in the image.
  • the relative position parameters of the first target area and the second target area in the image include at least one of the following:
  • the horizontal and vertical distances of the first target area relative to the second target area may be the horizontal and vertical distances between the reference point of the first target area and the reference point of the second target area, and the selection of the reference point of the target area , Can be the center point of the target area or a vertex of the target area.
  • the selection of a specific reference point is not limited in the present disclosure.
  • the lateral distance ⁇ x ij and the longitudinal distance ⁇ y ij of the first target area relative to the second target area are expressed as follows:
  • the first target area is the area where the text t i to be extracted is located, and the second target area is the area where the text t j is to be extracted.
  • the horizontal distance ⁇ x ij and the vertical distance ⁇ y ij can also be normalized to obtain the normalized horizontal and vertical distances.
  • the ⁇ x ij and ⁇ x ij and the vertical distance can be determined by the image size parameter.
  • ⁇ y ij is normalized. For example, when normalizing by the width W of the image, the relative position parameter is obtained The expression is as follows:
  • the high H of the image can also be used for normalization, which will not be repeated here.
  • the aspect ratio of the first target area is w i /h i
  • the aspect ratio of the second target area is w j /h j .
  • the relative size relationship between the first target area and the second target area may represent the relative size relationship between the size of the first target area and the size of the second target area. Since there are some specific relationships between the text sizes of certain fields, the relative position feature takes into account the relative size relationship between the first target area and the second target area, which can make the extraction result of the key information more accurate.
  • the size of the text "address” is shorter, and the size of the text “xx city xx street xx road xx number” is longer, so the difference between the two sizes is larger; while the text “total price” and “19.88 yuan” "The gap between the sizes is smaller. Therefore, the relative size relationship of the target area can reflect the field category corresponding to the text to a certain extent.
  • the relative position parameter includes the normalized horizontal and vertical distances, the aspect ratio of the first target area, and the relative size relationship between the first target area and the second target area. , Can make the extraction result of key information more accurate.
  • the relative position parameters can be characterized to obtain the relative position characteristics of the first target area and the second target area.
  • Characterizing the relative position parameters to obtain the relative position characteristics of the first target area and the second target area includes: mapping the relative position parameters to a D-dimensional space through a sine-cosine transformation matrix to obtain a D-dimensional feature vector, D is a positive integer; the D-dimensional feature vector is multiplied by a preset weight matrix to obtain a 1-dimensional weight value; the weight value is processed by the preset activation function to obtain the relative position feature.
  • the sine-cosine transformation matrix here is the transformation matrix used in Fourier sine transformation or cosine transformation.
  • the specific value of the preset weight matrix here can be determined by network training, and the initial value can be determined by random methods. During network training, the preset weight matrix will be tuned. The training process of the network will be described later, so I won’t go into details here.
  • the preset activation function here may be, for example, a linear rectification function (Rectified Linear Unit, ReLU), and the specific activation function may depend on the actual application scenario of the present disclosure, which is not limited in the present disclosure.
  • ReLU Rectified Linear Unit
  • M represents a sine-cosine transformation matrix
  • M(r ij ) represents that the relative position parameter r ij is mapped to a D-dimensional space through a sine-cosine transformation matrix M
  • W m is a preset weight matrix
  • ReLU represents a linear rectification function .
  • the relative position parameter can be converted into the data format required by the edge of the graph convolutional neural network through the feature processing, which is convenient for subsequent feature extraction through the graph convolutional neural network.
  • the target features in the embodiments of the present disclosure may include the visual features of the text to be extracted as a whole, or the text characters of the text to be extracted.
  • determining the target characteristics of each target area includes: determining the pixel data in the target area, extracting the pixel data to obtain the visual characteristics; determining the text characters in the target area, and comparing the text Characters perform feature extraction to obtain text character features; according to the extracted visual features and character features, the target features of the target area are determined.
  • the visual features can reflect the overall visual information of the text in the target area.
  • the specific extraction can be performed by a region of interest alignment (Region of Interest Align, RoI Align) method, and the present disclosure does not limit the specific way of extracting visual features.
  • the text characters can be recognized and extracted through text recognition technology. For example, it is possible to perform feature extraction on text characters through optical character recognition technology (Optical Character Recognition, OCR) to obtain text characters.
  • OCR Optical Character Recognition
  • the present disclosure does not limit the specific method of extracting text characters.
  • performing feature extraction on the text characters to obtain character features includes: mapping the text characters to a low-dimensional feature space through one-hot encoding; and then through bidirectional Bi-LSTM processes the text characters in the low-dimensional feature space to obtain the feature representation of the text, that is, obtain the character features of the text to be extracted.
  • discrete features text characters
  • a certain value of discrete features corresponds to a point in Euclidean space, which makes the calculation between features more reasonable.
  • determining the target feature of the target area according to the extracted visual features and character features includes: assigning different weights to the visual features and character features; and assigning weights to the visual features Combine (for example, add) with the character features to obtain the target feature of the target area.
  • weights can be optimized through network training, and the specific training process is described in detail later, and will not be repeated here.
  • the process of performing feature extraction on the text character s i to obtain the character feature t i can be expressed as the available formula (7).
  • W ⁇ R C ⁇ D represents the projection matrix of the one-hot encoding
  • Bi-LSTM represents the processing of the text characters after the one-hot encoding through the two-way long and short time series network. Represents the jth character in the text character s i.
  • the target feature n i can be obtained by referring to formulas (8) and (9).
  • ⁇ i ⁇ (W t t i +W v v i ) (8)
  • n i ⁇ i U t t i +(1- ⁇ i )U v v i (9)
  • W t ⁇ R 1 ⁇ Dt and W v ⁇ R 1 ⁇ Dv are one-dimensional projection matrices, which can be specifically optimized through network training, and ⁇ is the activation function.
  • U t ⁇ R Dh ⁇ Dt and U v ⁇ R Dh ⁇ Dt are projection parameters, which can also be obtained through network training.
  • the relative location feature and the target feature can be extracted through the graph convolutional neural network.
  • the relative position feature and the target feature are extracted through the graph convolutional neural network, and the extracted features are obtained, including: taking each target feature as the node of the graph, and taking each relative position feature as Connect the edges of two nodes to construct a connected graph; through the graph convolutional neural network, the connected graph is iteratively updated, and the connected graph that meets the convergence condition after the iterative update is used as the extracted feature.
  • the relative position feature When constructing a connected graph using the relative position feature of the target area as the edge connecting two nodes, the relative position feature will be used as a parameter of the adjacency matrix between the nodes.
  • the adjacency matrix can also include the semantic similarity of the nodes and other things. Parameters, this disclosure does not limit the specific settings of other parameters.
  • FIG. 2 is a schematic diagram of a connected graph provided in the present disclosure.
  • the nodes of the graph are target features
  • the edges connecting two nodes are the relative position features of the target area.
  • the connected graph constructed by the embodiments of the present disclosure includes not only the target features in the image, but also the relative position features between the target features in the image, which can characterize the characteristics of the text in the image as a whole, and therefore can improve the extraction of key information. The accuracy of the results.
  • the connected graph can be iteratively updated through the graph convolutional neural network, and the connected graph that meets the convergence condition after the iterative update is used as the extracted feature.
  • the feature of any node i is updated by projecting the feature value of each node through the adjacency matrix of each node connected to node i.
  • the feature value of each node will be It will no longer change with the increase of the number of iterations, that is, the eigenvalues of the nodes remain unchanged, at this time it can be regarded as meeting the convergence condition, and the connected graph meeting the convergence condition can be used as the extracted feature.
  • N l+1 ⁇ ((A l N l )W l ) (10)
  • N l is the feature of node N in the lth iteration
  • W l is the conversion matrix, which can be obtained through network training optimization
  • a l is the adjacency matrix of the node
  • the expression of the adjacency matrix A l ij of the nodes i and j as follows:
  • (n l i ) T represents the transposition of n l i, Represents the normalization parameters, which can be optimized through network training.
  • the extracted features determine the field corresponding to the text to be extracted, including: outputting the image convolutional neural network according to a plurality of pre-defined preset categories
  • the nodes in the connected graph are classified to obtain the category of the node.
  • the preset category includes: the category of the identifier of the characterizing text belonging to the preset field, and the category of the field value of the characterizing text belonging to the preset field; Type, to determine the identifier or field value of the preset field corresponding to the text to be extracted.
  • the recognized text there may be text that characterizes the identifier of the preset field, and there may also be text that characterizes the field value of the preset field.
  • the text that characterizes the identifier of the preset field is the text in the image used to indicate which field the field value belongs to, and the field value is the specific value under the field. For example, for the preset field "Total Price”, the image is identified The text “total price”, “total price” or “sub total”, etc., are all specific identifiers of the preset field “total price”; for the recognized text "19.88 yuan", “ ⁇ : 19.88", etc. , Are the field values of the preset fields.
  • two categories can be set to correspond to the preset field respectively.
  • one category is the category that characterizes the text belonging to the identifier of the preset field
  • the other category is the field that characterizes the text belongs to the preset field.
  • the category of the value When there are multiple different preset fields, each preset field can be set to 2 categories, so there will be multiple characterization texts belonging to the identification category of the preset field, and multiple characterization texts belonging to the preset field The category of the field value.
  • the preset fields can be set to "name”, “address”, “phone number”, “date”, “time”, “product category”, “product name”, “Commodity unit price”, “Single product total price”, “Taxes”, “Total total price”, “Reminder”, a total of 12 preset fields, then 24 categories can be preset, which respectively indicate the preset value of each preset field. Set the field identifier and the field value of each preset field. In addition, the category “Others” can be set to distinguish and extract texts that do not belong to the above categories, that is, a total of 25 categories are set.
  • the image processing method of the embodiment of the present disclosure may be implemented by a pre-built classification network, and the training steps of the classification network are as follows:
  • the label category includes: the category of the identifier representing the text belonging to a preset field, and the field value of the characterizing text belonging to the preset field Category
  • the classification network can be used to implement the image processing technology of the present disclosure.
  • the classification network can include the graph convolutional neural network described above.
  • the classification network can also include other networks.
  • the Bi-LSTM network for the networks included in the classification network of the present disclosure, may be determined according to the specific application scenarios of the embodiments of the present disclosure, which is not limited in the present disclosure.
  • FIG. 3 is a schematic structural diagram of a specific implementation of a classification network provided in this application.
  • the network includes a target feature extraction module, a relative position feature extraction module, a convolutional network feature extraction module, and a classification module. Extract the target feature of the image containing the text to be extracted through the target feature extraction module, and extract the relative position feature of the image through the relative position feature extraction module; input the target feature and relative position feature to the convolutional network feature extraction module for iterative update, and get The iteratively extracted features; then the iteratively extracted features are classified through the classification module to obtain the predicted category of the node.
  • the category characterizes the field corresponding to the text to be extracted, after the category of the text to be extracted is determined according to the extracted features, the field corresponding to the text to be extracted is determined.
  • the specific functions of each module please refer to the relevant discussion in this disclosure, which will not be repeated here.
  • the label category may be the preset category described above, which will not be repeated here.
  • the parameters in the classification network can be adjusted according to the loss of the first prediction category relative to the label category, so that the classification network can The difference between the predicted category and the labeled category of the sample image is the smallest.
  • the identification and identification value of whether two texts belong to the same preset field is also beneficial to the classification accuracy of the classification network.
  • the two texts respectively belonging to the identification and identification value of the same preset field are referred to as a field pair, for example, the text "total price” and "19.88 yuan" constitute a field pair.
  • the classification network when training the classification network, the classification network will also output the correspondence between the categories in the first prediction category, and at the same time, the correspondence between the texts will also be marked in the sample image. Then, the classification network can be trained according to the correspondence between the output of the classification network and the correspondence between the labeled texts to be extracted.
  • the loss function used during training may specifically be a cross entropy loss function (Cross Entropy Loss, CE), and the specific loss function may be selected according to actual requirements, which is not specifically limited in the present disclosure.
  • CE Cross Entropy Loss
  • the trained classification network can be used to determine the field corresponding to the text to be extracted during the extraction of key text information.
  • the trained classification network has higher accuracy when extracting text from images without adapted templates.
  • the recognized image includes at least one of the following: a receipt image, an invoice image, and a business card image.
  • a receipt image a receipt image
  • an invoice image a payment image
  • a business card image a payment image that specifies the payment amount.
  • the embodiments of the present disclosure can also be used to recognize other images, and the present disclosure does not specifically limit this.
  • a graph convolutional neural network can be used to determine the field corresponding to the text to be extracted in the image based on the relative position feature between the target regions and the feature of the text to be extracted.
  • the text extraction can be performed without relying on a fixed template. Compared with the method of text extraction based on a template, the accuracy of text extraction is higher when the text is extracted from an image without a suitable template.
  • the embodiments of the present disclosure when text extraction is performed, not only the text character features in the target area are used, but also the visual features of the target area are used, which reduces the influence of misrecognition of text characters on the final classification and improves the performance of text extraction. Accuracy; In addition, by establishing the spatial position relationship between the text areas, it is not dependent on the pre-designed templates, and can handle unseen templates, which has better scalability.
  • the image processing method can be executed by electronic equipment such as a terminal device or a server, and the terminal device can be a user equipment (UE), a mobile device, a user terminal, a terminal, a cellular phone, or a cordless
  • UE user equipment
  • PDAs personal digital assistants
  • the method can be implemented by a processor invoking computer-readable instructions stored in a memory.
  • the method can be executed by a server.
  • the present disclosure also provides image processing devices, electronic equipment, computer-readable storage media, and programs, all of which can be used to implement any of the image processing methods provided in the present disclosure.
  • image processing devices electronic equipment, computer-readable storage media, and programs, all of which can be used to implement any of the image processing methods provided in the present disclosure.
  • FIG. 4 shows a block diagram of an image processing device according to an embodiment of the present disclosure.
  • the image processing device 20 includes:
  • the recognition module 21 is configured to recognize an image and determine multiple target regions in the image, where the target region is the region where the text to be extracted is located;
  • the relative position feature determining module 22 is used to determine the relative position feature between each target area in the image
  • the target feature determining module 23 is configured to determine the target feature of each target area, where the target feature includes the feature of the text to be extracted;
  • the graph convolution module 24 is configured to perform feature extraction on the relative position feature and the target feature through the graph convolution neural network to obtain the extracted feature;
  • the field determination module 25 is configured to determine the field corresponding to the text to be extracted according to the extracted features.
  • a graph convolutional neural network can be used to determine the field corresponding to the text to be extracted in the image based on the relative position feature between the target regions and the feature of the text to be extracted.
  • the text extraction can be performed without relying on a fixed template. Compared with the method of text extraction based on a template, the accuracy of text extraction is higher when the text is extracted from an image without a suitable template.
  • the graph convolution module 24 includes: a first graph convolution sub-module and a second graph convolution sub-module, wherein:
  • the first graph convolution submodule is used to construct a connected graph by taking each of the target features as the nodes of the graph, and using each of the relative position features as the edges connecting the two nodes;
  • the second graph convolution submodule is used to iteratively update the connected graph through the graph convolutional neural network, and use the connected graph that meets the convergence condition after the iterative update as the extracted feature.
  • the constructed connected graph includes not only the target features in the image, but also the relative position features between the target features in the image, which can characterize the characteristics of the text in the image as a whole, and therefore can improve the key information The accuracy of the extraction results.
  • graph convolutional neural networks can represent images in the form of connected graphs and extract features.
  • Connected graph is composed of several nodes (Node) and edges (Edge) connecting two nodes. Edges are used to describe the relationship between different nodes. Therefore, the features extracted by the graph convolutional neural network can accurately characterize the relative position between the target regions and the features of the text to be extracted, so as to improve the accuracy of subsequent text extraction.
  • the field determination module 25 includes: a first field determination sub-module and a second field determination sub-module, where:
  • the first field determination sub-module is used to classify the nodes in the connected graph output by the graph convolutional neural network according to a plurality of pre-defined preset categories to obtain the category of the node, and the preset category includes: the characterization text belongs to The category of the identifier of the preset field, and the category of the field value of the characterizing text belonging to the preset field;
  • the second field determination submodule is used to determine the identifier or field value of the preset field corresponding to the text to be extracted according to the category of the node.
  • the identifier or field value of the text to be extracted corresponding to the preset field can be obtained, which improves The accuracy of text extraction is improved.
  • the relative position feature determination module 22 includes: a first relative position feature determination sub-module and a second relative position feature determination sub-module, wherein:
  • the first relative position feature determining sub-module is used to determine the relative position parameters of the first target area and the second target area in the image;
  • the second relative position feature determination sub-module is used to perform characterization processing on the relative position parameters to obtain the relative position features of the first target area and the second target area.
  • the relative position parameter includes at least one of the following:
  • the relative position parameter includes the horizontal distance and the vertical distance, the aspect ratio of the first target area, and the relative size relationship between the first target area and the second target area, so that The extraction result of key information is more accurate.
  • the second relative position feature determination submodule is used to map the relative position parameter to a D-dimensional space through a sine-cosine transformation matrix to obtain a D-dimensional eigenvector, where D is a positive integer Transform the D-dimensional feature vector into a 1-dimensional weight value through a preset weight matrix; process the weight value through a preset activation function to obtain a relative position feature.
  • the relative position parameter can be converted into the data format required by the edge of the graph convolutional neural network through the feature processing, which is convenient for subsequent feature extraction through the graph convolutional neural network.
  • the target feature determination module 23 includes a first target feature determination sub-module, a second target feature determination sub-module, and a third target feature determination sub-module, wherein:
  • the first target feature determination sub-module is used to determine pixel data in the target area, and perform feature extraction on the pixel data to obtain visual features;
  • the second target feature determination sub-module is used to determine text characters in the target area, and perform feature extraction on the text characters to obtain character features;
  • the third target feature determination sub-module is used to determine the target feature of the target area according to the extracted visual features and character features.
  • the third target feature determination submodule is used to assign different weights to the visual features and character features; to fuse the weighted visual features and character features to obtain the target area Target characteristics.
  • the device is implemented through a pre-built classification network, and the device further includes:
  • the first training module is configured to input the sample image into the classification network for processing to obtain the first prediction category of the text to be extracted in the sample image, and the corresponding relationship between each category in the first prediction category;
  • the second training module is configured to train the classification network according to the first prediction category and the label category of the sample image.
  • the label category includes: the category of the identifier that characterizes the text belonging to the preset field, and the characterization text The category of the field value belonging to the preset field;
  • the third training module is configured to train the classification network according to the corresponding relationship and the corresponding relationship between the labeled texts to be extracted.
  • the classification network can be trained more accurately by labeling the classification of the sample image and the corresponding relationship between each classification.
  • the trained classification network performs text extraction on the image without a suitable template. When the time, the accuracy is higher.
  • the image includes at least one of the following: a receipt image, an invoice image, and a business card image.
  • the functions or modules contained in the device provided in the embodiments of the present disclosure can be used to execute the methods described in the above method embodiments.
  • the functions or modules contained in the device provided in the embodiments of the present disclosure can be used to execute the methods described in the above method embodiments.
  • the embodiments of the present disclosure also provide a computer-readable storage medium on which computer program instructions are stored, and the computer program instructions implement the above-mentioned method when executed by a processor.
  • the computer-readable storage medium may be a volatile computer-readable storage medium or a non-volatile computer-readable storage medium.
  • An embodiment of the present disclosure also proposes an electronic device, including: a processor; a memory for storing executable instructions of the processor; wherein the processor is configured to call the instructions stored in the memory to execute the above method.
  • the embodiments of the present disclosure also provide a computer program product, including computer-readable code.
  • the processor in the device executes the image processing method for implementing the image processing method provided by any of the above embodiments. instruction.
  • the embodiments of the present disclosure also provide another computer program product for storing computer-readable instructions, which when executed, cause the computer to perform the operations of the image processing method provided by any of the foregoing embodiments.
  • the electronic device can be provided as a terminal, server or other form of device.
  • FIG. 5 shows a block diagram of an electronic device 800 according to an embodiment of the present disclosure.
  • the electronic device 800 may be a mobile phone, a computer, a digital broadcasting terminal, a messaging device, a game console, a tablet device, a medical device, a fitness device, a personal digital assistant, and other terminals.
  • the electronic device 800 may include one or more of the following components: a processing component 802, a memory 804, a power supply component 806, a multimedia component 808, an audio component 810, an input/output (I/O) interface 812, and a sensor component 814 , And communication component 816.
  • the processing component 802 generally controls the overall operations of the electronic device 800, such as operations associated with display, telephone calls, data communications, camera operations, and recording operations.
  • the processing component 802 may include one or more processors 820 to execute instructions to complete all or part of the steps of the foregoing method.
  • the processing component 802 may include one or more modules to facilitate the interaction between the processing component 802 and other components.
  • the processing component 802 may include a multimedia module to facilitate the interaction between the multimedia component 808 and the processing component 802.
  • the memory 804 is configured to store various types of data to support operations in the electronic device 800. Examples of such data include instructions for any application or method operating on the electronic device 800, contact data, phone book data, messages, images, videos, etc.
  • the memory 804 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable and Programmable Read Only Memory (EPROM), Programmable Read Only Memory (PROM), Read Only Memory (ROM), Magnetic Memory, Flash Memory, Magnetic Disk or Optical Disk.
  • SRAM static random access memory
  • EEPROM electrically erasable programmable read-only memory
  • EPROM erasable and Programmable Read Only Memory
  • PROM Programmable Read Only Memory
  • ROM Read Only Memory
  • Magnetic Memory Flash Memory
  • Magnetic Disk Magnetic Disk or Optical Disk.
  • the power supply component 806 provides power for various components of the electronic device 800.
  • the power supply component 806 may include a power management system, one or more power supplies, and other components associated with the generation, management, and distribution of power for the electronic device 800.
  • the multimedia component 808 includes a screen that provides an output interface between the electronic device 800 and the user.
  • the screen may include a liquid crystal display (LCD) and a touch panel (TP). If the screen includes a touch panel, the screen may be implemented as a touch screen to receive input signals from the user.
  • the touch panel includes one or more touch sensors to sense touch, sliding, and gestures on the touch panel. The touch sensor may not only sense the boundary of a touch or slide action, but also detect the duration and pressure related to the touch or slide operation.
  • the multimedia component 808 includes a front camera and/or a rear camera. When the electronic device 800 is in an operation mode, such as a shooting mode or a video mode, the front camera and/or the rear camera can receive external multimedia data. Each front camera and rear camera can be a fixed optical lens system or have focal length and optical zoom capabilities.
  • the audio component 810 is configured to output and/or input audio signals.
  • the audio component 810 includes a microphone (MIC), and when the electronic device 800 is in an operation mode, such as a call mode, a recording mode, and a voice recognition mode, the microphone is configured to receive an external audio signal.
  • the received audio signal may be further stored in the memory 804 or transmitted via the communication component 816.
  • the audio component 810 further includes a speaker for outputting audio signals.
  • the I/O interface 812 provides an interface between the processing component 802 and a peripheral interface module.
  • the above-mentioned peripheral interface module may be a keyboard, a click wheel, a button, and the like. These buttons may include, but are not limited to: home button, volume button, start button, and lock button.
  • the sensor component 814 includes one or more sensors for providing the electronic device 800 with various aspects of state evaluation.
  • the sensor component 814 can detect the on/off status of the electronic device 800 and the relative positioning of the components.
  • the component is the display and the keypad of the electronic device 800.
  • the sensor component 814 can also detect the electronic device 800 or the electronic device 800.
  • the position of the component changes, the presence or absence of contact between the user and the electronic device 800, the orientation or acceleration/deceleration of the electronic device 800, and the temperature change of the electronic device 800.
  • the sensor component 814 may include a proximity sensor configured to detect the presence of nearby objects when there is no physical contact.
  • the sensor component 814 may also include a light sensor, such as a CMOS or CCD image sensor, for use in imaging applications.
  • the sensor component 814 may also include an acceleration sensor, a gyroscope sensor, a magnetic sensor, a pressure sensor, or a temperature sensor.
  • the communication component 816 is configured to facilitate wired or wireless communication between the electronic device 800 and other devices.
  • the electronic device 800 can access a wireless network based on a communication standard, such as WiFi, 2G, or 3G, or a combination thereof.
  • the communication component 816 receives a broadcast signal or broadcast related information from an external broadcast management system via a broadcast channel.
  • the communication component 816 further includes a near field communication (NFC) module to facilitate short-range communication.
  • the NFC module can be implemented based on radio frequency identification (RFID) technology, infrared data association (IrDA) technology, ultra-wideband (UWB) technology, Bluetooth (BT) technology and other technologies.
  • RFID radio frequency identification
  • IrDA infrared data association
  • UWB ultra-wideband
  • Bluetooth Bluetooth
  • the electronic device 800 may be implemented by one or more application-specific integrated circuits (ASIC), digital signal processors (DSP), digital signal processing devices (DSPD), programmable logic devices (PLD), field-available A programmable gate array (FPGA), controller, microcontroller, microprocessor, or other electronic components are implemented to implement the above methods.
  • ASIC application-specific integrated circuits
  • DSP digital signal processors
  • DSPD digital signal processing devices
  • PLD programmable logic devices
  • FPGA field-available A programmable gate array
  • controller microcontroller, microprocessor, or other electronic components are implemented to implement the above methods.
  • a non-volatile computer-readable storage medium such as the memory 804 including computer program instructions, which can be executed by the processor 820 of the electronic device 800 to complete the foregoing method.
  • FIG. 6 shows a block diagram of an electronic device 1900 according to an embodiment of the present disclosure.
  • the electronic device 1900 may be provided as a server. 6
  • the electronic device 1900 includes a processing component 1922, which further includes one or more processors, and a memory resource represented by the memory 1932, for storing instructions executable by the processing component 1922, such as application programs.
  • the application program stored in the memory 1932 may include one or more modules each corresponding to a set of instructions.
  • the processing component 1922 is configured to execute instructions to perform the above-described methods.
  • the electronic device 1900 may also include a power supply component 1926 configured to perform power management of the electronic device 1900, a wired or wireless network interface 1950 configured to connect the electronic device 1900 to the network, and an input output (I/O) interface 1958 .
  • the electronic device 1900 can operate based on an operating system stored in the memory 1932, such as Windows ServerTM, Mac OS XTM, UnixTM, LinuxTM, FreeBSDTM or the like.
  • a non-volatile computer-readable storage medium is also provided, such as the memory 1932 including computer program instructions, which can be executed by the processing component 1922 of the electronic device 1900 to complete the foregoing method.
  • the present disclosure may be a system, method and/or computer program product.
  • the computer program product may include a computer-readable storage medium loaded with computer-readable program instructions for enabling a processor to implement various aspects of the present disclosure.
  • the computer-readable storage medium may be a tangible device that can hold and store instructions used by the instruction execution device.
  • the computer-readable storage medium may be, for example, but not limited to, an electrical storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination of the foregoing.
  • Non-exhaustive list of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM) Or flash memory), static random access memory (SRAM), portable compact disk read-only memory (CD-ROM), digital versatile disk (DVD), memory stick, floppy disk, mechanical encoding device, such as a printer with instructions stored thereon
  • RAM random access memory
  • ROM read-only memory
  • EPROM erasable programmable read-only memory
  • flash memory flash memory
  • SRAM static random access memory
  • CD-ROM compact disk read-only memory
  • DVD digital versatile disk
  • memory stick floppy disk
  • mechanical encoding device such as a printer with instructions stored thereon
  • the computer-readable storage medium used here is not interpreted as the instantaneous signal itself, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (for example, light pulses through fiber optic cables), or through wires Transmission of electrical signals.
  • the computer-readable program instructions described herein can be downloaded from a computer-readable storage medium to various computing/processing devices, or downloaded to an external computer or external storage device via a network, such as the Internet, a local area network, a wide area network, and/or a wireless network.
  • the network may include copper transmission cables, optical fiber transmission, wireless transmission, routers, firewalls, switches, gateway computers, and/or edge servers.
  • the network adapter card or network interface in each computing/processing device receives computer-readable program instructions from the network, and forwards the computer-readable program instructions for storage in the computer-readable storage medium in each computing/processing device .
  • the computer program instructions used to perform the operations of the present disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-related instructions, microcode, firmware instructions, state setting data, or in one or more programming languages.
  • Source code or object code written in any combination, the programming language includes object-oriented programming languages such as Smalltalk, C++, etc., and conventional procedural programming languages such as "C" language or similar programming languages.
  • Computer-readable program instructions can be executed entirely on the user's computer, partly on the user's computer, executed as a stand-alone software package, partly on the user's computer and partly executed on a remote computer, or entirely on the remote computer or server carried out.
  • the remote computer can be connected to the user's computer through any kind of network-including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computer (for example, using an Internet service provider to connect to the user's computer) connection).
  • LAN local area network
  • WAN wide area network
  • an electronic circuit such as a programmable logic circuit, a field programmable gate array (FPGA), or a programmable logic array (PLA), can be customized by using the status information of the computer-readable program instructions.
  • FPGA field programmable gate array
  • PDA programmable logic array
  • the computer-readable program instructions are executed to realize various aspects of the present disclosure.
  • These computer-readable program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, thereby producing a machine that makes these instructions when executed by the processor of the computer or other programmable data processing device , A device that implements the functions/actions specified in one or more blocks in the flowcharts and/or block diagrams is produced. It is also possible to store these computer-readable program instructions in a computer-readable storage medium. These instructions make computers, programmable data processing apparatuses, and/or other devices work in a specific manner. Thus, the computer-readable medium storing the instructions includes An article of manufacture, which includes instructions for implementing various aspects of the functions/actions specified in one or more blocks in the flowcharts and/or block diagrams.
  • each block in the flowchart or block diagram may represent a module, program segment, or part of an instruction, and the module, program segment, or part of an instruction contains one or more components for realizing the specified logical function.
  • Executable instructions may also occur in a different order from the order marked in the drawings. For example, two consecutive blocks can actually be executed substantially in parallel, or they can sometimes be executed in the reverse order, depending on the functions involved.
  • each block in the block diagram and/or flowchart, and the combination of the blocks in the block diagram and/or flowchart can be implemented by a dedicated hardware-based system that performs the specified functions or actions Or it can be realized by a combination of dedicated hardware and computer instructions.
  • the computer program product can be specifically implemented by hardware, software, or a combination thereof.
  • the computer program product is specifically embodied as a computer storage medium.
  • the computer program product is specifically embodied as a software product, such as a software development kit (SDK), etc. Wait.
  • SDK software development kit

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Computation (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • General Engineering & Computer Science (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Multimedia (AREA)
  • General Health & Medical Sciences (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Computing Systems (AREA)
  • Computational Linguistics (AREA)
  • Biophysics (AREA)
  • Molecular Biology (AREA)
  • Biomedical Technology (AREA)
  • Mathematical Physics (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Databases & Information Systems (AREA)
  • Medical Informatics (AREA)
  • Character Discrimination (AREA)
  • Image Analysis (AREA)

Abstract

一种图像处理方法及装置、电子设备和存储介质，所述方法包括：对图像进行识别，确定图像中的多个目标区域（S11），所述目标区域为待提取文本所在区域；确定图像中各目标区域之间的相对位置特征（S12）；确定各目标区域的目标特征（S13），所述目标特征包括所述待提取文本的特征；通过图卷积神经网络，对相对位置特征和目标特征进行特征提取，得到提取后的特征（S14）；根据提取后的特征，确定待提取文本对应的字段（S15）。

Description

图像处理方法及装置、电子设备和存储介质
本申请要求在2019年12月27日提交中国专利局、申请号为201911387827.1、申请名称为“图像处理方法及装置、电子设备和存储介质”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。
技术领域
本公开涉及计算机技术领域，尤其涉及一种图像处理方法及装置、电子设备和存储介质。
背景技术
图像中关键文字信息提取在自动化办公等场景中有着非常重要的作用，例如，通过对图像中的关键文字信息提取，可以实现诸如收据信息提取、发票信息提取、身份信息提取等功能。
在对图像中的文字进行提取时，会将识别出的文字对应到不同的字段中，以便后续对文字进行结构化存储、展示等操作。例如，识别出来的文字是“19.88元”，需要确定“19.88元”是对应字段“总价”，还是对应字段“单价”，以便后续将“19.88元”作为某个字段的值进行存储。
通常，会根据图像中文字的排布规则，预先定义模板，模板中定义了某个位置的文字和字段的对应关系，这样便可以确定识别出来的位于某个位置的文字对应的字段。例如，预先定义图像右下角的文字对应的字段为“总价”，这样可以确定图像右下角识别出来的“19.88元”对应的字段为“总价”。
发明内容
本公开提出了一种图像处理技术方案。
根据本公开的一方面，提供了一种图像处理方法，包括：对图像进行识别，确定所述图像中的多个目标区域，所述目标区域为待提取文本所在区域；确定所述图像中各目标区域之间的相对位置特征；确定各所述目标区域的目标特征，所述目标特征包括所述待提取文本的特征；通过图卷积神经网络，对所述相对位置特征和所述目标特征进行特征提取，得到提取后的特征；根据提取后的特征，确定所述待提取文本对应的字段。
在本公开实施例中，能够通过图卷积神经网络，基于各目标区域之间的相对位置特征以及待提取文本的特征，确定图像中的待提取文本对应的字段。可不依赖于固定的模板进行文本提取，相对于基于模板进行文本提取的方式，在对没有适配模板的图像进行文本提取时，准确性较高。
在一种可能的实现方式中，通过图卷积神经网络，对所述相对位置特征和所述目标特征进行特征提取，得到提取后的特征，包括：以各所述目标特征为图的节点，以各所述相对位置特征为连接两个节点的边，构建连通图；通过图卷积神经网络，对所述连通图进行迭代更新，将迭代更新后满足收敛条件的连通图作为提取后的特征。
在本公开实施例中，构建的连通图既包含了图像中的目标特征，也包含了图像中目标特征之间的相对位置特征，可以从整体上表征图像中文字的特征，因此能够提高关键信息提取结果的准确性。
图卷积神经网络在对特征进行提取时，能够以连通图的形式表示图像，对特征进行提取。连通图由若干个结点(Node)及连接两个结点的边(Edge)所构成，边用于刻画不同结点之间的关系。因此，通过图卷积神经网络提取后的特征，能够准确地表征各目标区域之间的相对位置和待提取文本的特征，以提高后续文本提取时的准确性。
在一种可能的实现方式中，根据提取后的特征，确定所述待提取文本对应的字段，包括：根据预先定义的多个预设类别，对图卷积神经网络输出的连通图中的节点进行分类，得到节点的类别，所述预设类别包括：表征文本属于预设字段的标识的类别，以及表征文本属于预设字段的字段值的类别；根据所述节点的类别，确定待提取文本对应于预设字段的标识或字段值。
在本公开实施例中，通过预先定义预设类别为预设字段的标识或字段，根据提取的特征对待提取文本进行分类，即可得到待提取文本对应于预设字段的标识或字段值，提高了文本提取时的准确性。
在一种可能的实现方式中，确定所述图像中各目标区域之间的相对位置特征，包括：确定图像中的第一目标区域和第二目标区域的相对位置参数；对所述相对位置参数进行特征化处理，得到第一目标区域和第二目标区域的相对位置特征。
在一种可能的实现方式中，所述相对位置参数包括下述至少一种：第一目标区域相对于第二目标区域的横向距离和纵向距离；所述第一目标区域的宽高比；所述第二目标区域的宽高比；所述第一目标区域和所述第二目标区域的相对尺寸关系。
在本公开实施例中，相对位置参数即包含了横向距离和纵向距离，也包含了第一目标区域的宽高比，也包含了第一目标区域和第二目标区域的相对尺寸关系，可以使得关键信息的提取结果更加准确。
在一种可能的实现方式中，对所述相对位置参数进行特征化处理，得到第一目标区域和第二目标区域的相对位置特征，包括：将所述相对位置参数通过正余弦变换矩阵映射到一个D维的空间，得到D维的特征向量，D为正整数；通过预设权重矩阵，将所述D维的特征向量转化为1维的权重值；通过预设激活函数对所述权重值进行处理，得到相对位置特征。
在本公开实施例中，通过特征化处理可以将相对位置参数转换为图卷积神经网络的边所需的数据格式，便于后续通过图卷积神经网络进行特征提取。
在一种可能的实现方式中，确定各所述目标区域的目标特征，包括：确定目标区域中的像素数据，对所述像素数据进行特征提取，得到视觉特征；确定目标区域中的文本字符，对所述文本字符进行特征提取，得到字符特征；根据提取到的视觉特征和字符特征，确定目标区域的目标特征。
在本公开实施例中，考虑到图像中会存在由于拍照视角、光线、遮挡等原因带来的干扰因素，因此，通过文字检测识别通常会有较多的误识，即可能会识别出错误的文本字符，这可能会影响关键信息提取的准确性。而通过视觉信息的提取，将视觉信息考虑到关键信息提取中，会降低文本误识对关键信息提取的影响。即使文本识别错误，但由于视觉信息不会改变太大，因此二者结合能够提高关键信息提取结果的准确性。
在一种可能的实现方式中，根据提取到的视觉特征和字符特征，确定目标区域的目标特征，包括：将所述视觉特征和字符特征赋予不同的权重；对赋予权重后的所述视觉特征和字符特征进行融合，得到目标区域的目标特征。
在本公开实施例中，通过对视觉特征和字符特征赋予不同的权重，能够提高关键信息提取结果的准确性。
在一种可能的实现方式中，所述方法通过预先构建的分类网络实现，所述分类网络的训练步骤如下：将样本图像输入所述分类网络中处理，得到样本图像中待提取文本的第一预测类别，以及所述第一预测类别中各个类别之间的对应关系；根据所述第一预测类别，以及所述样本图像的标注类别，训练所述分类网络，所述标注类别包括：表征文本属于预设字段的标识的类别，以及表征文本属于预设字段的字段值的类别；根据所述对应关系，以及标注的待提取文本之间的对应关系，训练所述分类网络。
在本公开实施例中，通过对样本图像的类别和各个类别之间的对应关系进行标注， 能够更准确地对分类网络进行训练，训练得到的分类网络在对没有适配模板的图像进行文本提取时，准确性较高。
在一种可能的实现方式中，所述图像包括下述至少一种：收据图像、发票图像、名片图像。
根据本公开的一方面，提供了一种图像处理装置，包括：识别模块，用于对图像进行识别，确定所述图像中的多个目标区域，所述目标区域为待提取文本所在区域；相对位置特征确定模块，用于确定所述图像中各目标区域之间的相对位置特征；目标特征确定模块，用于确定各所述目标区域的目标特征，所述目标特征包括所述待提取文本的特征；图卷积模块，用于通过图卷积神经网络，对所述相对位置特征和所述目标特征进行特征提取，得到提取后的特征；字段确定模块，用于根据提取后的特征，确定所述待提取文本对应的字段。
在本公开实施例中，能够通过图卷积神经网络，基于各目标区域之间的相对位置特征以及待提取文本的特征，确定图像中的待提取文本对应的字段。可不依赖于固定的模板进行文本提取，相对于基于模板进行文本提取的方式，在对没有适配模板的图像进行文本提取时，准确性较高。
在一种可能的实现方式中，图卷积模块包括：第一图卷积子模块和第二图卷积子模块，其中，第一图卷积子模块，用于以各所述目标特征为图的节点，以各所述相对位置特征为连接两个节点的边，构建连通图；第二图卷积子模块，用于通过图卷积神经网络，对所述连通图进行迭代更新，将迭代更新后满足收敛条件的连通图作为提取后的特征。
在本公开实施例中，构建的连通图既包含了图像中的目标特征，也包含了图像中目标特征之间的相对位置特征，可以从整体上表征图像中文字的特征，因此能够提高关键信息提取结果的准确性。
图卷积神经网络在对特征进行提取时，能够以连通图的形式表示图像，对特征进行提取。连通图由若干个结点及连接两个结点的边所构成，边用于刻画不同结点之间的关系。因此，通过图卷积神经网络提取后的特征，能够准确地表征各目标区域之间的相对位置和待提取文本的特征，以提高后续文本提取时的准确性。
在一种可能的实现方式中，字段确定模块包括：第一字段确定子模块和第二字段确定子模块，其中，第一字段确定子模块，用于根据预先定义的多个预设类别，对图卷积神经网络输出的连通图中的节点进行分类，得到节点的类别，所述预设类别包括：表征文本属于预设字段的标识的类别，以及表征文本属于预设字段的字段值的类别；第二字段确定子模块，用于根据所述节点的类别，确定待提取文本对应于预设字段的标识或字段值。
在本公开实施例中，通过预先定义预设类别为预设字段的标识或字段，根据提取的特征对待提取文本进行分类，即可得到待提取文本对应于预设字段的标识或字段值，提高了文本提取时的准确性。
在一种可能的实现方式中，相对位置特征确定模块包括：第一相对位置特征确定子模块和第二相对位置特征确定子模块，其中，第一相对位置特征确定子模块，用于确定图像中的第一目标区域和第二目标区域的相对位置参数；第二相对位置特征确定子模块，用于对所述相对位置参数进行特征化处理，得到第一目标区域和第二目标区域的相对位置特征。
在一种可能的实现方式中，所述相对位置参数包括下述至少一种：第一目标区域相对于第二目标区域的横向距离和纵向距离；所述第一目标区域的宽高比；所述第二目标区域的宽高比；所述第一目标区域和所述第二目标区域的相对尺寸关系。
在本公开实施例中，相对位置参数即包含了横向距离和纵向距离，也包含了第一目标区域的宽高比，也包含了第一目标区域和第二目标区域的相对尺寸关系，可以使得关 键信息的提取结果更加准确。
在一种可能的实现方式中，第二相对位置特征确定子模块，用于将所述相对位置参数通过正余弦变换矩阵映射到一个D维的空间，得到D维的特征向量，D为正整数；通过预设权重矩阵，将所述D维的特征向量转化为1维的权重值；通过预设激活函数对所述权重值进行处理，得到相对位置特征。
在本公开实施例中，通过特征化处理可以将相对位置参数转换为图卷积神经网络的边所需的数据格式，便于后续通过图卷积神经网络进行特征提取。
在一种可能的实现方式中，目标特征确定模块，包括第一目标特征确定子模块、第二目标特征确定子模块和第三目标特征确定子模块，其中，第一目标特征确定子模块，用于确定目标区域中的像素数据，对所述像素数据进行特征提取，得到视觉特征；第二目标特征确定子模块，用于确定目标区域中的文本字符，对所述文本字符进行特征提取，得到字符特征；第三目标特征确定子模块，用于根据提取到的视觉特征和字符特征，确定目标区域的目标特征。
在本公开实施例中，考虑到图像中会存在由于拍照视角、光线、遮挡等原因带来的干扰因素，因此，通过文字检测识别通常会有较多的误识，即可能会识别出错误的文本字符，这可能会影响关键信息提取的准确性。而通过视觉信息的提取，将视觉信息考虑到关键信息提取中，会降低文本误识对关键信息提取的影响。即使文本识别错误，但由于视觉信息不会改变太大，因此二者结合能够提高关键信息提取结果的准确性。
在一种可能的实现方式中，第三目标特征确定子模块，用于将所述视觉特征和字符特征赋予不同的权重；对赋予权重后的所述视觉特征和字符特征进行融合，得到目标区域的目标特征。
在本公开实施例中，通过对视觉特征和字符特征赋予不同的权重，能够提高关键信息提取结果的准确性。
在一种可能的实现方式中，所述装置通过预先构建的分类网络实现，所述装置还包括：第一训练模块，用于将样本图像输入所述分类网络中处理，得到样本图像中待提取文本的第一预测类别，以及所述第一预测类别中各个类别之间的对应关系；第二训练模块，用于根据所述第一预测类别，以及所述样本图像的标注类别，训练所述分类网络，所述标注类别包括：表征文本属于预设字段的标识的类别，以及表征文本属于预设字段的字段值的类别；第三训练模块，用于根据所述对应关系，以及标注的待提取文本之间的对应关系，训练所述分类网络。
在本公开实施例中，通过对样本图像的类别和各个类别之间的对应关系进行标注，能够更准确地对分类网络进行训练，训练得到的分类网络在对没有适配模板的图像进行文本提取时，准确性较高。
在一种可能的实现方式中，所述图像包括下述至少一种：收据图像、发票图像、名片图像。
根据本公开的一方面，提供了一种电子设备，包括：处理器；用于存储处理器可执行指令的存储器；其中，所述处理器被配置为调用所述存储器存储的指令，以执行上述方法。
根据本公开的一方面，提供了一种计算机可读存储介质，其上存储有计算机程序指令，所述计算机程序指令被处理器执行时实现上述方法。
根据本公开的一方面，提供了一种计算机程序，包括计算机可读代码，当所述计算机可读代码在电子设备中运行时，所述电子设备中的处理器执行用于实现上述方法的指令。
在本公开实施例中，能够通过图卷积神经网络，基于各目标区域之间的相对位置特征以及待提取文本的特征，确定图像中的待提取文本对应的字段。可不依赖于固定的模 板进行文本提取，相对于基于模板进行文本提取的方式，在对没有适配模板的图像进行文本提取时，准确性较高。
应当理解的是，以上的一般描述和后文的细节描述仅是示例性和解释性的，而非限制本公开。根据下面参考附图对示例性实施例的详细说明，本公开的其它特征及方面将变得清楚。
附图说明
此处的附图被并入说明书中并构成本说明书的一部分，这些附图示出了符合本公开的实施例，并与说明书一起用于说明本公开的技术方案。
图1示出根据本公开实施例的图像处理方法的流程图；
图2示出根据本公开实施例的一种连通图的结构示意图；
图3示出根据本公开实施例的一种分类网络的结构示意图；
图4示出根据本公开实施例的一种图像处理装置的框图；
图5示出根据本公开实施例的一种电子设备的框图；
图6示出根据本公开实施例的一种电子设备的框图。
具体实施方式
以下将参考附图详细说明本公开的各种示例性实施例、特征和方面。附图中相同的附图标记表示功能相同或相似的元件。尽管在附图中示出了实施例的各种方面，但是除非特别指出，不必按比例绘制附图。
在这里专用的词“示例性”意为“用作例子、实施例或说明性”。这里作为“示例性”所说明的任何实施例不必解释为优于或好于其它实施例。
本文中术语“和/或”，仅仅是一种描述关联对象的关联关系，表示可以存在三种关系，例如，A和/或B，可以表示：单独存在A，同时存在A和B，单独存在B这三种情况。另外，本文中术语“至少一种”表示多种中的任意一种或多种中的至少两种的任意组合，例如，包括A、B、C中的至少一种，可以表示包括从A、B和C构成的集合中选择的任意一个或多个元素。
另外，为了更好地说明本公开，在下文的具体实施方式中给出了众多的具体细节。本领域技术人员应当理解，没有某些具体细节，本公开同样可以实施。在一些实例中，对于本领域技术人员熟知的方法、手段、元件和电路未作详细描述，以便于凸显本公开的主旨。
随着人工智能技术的发展，基于图像进行关键信息提取的技术取得了长足的发展，在进行关键信息提取时，可以将图像中的文本识别出来，另外，还会确定识别出来的文本的结构化信息，即确定识别出的某一文本对应结构化数据中的哪个字段，以方便后续对识别出的数据进行结构化地存储、展示等操作。
为了提高关键信息提取的准确性，本公开实施例提供了一种图像处理方法，可以基于各目标区域之间的相对位置特征，以及待提取文本的特征，通过图卷积神经网络，确定图像中的待提取文本对应的字段。该方法可不依赖于固定的模板进行文字提取，相对于基于模板进行文字信息提取的方式，在对没有适配模板的图像进行文字信息提取时，准确性较高。
本公开实施例提供的图像处理方法，可应用于图像中关键信息的提取，可以实现诸如收据信息提取、发票信息提取、身份信息提取等功能，具备较高的应用价值。
图1示出根据本公开实施例的图像处理方法的流程图，如图1所示，所述图像处理方法包括：
步骤S11，对图像进行识别，确定图像中的多个目标区域。
目标区域为待提取文本所在区域。
由于待提取文本在图像上的分布往往是较为分散的，例如，文本“总价”和“19.88元”之间是有一定间隔的，因此，在确定目标区域时，可以根据文本在图像上的分布关系，以文本之间的间隔为依据，对图像进行划分，得到多个目标区域。另外也可以是根据其它方式对目标区域进行划分，具体划分方式可以视本公开的具体应用场景而定，本公开对此不作限制。
在进行目标区域的确定后，可以将构成一个词、构成一句话或表达某一含义的文本所在的区域确定为1个目标区域，例如，待提取文本“总价”所在的区域为1个目标区域，“19.88元”所在的区域为1个目标区域。
对于具体确定图像中目标区域的方式，本公开对此不作限制。
步骤S12，确定图像中各目标区域之间的相对位置特征。
相对位置特征能够表征各目标区域之间的相对位置关系，具体的相对位置特征可以根据2个目标区域的中心点确定，也可以是根据2个目标区域的某个顶点确定，本公开对此不作限制。另外，本公开中的相对位置特征还可以根据一些其它参数确定，具体将会在后文本公开可能的实现方式中进行论述，此处不做赘述。
步骤S13，确定各目标区域的目标特征。
目标特征包括待提取文本的特征。待提取文本的特征为待提取文本自身的特征，该特征即可以包括待提取文本整体上的视觉特征，也可以包括待提取文本的文本字符的特征，或者上述两个特征中的一个。
步骤S14，通过图卷积神经网络，对相对位置特征和目标特征进行特征提取，得到提取后的特征。
将相对位置特征和目标特征输入图卷积神经网络中，进行特征提取，可以得到提取后的特征。
图卷积神经网络在对特征进行提取时，能够以连通图的形式表示图像，对特征进行提取。连通图由若干个结点(Node)及连接两个结点的边(Edge)所构成，边用于刻画不同结点之间的关系。
因此，通过图卷积神经网络提取后的特征，能够准确地表征各目标区域之间的相对位置和待提取文本的特征，以提高后续文本提取时的准确性。
步骤S15，根据提取后的特征，确定待提取文本对应的字段。
根据提取后的特征确定待提取文本对应的字段时，具体可通过训练好的网络来实现，该网络可以根据提取后的特征对待提取文本进行分类，分类的类别用于表征与待提取文本对应的字段。在根据提取后的特征确定了待提取文本的类别后，即确定了待提取文本对应的字段。
后文将会对网络的训练过程进行描述，此处不做赘述。
根据本公开的实施例，能够通过图卷积神经网络，基于各目标区域之间的相对位置特征以及待提取文本的特征，确定图像中的待提取文本对应的字段。可不依赖于固定的模板进行文本提取，相对于基于模板进行文本提取的方式，在对没有适配模板的图像进行文本提取时，准确性较高。
在一种可能的实现方式中，确定图像中各目标区域之间的相对位置特征，包括：确定图像中的第一目标区域和第二目标区域的相对位置参数；对相对位置参数进行特征化处理，得到第一目标区域和第二目标区域的相对位置特征。
这里的第一目标区域和第二目标区域为图像中的任意两个目标区域。
其中，图像中的第一目标区域和第二目标区域的相对位置参数包括下述至少一种：
第一目标区域相对于第二目标区域的横向距离和纵向距离；
第一目标区域的宽高比；
第二目标区域的宽高比；
第一目标区域和第二目标区域的相对尺寸关系。
其中，第一目标区域相对于第二目标区域的横向距离和纵向距离，可以是第一目标区域的参考点与第二目标区域的参考点的横向距离和纵向距离，对于目标区域参考点的选取，可以是目标区域的中心点，也可以是目标区域的某个顶点，对具体参考点的选取，本公开不作限制。
为便于更清楚地理解相对位置特征的确定过程，下面通过具体的数学表达式来对相对位置特征的确定过程进行说明，需要说明的是，本公开中提供的具体数学表达式为本公开实施例在具体实施时的一种可能的实现方式，而不应当理解为对本公开实施例保护范围的限制。
对于一个待提取文本而言，其所在的目标区域往往是矩形的，那么对于待提取文本t i，可以表示为t i＝<x i,y i,h i,w i,s i>，其中，x i,y i分别表示目标区域的参考点在预设坐标系中的横坐标和纵坐标，h i,w i分别表示目标区域的高度和宽度，s i表示待提取文本的字符。
那么，在一种可能的实现方式中，第一目标区域相对于第二目标区域的横向距离Δx ij和纵向距离Δy ij，表达式如下：
Δx ij＝|x i-x j|              (1)
Δy ij＝|y i-y j|             (2)
其中，第一目标区域为待提取文本t i所在的区域，第二目标区域为待提取文本t j所在的区域。
在一种可能的实现方式中，还可以对横向距离Δx ij和纵向距离Δy ij进行归一化处理，得到归一化后的横向距离和纵向距离，具体可通过图像的尺寸参数对Δx ij和Δy ij进行归一化。例如，在通过图像的宽度W进行归一化时，得到相对位置参数
Figure PCTCN2020077247-appb-000001
的表达式如下：
Figure PCTCN2020077247-appb-000002
另外，也可用图像的高H来进行归一化，此处不作赘述。
通过对横向距离Δx ij和纵向距离Δy ij进行归一化处理，降低了被识别图像的放大或缩小对最终结果的影响，使得关键信息的提取结果更加准确。
在一种可能的实现方式中，第一目标区域的宽高比为w i/h i，第二目标区域的宽高比为w j/h j。
第一目标区域和第二目标区域的相对尺寸关系，可以表征第一目标区域尺寸和第二目标区域尺寸之间的相对大小关系。由于某些字段的文本的尺寸之间具备一些特定的关系，因此，相对位置特征中考虑了第一目标区域和第二目标区域的相对尺寸关系，可以使得关键信息的提取结果更加准确。
例如，文本“地址”的尺寸较短，文本“xx市xx街道xx路xx号”的尺寸较长，因此，这两个尺寸之间的差距较大；而文本“总价”和“19.88元”的尺寸之间的差距较小。因此，目标区域的相对尺寸关系能够在一定程度上反应文本对应的字段类别。
在一种可能的实现方式中，相对尺寸关系
Figure PCTCN2020077247-appb-000003
的表达式如下：
Figure PCTCN2020077247-appb-000004
在一种可能的实现方式中，对上述公式涉及的相对位置参数进行整合，得到的整合后的相对位置参数的表达式如下：
Figure PCTCN2020077247-appb-000005
该实现方式中，相对位置参数即包含了归一化后的横向距离和纵向距离，也包含了第一目标区域的宽高比，也包含了第一目标区域和第二目标区域的相对尺寸关系，可以使得关键信息的提取结果更加准确。
在一种可能的实现方式中，在得到相对位置参数后，可以对相对位置参数进行特征化处理，得到第一目标区域和第二目标区域的相对位置特征。
对相对位置参数进行特征化处理，得到第一目标区域和第二目标区域的相对位置特征，包括：将相对位置参数通过正余弦变换矩阵映射到一个D维的空间，得到D维的特征向量，D为正整数；将D维的特征向量乘以一个预设权重矩阵，得到一个1维的权重值；通过预设激活函数对权重值进行处理，得到相对位置特征。
这里的正余弦变换矩阵为傅里叶正弦变换或者余弦变换时所使用的变换矩阵。
这里的预设权重矩阵的具体值可通过网络训练来确定，初始值可以通过随机等方式确定，在进行网络训练时，将会对预设权重矩阵进行调优。后文将会对网络的训练过程进行描述，此处不做赘述。
这里的预设激活函数例如可以是线性整流函数(Rectified Linear Unit,ReLU)，具体的激活函数可以视本公开的实际应用场景而定，本公开对此不作限制。
为便于理解对相对位置参数进行特征化处理的过程，下面通过具体的表达式来说明特征化处理后的相对位置特征e ij，具体请参见公式(6)：
e ij＝ReLU(W mM(r ij))        (6)
其中，M表示正余弦变换矩阵，M(r ij)表示将所述相对位置参数r ij通过正余弦变换矩阵M映射到一个D维的空间，W m为预设权重矩阵，ReLU表示线性整流函数。
本公开实施例中，通过特征化处理可以将相对位置参数转换为图卷积神经网络的边所需的数据格式，便于后续通过图卷积神经网络进行特征提取。
如前文所述，本公开实施例中的目标特征即可以包括待提取文本整体上的视觉特征，也可以包括待提取文本的文本字符的特征。
那么，在一种可能的实现方式中，确定各目标区域的目标特征，包括：确定目标区域中的像素数据，对像素数据进行特征提取，得到视觉特征；确定目标区域中的文本字符，对文本字符进行特征提取，得到文本字符特征；根据提取到的视觉特征和字符特征，确定目标区域的目标特征。
其中，视觉特征可以反映目标区域中文本在整体上的视觉信息。在提取视觉特征时，具体可以通过感兴趣区域对齐(Region of Interest Align，RoI Align)方法来提取，对于具体提取视觉特征的方式，本公开不作限制。
本公开实施例中，考虑到图像中会存在由于拍照视角、光线、遮挡等原因带来的干扰因素，因此，通过文字检测识别通常会有较多的误识，即可能会识别出错误的文本字符，这可能会影响关键信息提取的准确性。而通过视觉信息的提取，将视觉信息考虑到关键信息提取中，会降低文本误识对关键信息提取的影响。即使文本识别错误，但由于视觉信息不会改变太大，因此二者结合能够提高关键信息提取结果的准确性。
确定目标区域中的文本字符时，可以通过文字识别技术对文本字符进行识别提取。 例如，可以通过光学字符识别技术(Optical Character Recognition，OCR)对文本字符进行特征提取，得到文本字符。对于具体提取文本字符的方式，本公开不做限制。
在一种可能的实现方式中，对所述文本字符进行特征提取，得到字符特征，包括：通过独热(one-hot)编码的方式将文本字符映射到一个低维特征空间；然后通过双向的长短时序网络(Bi-LSTM)对低维特征空间中的文本字符进行处理，得到文本的特征表示，即得到了待提取文本的字符特征。
通过独热编码，可以将离散特征(文本字符)的取值扩展到欧式空间，离散特征的某个取值对应于欧式空间中的某个点，会让特征之间的计算更加合理。
在一种可能的实现方式中，根据提取到的视觉特征和字符特征，确定目标区域的目标特征，包括：将所述视觉特征和字符特征赋予不同的权重；对赋予权重后的所述视觉特征和字符特征进行融合(例如相加)，得到目标区域的目标特征。
考虑到视觉特征和字符特征对提取结果的影响可能不同，因此这里通过对视觉特征和字符特征赋予不同的权重，来提高提取结果的准确性。这里的权重可以是通过网络训练进行优化得到的，具体训练过程详见后文描述，此处不做赘述。
为便于理解对文字字符进行特征化处理的过程，下面通过具体的表达式来说明特征化处理后的字符特征。
对于一个待提取文本而言，对文本字符s i进行特征提取得到字符特征t i的过程可表示为可用公式(7)。
Figure PCTCN2020077247-appb-000006
其中，W∈R C×D表示独热编码的投影矩阵，Bi-LSTM表示通过双向长短时序网络对独热编码后的文本字符进行处理，
Figure PCTCN2020077247-appb-000007
表示文本字符s i中的第j个字符。
通过对字符特征t i赋予权重α i，对视觉特征v i赋予权重(1-α i)，得到目标特征n i具体详见公式(8)和(9)。
α i＝σ(W tt i+W vv i)          (8)
n i＝α i U tt i+(1-α i)U vv i         (9)
其中，W t∈R 1×Dt和W v∈R 1×Dv为一维投影矩阵，具体可以通过网络训练进行优化得到，σ为激活函数。U t∈R Dh×Dt和U v∈R Dh×Dt为投影参数，也可以通过网络训练进行优化得到。
在得到目标特征n i和相对位置特征e ij后，即可通过图卷积神经网络，对相对位置特征和目标特征进行特征提取。
在一种可能的实现方式中，通过图卷积神经网络，对相对位置特征和目标特征进行特征提取，得到提取后的特征，包括：以各目标特征为图的节点，以各相对位置特征为连接两个节点的边，构建连通图；通过图卷积神经网络，对连通图进行迭代更新，将迭代更新后满足收敛条件的连通图作为提取后的特征。
在将目标区域的相对位置特征作为连接两个节点的边构建连通图时，会将相对位置特征作为节点之间的邻接矩阵的一个参数，当然邻接矩阵中还可以包含节点的语义相似性等其它参数，本公开对其它参数的具体设置不作限制。
请参阅图2，为本公开提供的一种连通图的示意图，该连通图中，图的节点为各目标特征，连接两个节点的边为目标区域的相对位置特征。
本公开实施例构建的连通图中，既包含了图像中的目标特征，也包含了图像中目标特征之间的相对位置特征，可以从整体上表征图像中文字的特征，因此能够提高关键信息提取结果的准确性。
在构建好连通图后，可以通过图卷积神经网络，对连通图进行迭代更新，将迭代更新后满足收敛条件的连通图作为提取后的特征。每次迭代过程中，任意节点i的特征，是通过与节点i相连的各节点的邻接矩阵，对各节点的特征值进行投影来更新的，在经过多次迭代后，各节点的特征值将不会再随迭代次数的增加而变化，即节点的特征值保持不变，此时即可视为满足收敛条件，满足收敛条件的连通图即可作为提取后的特征。
为便于理解，第l+1次迭代时节点N的特征N l+1的表达式如下：
N l+1＝σ((A lN l)W l)         (10)
其中，N l为第l次迭代时节点N的特征，W l为转换矩阵，可以通过网络训练进行优化得到，A l为节点的邻接矩阵，节点i和j的邻接矩阵A l ij的表达式如下：
Figure PCTCN2020077247-appb-000008
Figure PCTCN2020077247-appb-000009
其中，(n l i) T表示n l i的转置，
Figure PCTCN2020077247-appb-000010
表示归一化参数，可以通过网络训练进行优化得到。
在得到提取后的特征后，在一种可能的实现方式中，根据提取后的特征，确定待提取文本对应的字段，包括：根据预先定义的多个预设类别，对图卷积神经网络输出的连通图中的节点进行分类，得到节点的类别，所述预设类别包括：表征文本属于预设字段的标识的类别，以及表征文本属于预设字段的字段值的类别；根据所述节点的类别，确定待提取文本对应于预设字段的标识或字段值。
由于识别到的文本中，即可能有表征预设字段的标识的文本，也可能有表征预设字段的字段值的文本。表征预设字段的标识的文本，为图像中用来指示字段值属于哪个字段的文本，而字段值是字段下的具体值，例如，对于预设字段“总价”而言，图像中识别出的文本“总价”、“总价格”或者“sub total”等，都是预设字段“总价”的一种具体的标识；对于识别出的文本“19.88元”、“￥：19.88”等，都是预设字段的字段值。
因此，对于某一个预设字段，可以设置2个类别分别对应该预设字段，其中，1个类别是表征文本属于预设字段的标识的类别，1个类别是表征文本属于预设字段的字段值的类别。当有多个不同的预设字段时，每一预设字段均可以设置2个类别，如此便会有多个表征文本属于预设字段的标识的类别，以及多个表征文本属于预设字段的字段值的类别。
例如，在针对商品购物小票进行识别时，预设字段可设置为“名称”、“地址”、“电话号码”、“日期”、“时间”、“商品类目”、“商品名称”、“商品单价”、“单品总价”、“税费”、“合计总价”、“提示”，共计12个预设字段，那么可以预设24个类别，分别表示各预设字段的预设字段标识，以及各预设字段的字段值。另外，还可以设置类别“其它”，以将不属于上述类别的文本进行区分提取，即共计设置25个类别。
上述举例中的25个具体预设类别示例如下：
名称-标识；名称-字段值；地址-标识；地址-字段值；电话号码-标识；电话号码-字段值；日期-标识；日期-字段值、时间-标识；时间-字段值、商品类目-标识；商品类目-字段值；商品名称-标识；商品名称-字段值；商品单价-标识；商品单价-字段值；单品总价-标识；单品总价-字段值；税费-标识；税费-字段值；合计总价-标识；合计总价-字段值；提示-标识；提示-字段值；其它。
在一种可能的实现方式中，本公开实施例的图像处理方法可通过预先构建的分类网络实现，该分类网络的训练步骤如下：
将样本图像输入所述分类网络中处理，得到样本图像中待提取文本的第一预测类别，以及所述第一预测类别中各个类别之间的对应关系；
根据所述第一预测类别，以及所述样本图像的标注类别，训练所述分类网络，所述标注类别包括：表征文本属于预设字段的标识的类别，以及表征文本属于预设字段的字段值的类别；
根据所述对应关系，以及标注的标注类别之间的对应关系，训练所述分类网络。
该分类网络可用于实现本公开的图像处理技术，该分类网络中可以包含前文所述的图卷积神经网络，另外，为实现本公开的各项功能，该分类网络中还可以包含其它网络，例如，Bi-LSTM网络，对于本公开的分类网络中包含的网络，可根据本公开实施例的具体应用场景而定，本公开对此不做限制。
请参阅图3，为本申请提供的一种分类网络的具体实现方式的结构示意图，网络中包含目标特征提取模块、相对位置特征提取模块、卷积网络特征提取模块、分类模块。通过目标特征提取模块提取包含待提取文本的图像的目标特征，通过相对位置特征提取模块提取图像的相对位置特征；将目标特征和相对位置特征输入至卷积网络特征提取模块，进行迭代更新，得到迭代提取后的特征；然后通过分类模块对迭代提取后的特征进行分类，得到节点的预测类别。由于类别表征与待提取文本对应的字段，在根据提取后的特征确定了待提取文本的类别后，即确定了待提取文本对应的字段。对于各个模块具体功能的实现，请参考本公开中的相关论述，此处不做赘述。
在上述训练过程中，标注类别可以是上文所述的预设类别，此处不再赘述。
在根据所述第一预测类别，以及所述样本图像的标注类别，训练所述分类网络时，可以根据第一预测类别相对于标注类别的损失，调整分类网络中的参数，以使分类网络对样本图像的预测类别与标注类别之间的差异最小。
此外，在训练时，利用两个文本是否分别属于同一预设字段的标识和标识值，对分类网络的分类准确度也有益处。为便于后续描述，这里将分别属于同一预设字段的标识和标识值的两个文本称为字段对，例如：文本“总价”和“19.88元”构成字段对。
因此，在训练所述分类网络时，分类网络还会输出第一预测类别中各个类别之间的对应关系，同时，样本图像中也会对文本之间的对应关系进行标注。那么，便可以根据分类网络输出的对应关系，以及标注的待提取文本之间的对应关系，训练所述分类网络。
训练时所使用的损失函数具体可以是交叉熵损失函数(Cross Entropy Loss，CE)，具体的损失函数可以根据实际需求选择，本公开对此不作具体限定。
根据本公开的实施例，训练后的分类网络可用于在文字关键信息提取时确定待提取文本对应的字段，具体详见本公开提供的实施例，由于在训练时利用了待提取文本之间的对应关系，因此训练得到的分类网络在对没有适配模板的图像进行文本提取时，准确性较高。
在一种可能的实现方式中，识别的图像包括下述至少一种：收据图像、发票图像、名片图像。当然，在实际应用中，本公开的实施例也可用于对其它图像的识别，本公开对此不作具体限定。
根据本公开的实施例，能够通过图卷积神经网络，基于各目标区域之间的相对位置特征以及待提取文本的特征，确定图像中的待提取文本对应的字段。可不依赖于固定的模板进行文本提取，相对于基于模板进行文本提取的方式，在对没有适配模板的图像进行文本提取时，准确性较高。
根据本公开的实施例，在进行文本提取时，不仅利用目标区域中的文本字符特征，还利用了目标区域的视觉特征，降低了文本字符误识对最后分类的影响，提高了文本提取时的准确性；另外，通过建立了文本区域之间的空间位置联系，可以不依赖于预先设计好的模板，可以处理未见过的模板，有更好的拓展性。
在一种可能的实现方式中，所述图像处理方法可以由终端设备或服务器等电子设备执行，终端设备可以为用户设备(User Equipment，UE)、移动设备、用户终端、终端、蜂窝电话、无绳电话、个人数字处理(Personal Digital Assistant，PDA)、手持设备、计算设备、车载设备、可穿戴设备等，所述方法可以通过处理器调用存储器中存储的计算机可读指令的方式来实现。或者，可通过服务器执行所述方法。
可以理解，本公开提及的上述各个方法实施例，在不违背原理逻辑的情况下，均可以彼此相互结合形成结合后的实施例，限于篇幅，本公开不再赘述。本领域技术人员可以理解，在具体实施方式的上述方法中，各步骤的具体执行顺序应当以其功能和可能的内在逻辑确定。
此外，本公开还提供了图像处理装置、电子设备、计算机可读存储介质、程序，上述均可用来实现本公开提供的任一种图像处理方法，相应技术方案和描述参见方法部分的相应记载，不再赘述。
图4示出根据本公开实施例的图像处理装置的框图，如图4所示，所述图像处理装置20包括：
识别模块21，用于对图像进行识别，确定所述图像中的多个目标区域，所述目标区域为待提取文本所在区域；
相对位置特征确定模块22，用于确定所述图像中各目标区域之间的相对位置特征；
目标特征确定模块23，用于确定各所述目标区域的目标特征，所述目标特征包括所述待提取文本的特征；
图卷积模块24，用于通过图卷积神经网络，对所述相对位置特征和所述目标特征进行特征提取，得到提取后的特征；
字段确定模块25，用于根据提取后的特征，确定所述待提取文本对应的字段。
在本公开实施例中，能够通过图卷积神经网络，基于各目标区域之间的相对位置特征以及待提取文本的特征，确定图像中的待提取文本对应的字段。可不依赖于固定的模板进行文本提取，相对于基于模板进行文本提取的方式，在对没有适配模板的图像进行文本提取时，准确性较高。
在一种可能的实现方式中，图卷积模块24包括：第一图卷积子模块和第二图卷积子模块，其中：
第一图卷积子模块，用于以各所述目标特征为图的节点，以各所述相对位置特征为连接两个节点的边，构建连通图；
第二图卷积子模块，用于通过图卷积神经网络，对所述连通图进行迭代更新，将迭代更新后满足收敛条件的连通图作为提取后的特征。
在本公开实施例中，构建的连通图既包含了图像中的目标特征，也包含了图像中目标特征之间的相对位置特征，可以从整体上表征图像中文字的特征，因此能够提高关键信息提取结果的准确性。
图卷积神经网络在对特征进行提取时，能够以连通图的形式表示图像，对特征进行提取。连通图由若干个结点(Node)及连接两个结点的边(Edge)所构成，边用于刻画不同结点之间的关系。因此，通过图卷积神经网络提取后的特征，能够准确地表征各目标区域之间的相对位置和待提取文本的特征，以提高后续文本提取时的准确性。
在一种可能的实现方式中，字段确定模块25包括：第一字段确定子模块和第二字段确定子模块，其中：
第一字段确定子模块，用于根据预先定义的多个预设类别，对图卷积神经网络输出的连通图中的节点进行分类，得到节点的类别，所述预设类别包括：表征文本属于预设 字段的标识的类别，以及表征文本属于预设字段的字段值的类别；
第二字段确定子模块，用于根据所述节点的类别，确定待提取文本对应于预设字段的标识或字段值。
在本公开实施例中，通过预先定义预设类别为预设字段的标识或字段，根据提取的特征对待提取文本进行分类，即可得到待提取文本对应于预设字段的标识或字段值，提高了文本提取时的准确性。
在一种可能的实现方式中，相对位置特征确定模块22包括：第一相对位置特征确定子模块和第二相对位置特征确定子模块，其中：
第一相对位置特征确定子模块，用于确定图像中的第一目标区域和第二目标区域的相对位置参数；
第二相对位置特征确定子模块，用于对所述相对位置参数进行特征化处理，得到第一目标区域和第二目标区域的相对位置特征。
在一种可能的实现方式中，所述相对位置参数包括下述至少一种：
第一目标区域相对于第二目标区域的横向距离和纵向距离；
所述第一目标区域的宽高比；
所述第二目标区域的宽高比；
所述第一目标区域和所述第二目标区域的相对尺寸关系。
在本公开实施例中，相对位置参数即包含了横向距离和纵向距离，也包含了第一目标区域的宽高比，也包含了第一目标区域和第二目标区域的相对尺寸关系，可以使得关键信息的提取结果更加准确。
在一种可能的实现方式中，第二相对位置特征确定子模块，用于将所述相对位置参数通过正余弦变换矩阵映射到一个D维的空间，得到D维的特征向量，D为正整数；通过预设权重矩阵，将所述D维的特征向量转化为1维的权重值；通过预设激活函数对所述权重值进行处理，得到相对位置特征。
在本公开实施例中，通过特征化处理可以将相对位置参数转换为图卷积神经网络的边所需的数据格式，便于后续通过图卷积神经网络进行特征提取。
在一种可能的实现方式中，目标特征确定模块23，包括第一目标特征确定子模块、第二目标特征确定子模块和第三目标特征确定子模块，其中：
第一目标特征确定子模块，用于确定目标区域中的像素数据，对所述像素数据进行特征提取，得到视觉特征；
第二目标特征确定子模块，用于确定目标区域中的文本字符，对所述文本字符进行特征提取，得到字符特征；
第三目标特征确定子模块，用于根据提取到的视觉特征和字符特征，确定目标区域的目标特征。
在本公开实施例中，考虑到图像中会存在由于拍照视角、光线、遮挡等原因带来的干扰因素，因此，通过文字检测识别通常会有较多的误识，即可能会识别出错误的文本字符，这可能会影响关键信息提取的准确性。而通过视觉信息的提取，将视觉信息考虑到关键信息提取中，会降低文本误识对关键信息提取的影响。即使文本识别错误，但由于视觉信息不会改变太大，因此二者结合能够提高关键信息提取结果的准确性。
在一种可能的实现方式中，第三目标特征确定子模块，用于将所述视觉特征和字符特征赋予不同的权重；对赋予权重后的所述视觉特征和字符特征进行融合，得到目标区域的目标特征。
在本公开实施例中，通过对视觉特征和字符特征赋予不同的权重，能够提高关键信息提取结果的准确性。
在一种可能的实现方式中，所述装置通过预先构建的分类网络实现，所述装置还包 括：
第一训练模块，用于将样本图像输入所述分类网络中处理，得到样本图像中待提取文本的第一预测类别，以及所述第一预测类别中各个类别之间的对应关系；
第二训练模块，用于根据所述第一预测类别，以及所述样本图像的标注类别，训练所述分类网络，所述标注类别包括：表征文本属于预设字段的标识的类别，以及表征文本属于预设字段的字段值的类别；
第三训练模块，用于根据所述对应关系，以及标注的待提取文本之间的对应关系，训练所述分类网络。
在本公开实施例中，通过对样本图像的类别和各个类别之间的对应关系进行标注，能够更准确地对分类网络进行训练，训练得到的分类网络在对没有适配模板的图像进行文本提取时，准确性较高。
在一种可能的实现方式中，所述图像包括下述至少一种：收据图像、发票图像、名片图像。
在一些实施例中，本公开实施例提供的装置具有的功能或包含的模块可以用于执行上文方法实施例描述的方法，其具体实现可以参照上文方法实施例的描述，为了简洁，这里不再赘述。
本公开实施例还提出一种计算机可读存储介质，其上存储有计算机程序指令，所述计算机程序指令被处理器执行时实现上述方法。计算机可读存储介质可以是易失性计算机可读存储介质或非易失性计算机可读存储介质。
本公开实施例还提出一种电子设备，包括：处理器；用于存储处理器可执行指令的存储器；其中，所述处理器被配置为调用所述存储器存储的指令，以执行上述方法。
本公开实施例还提供了一种计算机程序产品，包括计算机可读代码，当计算机可读代码在设备上运行时，设备中的处理器执行用于实现如上任一实施例提供的图像处理方法的指令。
本公开实施例还提供了另一种计算机程序产品，用于存储计算机可读指令，指令被执行时使得计算机执行上述任一实施例提供的图像处理方法的操作。
电子设备可以被提供为终端、服务器或其它形态的设备。
图5示出根据本公开实施例的一种电子设备800的框图。例如，电子设备800可以是移动电话，计算机，数字广播终端，消息收发设备，游戏控制台，平板设备，医疗设备，健身设备，个人数字助理等终端。
参照图5，电子设备800可以包括以下一个或多个组件：处理组件802，存储器804，电源组件806，多媒体组件808，音频组件810，输入/输出(I/O)的接口812，传感器组件814，以及通信组件816。
处理组件802通常控制电子设备800的整体操作，诸如与显示，电话呼叫，数据通信，相机操作和记录操作相关联的操作。处理组件802可以包括一个或多个处理器820来执行指令，以完成上述的方法的全部或部分步骤。此外，处理组件802可以包括一个或多个模块，便于处理组件802和其他组件之间的交互。例如，处理组件802可以包括多媒体模块，以方便多媒体组件808和处理组件802之间的交互。
存储器804被配置为存储各种类型的数据以支持在电子设备800的操作。这些数据的示例包括用于在电子设备800上操作的任何应用程序或方法的指令，联系人数据，电话簿数据，消息，图像，视频等。存储器804可以由任何类型的易失性或非易失性存储设备或者它们的组合实现，如静态随机存取存储器(SRAM)，电可擦除可编程只读存储器(EEPROM)，可擦除可编程只读存储器(EPROM)，可编程只读存储器(PROM)，只读存储器(ROM)，磁存储器，快闪存储器，磁盘或光盘。
电源组件806为电子设备800的各种组件提供电力。电源组件806可以包括电源管理系统，一个或多个电源，及其他与为电子设备800生成、管理和分配电力相关联的组件。
多媒体组件808包括在所述电子设备800和用户之间的提供一个输出接口的屏幕。在一些实施例中，屏幕可以包括液晶显示器(LCD)和触摸面板(TP)。如果屏幕包括触摸面板，屏幕可以被实现为触摸屏，以接收来自用户的输入信号。触摸面板包括一个或多个触摸传感器以感测触摸、滑动和触摸面板上的手势。所述触摸传感器可以不仅感测触摸或滑动动作的边界，而且还检测与所述触摸或滑动操作相关的持续时间和压力。在一些实施例中，多媒体组件808包括一个前置摄像头和/或后置摄像头。当电子设备800处于操作模式，如拍摄模式或视频模式时，前置摄像头和/或后置摄像头可以接收外部的多媒体数据。每个前置摄像头和后置摄像头可以是一个固定的光学透镜系统或具有焦距和光学变焦能力。
音频组件810被配置为输出和/或输入音频信号。例如，音频组件810包括一个麦克风(MIC)，当电子设备800处于操作模式，如呼叫模式、记录模式和语音识别模式时，麦克风被配置为接收外部音频信号。所接收的音频信号可以被进一步存储在存储器804或经由通信组件816发送。在一些实施例中，音频组件810还包括一个扬声器，用于输出音频信号。
I/O接口812为处理组件802和外围接口模块之间提供接口，上述外围接口模块可以是键盘，点击轮，按钮等。这些按钮可包括但不限于：主页按钮、音量按钮、启动按钮和锁定按钮。
传感器组件814包括一个或多个传感器，用于为电子设备800提供各个方面的状态评估。例如，传感器组件814可以检测到电子设备800的打开/关闭状态，组件的相对定位，例如所述组件为电子设备800的显示器和小键盘，传感器组件814还可以检测电子设备800或电子设备800一个组件的位置改变，用户与电子设备800接触的存在或不存在，电子设备800方位或加速/减速和电子设备800的温度变化。传感器组件814可以包括接近传感器，被配置用来在没有任何的物理接触时检测附近物体的存在。传感器组件814还可以包括光传感器，如CMOS或CCD图像传感器，用于在成像应用中使用。在一些实施例中，该传感器组件814还可以包括加速度传感器，陀螺仪传感器，磁传感器，压力传感器或温度传感器。
通信组件816被配置为便于电子设备800和其他设备之间有线或无线方式的通信。电子设备800可以接入基于通信标准的无线网络，如WiFi，2G或3G，或它们的组合。在一个示例性实施例中，通信组件816经由广播信道接收来自外部广播管理系统的广播信号或广播相关信息。在一个示例性实施例中，所述通信组件816还包括近场通信(NFC)模块，以促进短程通信。例如，在NFC模块可基于射频识别(RFID)技术，红外数据协会(IrDA)技术，超宽带(UWB)技术，蓝牙(BT)技术和其他技术来实现。
在示例性实施例中，电子设备800可以被一个或多个应用专用集成电路(ASIC)、数字信号处理器(DSP)、数字信号处理设备(DSPD)、可编程逻辑器件(PLD)、现场可编程门阵列(FPGA)、控制器、微控制器、微处理器或其他电子元件实现，用于执行上述方法。
在示例性实施例中，还提供了一种非易失性计算机可读存储介质，例如包括计算机程序指令的存储器804，上述计算机程序指令可由电子设备800的处理器820执行以完成上述方法。
图6示出根据本公开实施例的一种电子设备1900的框图。例如，电子设备1900可以被提供为一服务器。参照图6，电子设备1900包括处理组件1922，其进一步包括一个或多个处理器，以及由存储器1932所代表的存储器资源，用于存储可由处理组件1922的执行的指令，例如应用程序。存储器1932中存储的应用程序可以包括一个或一个以上的每一个 对应于一组指令的模块。此外，处理组件1922被配置为执行指令，以执行上述方法。
电子设备1900还可以包括一个电源组件1926被配置为执行电子设备1900的电源管理，一个有线或无线网络接口1950被配置为将电子设备1900连接到网络，和一个输入输出(I/O)接口1958。电子设备1900可以操作基于存储在存储器1932的操作系统，例如Windows ServerTM，Mac OS XTM，UnixTM,LinuxTM，FreeBSDTM或类似。
在示例性实施例中，还提供了一种非易失性计算机可读存储介质，例如包括计算机程序指令的存储器1932，上述计算机程序指令可由电子设备1900的处理组件1922执行以完成上述方法。
本公开可以是系统、方法和/或计算机程序产品。计算机程序产品可以包括计算机可读存储介质，其上载有用于使处理器实现本公开的各个方面的计算机可读程序指令。
计算机可读存储介质可以是可以保持和存储由指令执行设备使用的指令的有形设备。计算机可读存储介质例如可以是――但不限于――电存储设备、磁存储设备、光存储设备、电磁存储设备、半导体存储设备或者上述的任意合适的组合。计算机可读存储介质的更具体的例子(非穷举的列表)包括：便携式计算机盘、硬盘、随机存取存储器(RAM)、只读存储器(ROM)、可擦式可编程只读存储器(EPROM或闪存)、静态随机存取存储器(SRAM)、便携式压缩盘只读存储器(CD-ROM)、数字多功能盘(DVD)、记忆棒、软盘、机械编码设备、例如其上存储有指令的打孔卡或凹槽内凸起结构、以及上述的任意合适的组合。这里所使用的计算机可读存储介质不被解释为瞬时信号本身，诸如无线电波或者其他自由传播的电磁波、通过波导或其他传输媒介传播的电磁波(例如，通过光纤电缆的光脉冲)、或者通过电线传输的电信号。
这里所描述的计算机可读程序指令可以从计算机可读存储介质下载到各个计算/处理设备，或者通过网络、例如因特网、局域网、广域网和/或无线网下载到外部计算机或外部存储设备。网络可以包括铜传输电缆、光纤传输、无线传输、路由器、防火墙、交换机、网关计算机和/或边缘服务器。每个计算/处理设备中的网络适配卡或者网络接口从网络接收计算机可读程序指令，并转发该计算机可读程序指令，以供存储在各个计算/处理设备中的计算机可读存储介质中。
用于执行本公开操作的计算机程序指令可以是汇编指令、指令集架构(ISA)指令、机器指令、机器相关指令、微代码、固件指令、状态设置数据、或者以一种或多种编程语言的任意组合编写的源代码或目标代码，所述编程语言包括面向对象的编程语言—诸如Smalltalk、C++等，以及常规的过程式编程语言—诸如“C”语言或类似的编程语言。计算机可读程序指令可以完全地在用户计算机上执行、部分地在用户计算机上执行、作为一个独立的软件包执行、部分在用户计算机上部分在远程计算机上执行、或者完全在远程计算机或服务器上执行。在涉及远程计算机的情形中，远程计算机可以通过任意种类的网络—包括局域网(LAN)或广域网(WAN)—连接到用户计算机，或者，可以连接到外部计算机(例如利用因特网服务提供商来通过因特网连接)。在一些实施例中，通过利用计算机可读程序指令的状态信息来个性化定制电子电路，例如可编程逻辑电路、现场可编程门阵列(FPGA)或可编程逻辑阵列(PLA)，该电子电路可以执行计算机可读程序指令，从而实现本公开的各个方面。
这里参照根据本公开实施例的方法、装置(系统)和计算机程序产品的流程图和/或框图描述了本公开的各个方面。应当理解，流程图和/或框图的每个方框以及流程图和/或框图中各方框的组合，都可以由计算机可读程序指令实现。
这些计算机可读程序指令可以提供给通用计算机、专用计算机或其它可编程数据处理装置的处理器，从而生产出一种机器，使得这些指令在通过计算机或其它可编程数据处理装置的处理器执行时，产生了实现流程图和/或框图中的一个或多个方框中规定的功能/动作的装置。也可以把这些计算机可读程序指令存储在计算机可读存储介质中，这些 指令使得计算机、可编程数据处理装置和/或其他设备以特定方式工作，从而，存储有指令的计算机可读介质则包括一个制造品，其包括实现流程图和/或框图中的一个或多个方框中规定的功能/动作的各个方面的指令。
也可以把计算机可读程序指令加载到计算机、其它可编程数据处理装置、或其它设备上，使得在计算机、其它可编程数据处理装置或其它设备上执行一系列操作步骤，以产生计算机实现的过程，从而使得在计算机、其它可编程数据处理装置、或其它设备上执行的指令实现流程图和/或框图中的一个或多个方框中规定的功能/动作。
附图中的流程图和框图显示了根据本公开的多个实施例的系统、方法和计算机程序产品的可能实现的体系架构、功能和操作。在这点上，流程图或框图中的每个方框可以代表一个模块、程序段或指令的一部分，所述模块、程序段或指令的一部分包含一个或多个用于实现规定的逻辑功能的可执行指令。在有些作为替换的实现中，方框中所标注的功能也可以以不同于附图中所标注的顺序发生。例如，两个连续的方框实际上可以基本并行地执行，它们有时也可以按相反的顺序执行，这依所涉及的功能而定。也要注意的是，框图和/或流程图中的每个方框、以及框图和/或流程图中的方框的组合，可以用执行规定的功能或动作的专用的基于硬件的系统来实现，或者可以用专用硬件与计算机指令的组合来实现。
该计算机程序产品可以具体通过硬件、软件或其结合的方式实现。在一个可选实施例中，所述计算机程序产品具体体现为计算机存储介质，在另一个可选实施例中，计算机程序产品具体体现为软件产品，例如软件开发包(Software Development Kit，SDK)等等。
以上已经描述了本公开的各实施例，上述说明是示例性的，并非穷尽性的，并且也不限于所披露的各实施例。在不偏离所说明的各实施例的范围和精神的情况下，对于本技术领域的普通技术人员来说许多修改和变更都是显而易见的。本文中所用术语的选择，旨在最好地解释各实施例的原理、实际应用或对市场中的技术的改进，或者使本技术领域的其它普通技术人员能理解本文披露的各实施例。

Claims (23)

  1. 一种图像处理方法，其中，包括：
    对图像进行识别，确定所述图像中的多个目标区域，所述目标区域为待提取文本所在区域；
    确定所述图像中各目标区域之间的相对位置特征；
    确定各所述目标区域的目标特征，所述目标特征包括所述待提取文本的特征；
    通过图卷积神经网络，对所述相对位置特征和所述目标特征进行特征提取，得到提取后的特征；
    根据提取后的特征，确定所述待提取文本对应的字段。
  2. 根据权利要求1所述方法，其中，通过图卷积神经网络，对所述相对位置特征和所述目标特征进行特征提取，得到提取后的特征，包括：
    以各所述目标特征为图的节点，以各所述相对位置特征为连接两个节点的边，构建连通图；
    通过图卷积神经网络，对所述连通图进行迭代更新，将迭代更新后满足收敛条件的连通图作为提取后的特征。
  3. 根据权利要求2所述方法，其中，根据提取后的特征，确定所述待提取文本对应的字段，包括：
    根据预先定义的多个预设类别，对图卷积神经网络输出的连通图中的节点进行分类，得到节点的类别，所述预设类别包括：表征文本属于预设字段的标识的类别，以及表征文本属于预设字段的字段值的类别；
    根据所述节点的类别，确定待提取文本对应于预设字段的标识或字段值。
  4. 根据权利要求1-3中任一项所述方法，其中，确定所述图像中各目标区域之间的相对位置特征，包括：
    确定图像中的第一目标区域和第二目标区域的相对位置参数；
    对所述相对位置参数进行特征化处理，得到第一目标区域和第二目标区域的相对位置特征。
  5. 根据权利要求4所述方法，其中，所述相对位置参数包括下述至少一种：
    第一目标区域相对于第二目标区域的横向距离和纵向距离；
    所述第一目标区域的宽高比；
    所述第二目标区域的宽高比；
    所述第一目标区域和所述第二目标区域的相对尺寸关系。
  6. 根据权利要求4或5任一所述方法，其中，对所述相对位置参数进行特征化处理，得到第一目标区域和第二目标区域的相对位置特征，包括：
    将所述相对位置参数通过正余弦变换矩阵映射到一个D维的空间，得到D维的特征向量，D为正整数；
    通过预设权重矩阵，将所述D维的特征向量转化为1维的权重值；
    通过预设激活函数对所述权重值进行处理，得到相对位置特征。
  7. 根据权利要求1-6中任一项所述方法，其中，确定各所述目标区域的目标特征，包括：
    确定目标区域中的像素数据，对所述像素数据进行特征提取，得到视觉特征；
    确定目标区域中的文本字符，对所述文本字符进行特征提取，得到字符特征；
    根据提取到的视觉特征和字符特征，确定目标区域的目标特征。
  8. 根据权利要求7所述方法，其中，根据提取到的视觉特征和字符特征，确定目标区域的目标特征，包括：
    将所述视觉特征和字符特征赋予不同的权重；
    对赋予权重后的所述视觉特征和字符特征进行融合，得到目标区域的目标特征。
  9. 根据权利要求1-8中任一项所述方法，其中，所述方法通过预先构建的分类网络实现，所述分类网络的训练步骤如下：
    将样本图像输入所述分类网络中处理，得到样本图像中待提取文本的第一预测类别，以及所述第一预测类别中各个类别之间的对应关系；
    根据所述第一预测类别，以及所述样本图像的标注类别，训练所述分类网络，所述标注类别包括：表征文本属于预设字段的标识的类别，以及表征文本属于预设字段的字段值的类别；
    根据所述对应关系，以及标注的待提取文本之间的对应关系，训练所述分类网络。
  10. 根据权利要求1-9中任一项所述方法，其中，所述图像包括下述至少一种：收据图像、发票图像、名片图像。
  11. 一种图像处理装置，其中，包括：
    识别模块，用于对图像进行识别，确定所述图像中的多个目标区域，所述目标区域为待提取文本所在区域；
    相对位置特征确定模块，用于确定所述图像中各目标区域之间的相对位置特征；
    目标特征确定模块，用于确定各所述目标区域的目标特征，所述目标特征包括所述待提取文本的特征；
    图卷积模块，用于通过图卷积神经网络，对所述相对位置特征和所述目标特征进行特征提取，得到提取后的特征；
    字段确定模块，用于根据提取后的特征，确定所述待提取文本对应的字段。
  12. 根据权利要求11所述装置，其中，所述图卷积模块包括：第一图卷积子模块和第二图卷积子模块，其中：
    第一图卷积子模块，用于以各所述目标特征为图的节点，以各所述相对位置特征为连接两个节点的边，构建连通图；
    第二图卷积子模块，用于通过图卷积神经网络，对所述连通图进行迭代更新，将迭代更新后满足收敛条件的连通图作为提取后的特征。
  13. 根据权利要求12所述装置，其中，所述字段确定模块包括：第一字段确定子模块和第二字段确定子模块，其中：
    第一字段确定子模块，用于根据预先定义的多个预设类别，对图卷积神经网络输出的连通图中的节点进行分类，得到节点的类别，所述预设类别包括：表征文本属于预设字段的标识的类别，以及表征文本属于预设字段的字段值的类别；
    第二字段确定子模块，用于根据所述节点的类别，确定待提取文本对应于预设字段的标识或字段值。
  14. 根据权利要求11-13中任一项所述装置，其中，相对位置特征确定模块包括：第一相对位置特征确定子模块和第二相对位置特征确定子模块，其中：
    第一相对位置特征确定子模块，用于确定图像中的第一目标区域和第二目标区域的相对位置参数；
    第二相对位置特征确定子模块，用于对所述相对位置参数进行特征化处理，得到第一目标区域和第二目标区域的相对位置特征。
  15. 根据权利要求14所述装置，其中，所述相对位置参数包括下述至少一种：
    第一目标区域相对于第二目标区域的横向距离和纵向距离；
    所述第一目标区域的宽高比；
    所述第二目标区域的宽高比；
    所述第一目标区域和所述第二目标区域的相对尺寸关系。
  16. 根据权利要求14或15任一所述装置，其中，第二相对位置特征确定子模块，用于将所述相对位置参数通过正余弦变换矩阵映射到一个D维的空间，得到D维的特征向 量，D为正整数；通过预设权重矩阵，将所述D维的特征向量转化为1维的权重值；通过预设激活函数对所述权重值进行处理，得到相对位置特征。
  17. 根据权利要求11-16中任一项所述装置，其中，目标特征确定模块，包括第一目标特征确定子模块、第二目标特征确定子模块和第三目标特征确定子模块，其中：
    第一目标特征确定子模块，用于确定目标区域中的像素数据，对所述像素数据进行特征提取，得到视觉特征；
    第二目标特征确定子模块，用于确定目标区域中的文本字符，对所述文本字符进行特征提取，得到字符特征；
    第三目标特征确定子模块，用于根据提取到的视觉特征和字符特征，确定目标区域的目标特征。
  18. 根据权利要求17所述装置，其中，第三目标特征确定子模块，用于将所述视觉特征和字符特征赋予不同的权重；对赋予权重后的所述视觉特征和字符特征进行融合，得到目标区域的目标特征。
  19. 根据权利要求11-18中任一项所述装置，其中，所述装置通过预先构建的分类网络实现，所述装置还包括：
    第一训练模块，用于将样本图像输入所述分类网络中处理，得到样本图像中待提取文本的第一预测类别，以及所述第一预测类别中各个类别之间的对应关系；
    第二训练模块，用于根据所述第一预测类别，以及所述样本图像的标注类别，训练所述分类网络，所述标注类别包括：表征文本属于预设字段的标识的类别，以及表征文本属于预设字段的字段值的类别；
    第三训练模块，用于根据所述对应关系，以及标注的待提取文本之间的对应关系，训练所述分类网络。
  20. 根据权利要求11-19中任一项所述装置，其中，所述图像包括下述至少一种：收据图像、发票图像、名片图像。
  21. 一种电子设备，其中，包括：
    处理器；
    用于存储处理器可执行指令的存储器；
    其中，所述处理器被配置为调用所述存储器存储的指令，以执行权利要求1至10中任意一项所述的方法。
  22. 一种计算机可读存储介质，其上存储有计算机程序指令，其中，所述计算机程序指令被处理器执行时实现权利要求1至10中任意一项所述的方法。
  23. 一种计算机程序，其中，包括计算机可读代码，当所述计算机可读代码在电子设备中运行时，所述电子设备中的处理器执行用于实现权利要求1-10中的任一权利要求所述的方法。
PCT/CN2020/077247 2019-12-27 2020-02-28 图像处理方法及装置、电子设备和存储介质 WO2021128578A1 (zh)

Priority Applications (2)

Application Number Priority Date Filing Date Title
KR1020217020203A KR20210113192A (ko) 2019-12-27 2020-02-28 이미지 처리 방법 및 장치, 전자 기기 및 기억 매체
JP2021538344A JP7097513B2 (ja) 2019-12-27 2020-02-28 画像処理方法及び装置、電子機器並びに記憶媒体

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201911387827.1 2019-12-27
CN201911387827.1A CN111191715A (zh) 2019-12-27 2019-12-27 图像处理方法及装置、电子设备和存储介质

Publications (1)

Publication Number Publication Date
WO2021128578A1 true WO2021128578A1 (zh) 2021-07-01

Family

ID=70707802

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/077247 WO2021128578A1 (zh) 2019-12-27 2020-02-28 图像处理方法及装置、电子设备和存储介质

Country Status (5)

Country Link
JP (1) JP7097513B2 (zh)
KR (1) KR20210113192A (zh)
CN (1) CN111191715A (zh)
TW (1) TWI736230B (zh)
WO (1) WO2021128578A1 (zh)

Cited By (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113506322A (zh) * 2021-07-15 2021-10-15 清华大学 图像处理方法及装置、电子设备和存储介质
CN113688686A (zh) * 2021-07-26 2021-11-23 厦门大学 基于图卷积神经网络的虚拟现实视频质量评价方法
CN113705559A (zh) * 2021-08-31 2021-11-26 平安银行股份有限公司 基于人工智能的文字识别方法及装置、电子设备
CN114283403A (zh) * 2021-12-24 2022-04-05 北京有竹居网络技术有限公司 一种图像检测方法、装置、存储介质及设备
CN114511864A (zh) * 2022-04-19 2022-05-17 腾讯科技（深圳）有限公司 文本信息提取方法、目标模型的获取方法、装置及设备
CN114708961A (zh) * 2022-03-18 2022-07-05 北京理工大学珠海学院 个人生理和心理特性类别测评装置及方法
CN114724133A (zh) * 2022-04-18 2022-07-08 北京百度网讯科技有限公司 文字检测和模型训练方法、装置、设备及存储介质
CN114863245A (zh) * 2022-05-26 2022-08-05 中国平安人寿保险股份有限公司 图像处理模型的训练方法和装置、电子设备及介质
WO2023005468A1 (zh) * 2021-07-30 2023-02-02 上海商汤智能科技有限公司 检测呼吸率的方法、装置、存储介质及电子设备

Families Citing this family (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112801099B (zh) * 2020-06-02 2024-05-24 腾讯科技（深圳）有限公司 一种图像处理方法、装置、终端设备及介质
CN111695517B (zh) * 2020-06-12 2023-08-18 北京百度网讯科技有限公司 图像的表格提取方法、装置、电子设备及存储介质
CN112069877B (zh) * 2020-07-21 2022-05-03 北京大学 一种基于边缘信息和注意力机制的人脸信息识别方法
CN112016438B (zh) * 2020-08-26 2021-08-10 北京嘀嘀无限科技发展有限公司 一种基于图神经网络识别证件的方法及系统
CN112784720A (zh) * 2021-01-13 2021-05-11 浙江诺诺网络科技有限公司 基于银行回单的关键信息提取方法、装置、设备及介质
CN113807369B (zh) * 2021-09-26 2024-09-17 北京市商汤科技开发有限公司 目标重识别方法及装置、电子设备和存储介质
CN114037985A (zh) * 2021-11-04 2022-02-11 北京有竹居网络技术有限公司 信息提取方法、装置、设备、介质及产品
KR102485944B1 (ko) 2021-11-19 2023-01-10 주식회사 스탠다임 트랜스포머 신경망에서의 그래프 인코딩 방법
CN116383428B (zh) * 2023-03-31 2024-04-05 北京百度网讯科技有限公司 一种图文编码器训练方法、图文匹配方法及装置

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20160098608A1 (en) * 2013-09-05 2016-04-07 Ebay Inc. System and method for scene text recognition
CN108921166A (zh) * 2018-06-22 2018-11-30 深源恒际科技有限公司 基于深度神经网络的医疗票据类文本检测识别方法及系统
CN109308476A (zh) * 2018-09-06 2019-02-05 邬国锐 票据信息处理方法、系统及计算机可读存储介质
CN109919014A (zh) * 2019-01-28 2019-06-21 平安科技（深圳）有限公司 Ocr识别方法及其电子设备
CN110033000A (zh) * 2019-03-21 2019-07-19 华中科技大学 一种票据图像的文本检测与识别方法

Family Cites Families (17)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2000132639A (ja) 1998-10-27 2000-05-12 Nippon Telegr & Teleph Corp <Ntt> 文字抽出認識方法及び装置及びこの方法を記録した記録媒体
US7756871B2 (en) * 2004-10-13 2010-07-13 Hewlett-Packard Development Company, L.P. Article extraction
CN101894123A (zh) * 2010-05-11 2010-11-24 清华大学 基于子图的链接相似度的快速近似计算系统和方法
CN105786980B (zh) * 2016-02-14 2019-12-20 广州神马移动信息科技有限公司 对描述同一实体的不同实例进行合并的方法、装置及设备
CN107679153A (zh) * 2017-09-27 2018-02-09 国家电网公司信息通信分公司 一种专利分类方法及装置
JP7068570B2 (ja) 2017-12-11 2022-05-17 富士通株式会社 生成プログラム、情報処理装置及び生成方法
JP6928876B2 (ja) 2017-12-15 2021-09-01 京セラドキュメントソリューションズ株式会社 フォーム種別学習システムおよび画像処理装置
CN109977723B (zh) * 2017-12-22 2021-10-22 苏宁云商集团股份有限公司 大票据图片文字识别方法
CN108549850B (zh) * 2018-03-27 2021-07-16 联想(北京)有限公司 一种图像识别方法及电子设备
JP7063080B2 (ja) 2018-04-20 2022-05-09 富士通株式会社 機械学習プログラム、機械学習方法および機械学習装置
CN109086756B (zh) * 2018-06-15 2021-08-03 众安信息技术服务有限公司 一种基于深度神经网络的文本检测分析方法、装置及设备
CN110619325B (zh) * 2018-06-20 2024-03-08 北京搜狗科技发展有限公司 一种文本识别方法及装置
WO2020113437A1 (zh) * 2018-12-04 2020-06-11 区链通网络有限公司 图结构处理方法、系统、网络设备及存储介质
CN110276396B (zh) * 2019-06-21 2022-12-06 西安电子科技大学 基于物体显著性和跨模态融合特征的图片描述生成方法
CN110598759A (zh) * 2019-08-23 2019-12-20 天津大学 一种基于多模态融合的生成对抗网络的零样本分类方法
CN110569846A (zh) * 2019-09-16 2019-12-13 北京百度网讯科技有限公司 图像文字识别方法、装置、设备及存储介质
CN110610166B (zh) * 2019-09-18 2022-06-07 北京猎户星空科技有限公司 文本区域检测模型训练方法、装置、电子设备和存储介质

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20160098608A1 (en) * 2013-09-05 2016-04-07 Ebay Inc. System and method for scene text recognition
CN108921166A (zh) * 2018-06-22 2018-11-30 深源恒际科技有限公司 基于深度神经网络的医疗票据类文本检测识别方法及系统
CN109308476A (zh) * 2018-09-06 2019-02-05 邬国锐 票据信息处理方法、系统及计算机可读存储介质
CN109919014A (zh) * 2019-01-28 2019-06-21 平安科技（深圳）有限公司 Ocr识别方法及其电子设备
CN110033000A (zh) * 2019-03-21 2019-07-19 华中科技大学 一种票据图像的文本检测与识别方法

Cited By (15)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113506322A (zh) * 2021-07-15 2021-10-15 清华大学 图像处理方法及装置、电子设备和存储介质
CN113506322B (zh) * 2021-07-15 2024-04-12 清华大学 图像处理方法及装置、电子设备和存储介质
CN113688686B (zh) * 2021-07-26 2023-10-27 厦门大学 基于图卷积神经网络的虚拟现实视频质量评价方法
CN113688686A (zh) * 2021-07-26 2021-11-23 厦门大学 基于图卷积神经网络的虚拟现实视频质量评价方法
WO2023005468A1 (zh) * 2021-07-30 2023-02-02 上海商汤智能科技有限公司 检测呼吸率的方法、装置、存储介质及电子设备
CN113705559A (zh) * 2021-08-31 2021-11-26 平安银行股份有限公司 基于人工智能的文字识别方法及装置、电子设备
CN113705559B (zh) * 2021-08-31 2024-05-10 平安银行股份有限公司 基于人工智能的文字识别方法及装置、电子设备
CN114283403A (zh) * 2021-12-24 2022-04-05 北京有竹居网络技术有限公司 一种图像检测方法、装置、存储介质及设备
CN114283403B (zh) * 2021-12-24 2024-01-16 北京有竹居网络技术有限公司 一种图像检测方法、装置、存储介质及设备
CN114708961A (zh) * 2022-03-18 2022-07-05 北京理工大学珠海学院 个人生理和心理特性类别测评装置及方法
CN114724133A (zh) * 2022-04-18 2022-07-08 北京百度网讯科技有限公司 文字检测和模型训练方法、装置、设备及存储介质
CN114724133B (zh) * 2022-04-18 2024-02-02 北京百度网讯科技有限公司 文字检测和模型训练方法、装置、设备及存储介质
WO2023202268A1 (zh) * 2022-04-19 2023-10-26 腾讯科技（深圳）有限公司 文本信息提取方法、目标模型的获取方法、装置及设备
CN114511864A (zh) * 2022-04-19 2022-05-17 腾讯科技（深圳）有限公司 文本信息提取方法、目标模型的获取方法、装置及设备
CN114863245A (zh) * 2022-05-26 2022-08-05 中国平安人寿保险股份有限公司 图像处理模型的训练方法和装置、电子设备及介质

Also Published As

Publication number Publication date
CN111191715A (zh) 2020-05-22
JP2022518889A (ja) 2022-03-17
JP7097513B2 (ja) 2022-07-07
TW202125307A (zh) 2021-07-01
KR20210113192A (ko) 2021-09-15
TWI736230B (zh) 2021-08-11

Similar Documents

Publication Publication Date Title
WO2021128578A1 (zh) 图像处理方法及装置、电子设备和存储介质
TWI728621B (zh) 圖像處理方法及其裝置、電子設備、電腦可讀儲存媒體和電腦程式
WO2021155632A1 (zh) 图像处理方法及装置、电子设备和存储介质
US11120078B2 (en) Method and device for video processing, electronic device, and storage medium
TWI749423B (zh) 圖像處理方法及裝置、電子設備和電腦可讀儲存介質
WO2021051857A1 (zh) 目标对象匹配方法及装置、电子设备和存储介质
WO2021056808A1 (zh) 图像处理方法及装置、电子设备和存储介质
WO2020232977A1 (zh) 神经网络训练方法及装置以及图像处理方法及装置
WO2021208667A1 (zh) 图像处理方法及装置、电子设备和存储介质
WO2021051650A1 (zh) 人脸和人手关联检测方法及装置、电子设备和存储介质
JP6007354B2 (ja) 直線検出方法、装置、プログラム、及び記録媒体
CN110009090B (zh) 神经网络训练与图像处理方法及装置
US11288531B2 (en) Image processing method and apparatus, electronic device, and storage medium
CN110443366B (zh) 神经网络的优化方法及装置、目标检测方法及装置
CN110532956B (zh) 图像处理方法及装置、电子设备和存储介质
CN111259967B (zh) 图像分类及神经网络训练方法、装置、设备及存储介质
WO2021208666A1 (zh) 字符识别方法及装置、电子设备和存储介质
WO2023115911A1 (zh) 对象重识别方法及装置、电子设备、存储介质和计算机程序产品
WO2020147414A1 (zh) 网络优化方法及装置、图像处理方法及装置、存储介质
CN112926310B (zh) 一种关键词提取方法及装置
WO2021082463A1 (zh) 数据处理方法及装置、电子设备和存储介质
WO2022141969A1 (zh) 图像分割方法及装置、电子设备、存储介质和程序
CN105824955A (zh) 短信聚类方法及装置
TW201344577A (zh) 利用圖像辨識導引安裝應用程式的方法及電子裝置
US20220270352A1 (en) Methods, apparatuses, devices, storage media and program products for determining performance parameters

Legal Events

Date Code Title Description
ENP Entry into the national phase

Ref document number: 2021538344

Country of ref document: JP

Kind code of ref document: A

121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 20906061

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205A DATED 31.10.2022)

122 Ep: pct application non-entry in european phase

Ref document number: 20906061

Country of ref document: EP

Kind code of ref document: A1