WO2017148265A1 - 一种单词的分割方法和装置 - Google Patents

一种单词的分割方法和装置 Download PDF

Info

Publication number
WO2017148265A1
WO2017148265A1 PCT/CN2017/073718 CN2017073718W WO2017148265A1 WO 2017148265 A1 WO2017148265 A1 WO 2017148265A1 CN 2017073718 W CN2017073718 W CN 2017073718W WO 2017148265 A1 WO2017148265 A1 WO 2017148265A1
Authority
WO
WIPO (PCT)
Prior art keywords
image
tested
probability value
word
feature vector
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Ceased
Application number
PCT/CN2017/073718
Other languages
English (en)
French (fr)
Inventor
周文猛
程孟力
毛旭东
施兴
褚崴
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Alibaba Group Holding Ltd
Original Assignee
Alibaba Group Holding Ltd
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 Alibaba Group Holding Ltd filed Critical Alibaba Group Holding Ltd
Priority to US16/081,011 priority Critical patent/US10817741B2/en
Publication of WO2017148265A1 publication Critical patent/WO2017148265A1/zh
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; 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
    • G06V30/158—Segmentation of character regions using character size, text spacings or pitch estimation
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; 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 OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; 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 OR CALCULATING; 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 OR CALCULATING; 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
    • G06N3/09—Supervised learning
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N5/00—Computing arrangements using knowledge-based models
    • G06N5/04—Inference or reasoning models
    • G06N5/046—Forward inferencing; Production systems
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/10—Segmentation; Edge detection
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; 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
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; 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 OR CALCULATING; 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

Definitions

  • the present application relates to the field of Internet technologies, and in particular, to a method and an apparatus for segmenting words.
  • OCR Optical Character Recognition
  • the present application provides a method and device for segmenting words to perform word segmentation on an image, and quickly and accurately detect a text region in the image to improve the user experience.
  • the technical solutions are as follows:
  • the application provides a method for segmenting words, the method comprising the following steps:
  • Obtaining a sample image where the sample image includes a word interval mark or a non-word interval mark; processing the sample image by using a convolutional neural network to obtain a first feature vector corresponding to the sample image, and corresponding to the first feature vector Word interval probability value and/or non-word interval probability value;
  • the acquiring the sample image includes: extracting at least two letter images from the font library, and splicing the at least two letter images to obtain a stitched image, and setting a word interval mark or a non-word interval for the stitched image. Marking the stitched image as the sample image; or performing one or any combination of the following operations on the stitched image: a panning operation, a rotation operation, a stretching operation, adding Gaussian noise, and performing the operation The subsequent image is labeled as the sample image.
  • the method further includes converting the sample image into a grayscale sample image and converting the grayscale sample image to a size of M*N before processing the sample image using a convolutional neural network.
  • the image to be tested is converted into the image to be tested of the size of M*L, and the image to be tested of the size M*L is normalized;
  • the M is a positive integer
  • the N is a positive integer
  • the L is a positive integer
  • the ratio of the M to the L is the same as the aspect ratio of the image to be tested.
  • Processing the image to be tested by using the convolutional neural network to obtain a second feature vector corresponding to the image to be tested including:
  • the K+1 different images to be tested are processed by using a convolutional neural network to obtain K+1 different feature vectors corresponding to the images to be tested, and the K+1 different images to be tested are respectively corresponding.
  • the feature vectors are combined to obtain a second feature vector corresponding to the image to be tested.
  • the process of obtaining the word interval probability value or the non-word interval probability value corresponding to the second feature vector includes:
  • performing word segmentation on the image to be tested includes:
  • the position of the image to be tested corresponding to the second feature vector is taken as the abscissa, and the word interval probability value corresponding to the second feature vector is taken as the ordinate to obtain a probability value result map; the word is found from the probability value result map. And spacing the maximum value of the probability value, and using the maximum value position to perform word segmentation on the image to be tested; or
  • the position of the image to be tested corresponding to the second feature vector is taken as the abscissa, and the non-word interval probability value corresponding to the second feature vector is taken as the ordinate, and a probability value result map is obtained; and the probability value result map is found.
  • the application provides a word segmentation device, and the device specifically includes:
  • a training module configured to acquire a sample image, where the sample image includes a word interval mark or a non-word interval mark; processing the sample image by using a convolutional neural network to obtain a first feature vector corresponding to the sample image, a word interval probability value and/or a non-word interval probability value corresponding to the first feature vector;
  • a prediction module configured to acquire an image to be tested, and process the image to be tested by using the convolutional neural network to obtain a second feature vector corresponding to the image to be tested and a word interval corresponding to the second feature vector Probability value or non-word interval probability value; performing word segmentation on the image to be tested by using the currently obtained word interval probability value or non-word interval probability value.
  • the training module is specifically configured to: extract at least two letter images from the font library in the process of acquiring the sample image, and perform splicing on the at least two letter images to obtain a spliced image, and The image sets a word interval mark or a non-word interval mark;
  • Marking the stitched image as the sample image or performing one or any combination of the following operations on the stitched image: a panning operation, a rotation operation, a stretching operation, adding Gaussian noise, and performing the operation
  • the image is labeled as the sample image.
  • the training module is further configured to convert the sample image into a grayscale sample image and convert the grayscale sample image into a size of M*N before processing the sample image by using a convolutional neural network. a sample image, and normalizing the sample image of size M*N;
  • the prediction module is further configured to convert the image to be tested into a grayscale image to be tested before converting the image to be tested by using the convolutional neural network, and convert the grayscale image to be tested into a size of M*L.
  • the image to be tested is normalized to the image to be tested of size M*L;
  • the M is a positive integer
  • the N is a positive integer
  • the L is a positive integer
  • the ratio of the M to the L is the same as the aspect ratio of the image to be tested.
  • the prediction module is specifically configured to: when the image to be tested is processed by using the convolutional neural network to obtain a second feature vector corresponding to the image to be tested, the image to be tested is translated to the left K times or K times to the right, the translation position is different for each translation to obtain K+1 different images to be tested, and the K is a positive integer;
  • K+1 different images to be tested are processed by convolutional neural network, and K+1 different samples to be tested are obtained.
  • the feature vectors corresponding to the images are respectively combined, and the feature vectors corresponding to the K+1 different images to be tested are combined to obtain a second feature vector corresponding to the image to be tested.
  • the prediction module is specifically configured to obtain, by using the convolutional neural network, multiple matching of the second feature vector in a process of obtaining a word interval probability value or a non-word interval probability value corresponding to the second feature vector a first feature vector, and a word interval probability value or a non-word interval probability value corresponding to the plurality of first feature vectors is used as a word interval probability value or a non-word interval probability value corresponding to the second feature vector.
  • the prediction module is specifically configured to: in the process of performing word segmentation on the image to be tested by using a currently obtained word interval probability value or a non-word interval probability value, position the image to be tested corresponding to the second feature vector As the abscissa, a word interval probability value corresponding to the second feature vector is taken as an ordinate, and a probability value result map is obtained; a maximum value position of the word interval probability value is found from the probability value result map, and the a maximum value position is used to perform word segmentation on the image to be tested; or, a position of the image to be tested corresponding to the second feature vector is taken as an abscissa, and a non-word interval probability value corresponding to the second feature vector is used as an ordinate. Obtaining a probability value result map; finding a minimum value position of the non-word interval probability value from the probability value result map, and performing word segmentation on the image to be tested by using the minimum value position.
  • the word segmentation of the image by the convolutional neural network can effectively divide the image including multiple words into several small images, ensuring that each small image contains only one word, and It does not divide a word into multiple small images, so that word segmentation can be accurately performed, the accuracy of word segmentation can be improved, and the speed of word segmentation can be improved, and the text region in the image can be detected quickly and accurately.
  • User experience can effectively divide the image including multiple words into several small images, ensuring that each small image contains only one word, and It does not divide a word into multiple small images, so that word segmentation can be accurately performed, the accuracy of word segmentation can be improved, and the speed of word segmentation can be improved, and the text region in the image can be detected quickly and accurately.
  • 1 and 2 are schematic structural diagrams of a convolutional neural network in an embodiment of the present application
  • FIG. 3 is a flowchart of a method for dividing a word in an embodiment of the present application.
  • FIG. 4 is a schematic diagram of a sample image in an embodiment of the present application.
  • FIG. 5 is a schematic diagram of a probability value result diagram in an embodiment of the present application.
  • FIG. 6 is a hardware structural diagram of an electronic device in an embodiment of the present application.
  • Fig. 7 is a view showing the configuration of a word dividing device in an embodiment of the present application.
  • first, second, third, etc. may be used to describe various information in this application, such information should not be limited to these terms. These terms are only used to distinguish the same type of information from each other.
  • first information may also be referred to as the second information without departing from the scope of the present application.
  • second information may also be referred to as the first information.
  • word “if” may be interpreted to mean “at time” or "when” or "in response to determination.”
  • Step 1. Convert the image to a grayscale image.
  • Step 2. Perform median filtering on the grayscale image.
  • Step 3: Perform binarization processing on the median-filtered image, and mark the connected component, and use each connected area as a preliminary candidate area for characters.
  • Step 4: According to the prior knowledge of the characters in the image, the culling decision condition is set, and the candidate regions that are not the characters are excluded by the feature similarity of the text candidate region, the stroke width and the height of the character.
  • Step 5 Perform morphological closure on the candidate region that is not proposed in step 4, and use the convolutional neural network to determine the text region, and reject the region that is not a text.
  • Step 6. Perform statistics according to the interval of the text area. If the interval between adjacent areas is greater than the manually set distance threshold, the two areas are separated.
  • the binarization process is very poor in reliability for images of complex backgrounds, and it is easy to fail, resulting in subsequent errors.
  • the method introduces a lot of a priori artificial knowledge and rules, for example, using the prior knowledge of the text to set the decision condition, and using the width of the stroke of the text to eliminate the candidate text area, etc., these features change with the scene. It is difficult to adapt to all scenarios through manual rules.
  • the method divides the word by setting the artificial threshold based on the distance between the characters, and it is easy to separate the words with a slightly larger word spacing.
  • the embodiment of the present application proposes a method for segmenting a word.
  • the image containing multiple words can be effectively segmented into several small images to ensure that each small image is only Contains a word, and does not split a word into multiple small images, so as to accurately segment the word, improve the accuracy of word segmentation, and improve the speed of word segmentation, which can quickly and accurately detect the image.
  • the text area enhances the user experience.
  • it is not necessary to perform binarization processing on the image, improve the reliability of the algorithm, and avoid the problem of result error caused by the binarization process.
  • the convolutional neural network By providing a large number of sample images to the convolutional neural network, the interval between words and words can be automatically distinguished. This method is completely driven by data and does not require manual intervention. Moreover, acceleration is performed by a convolutional neural network, and the recognition speed is fast under the premise of ensuring the recognition accuracy.
  • the convolutional neural networks referred to in the embodiments of the present application are all full convolutional neural networks.
  • FIG. 1 which is a schematic diagram of a convolutional neural network
  • the input of the convolutional neural network is a two-dimensional image
  • the output of the convolutional neural network is a word interval probability value and/or a non-word interval probability value, for example, convolution.
  • the output of the neural network may be two neurons, one neuron outputs a word interval probability value indicating that the current image is a probability value of the word interval, and the other neuron outputs a non-word interval probability value indicating that the current image is not a word interval Probability value.
  • the structural unit of the convolutional neural network may include a convolutional layer, a pooled layer, a fully connected layer, and the like.
  • the convolutional layer by convolving the image using a convolution kernel, the image features can be enhanced, and the convolutional layer is convoluted using a convolution kernel in a spatial range, and the convolution kernel can be an m*n size.
  • the matrix, the input of the convolutional layer is convolved with the convolution kernel, and the output of the convolutional layer is obtained.
  • the convolution operation is actually a filtering process.
  • the gray value f(x, y) of the point (x, y) on the image is convolved with the convolution kernel w(x, y). .
  • a 4*4 convolution kernel is provided, and the 4*4 convolution kernel contains 16 values, and the size of the 16 values can be arbitrarily configured according to actual needs. Sliding sequentially on the image according to the size of 4*4, a plurality of 4*4 sliding windows are obtained, and the 4*4 convolution kernel is convoluted with each sliding window to obtain multiple convolution features, and these convolutions are obtained.
  • the feature is the output of the convolutional layer and is provided to the pooling layer.
  • the pooling layer it is actually a downsampling process, which reduces the amount of calculation and maintains the feature by performing operations such as maximizing, minimizing, and averaging multiple convolutional features (output of the convolutional layer). transsexual.
  • the image is subsampled using the principle of image local correlation, which can reduce the amount of data processing and retain useful information.
  • the convolution feature can be pooled using the following formula: And get the characteristics of the pooled processing. Where s represents the window size (s*s) at the time of the corresponding pooling process, m and n are both set values, j and k are convolution features of the convolutional layer output, and i represents the i-th image.
  • the characteristics of the pooled layer output can usually be performed using an activation function such as a nonlinear function. Mapping, thereby introducing nonlinear factors, enables the convolutional neural network to enhance expression capabilities through a combination of nonlinearities.
  • the activation function may include, but is not limited to, a sigmoid (S-type) function, a ReLU (Rectified Linear Units) function, and the like. Taking the ReLU function as an example, the ReLU function can set all the features output by the pooling layer to less than 0, and the features greater than 0 remain unchanged.
  • the fully connected layer is used to perform full join processing on all features input to the fully connected layer, thereby obtaining a feature vector, and the feature vector includes a plurality of features. Further, the fully connected layer may also adopt a 1*1 convolution layer to form a fully convolved network.
  • the convolutional neural network is exemplified by two convolutional layers, two pooled layers, and two fully connected layers.
  • the number of convolutional layers, the number of pooled layers, and The number of fully connected layers can also be other values.
  • the number of convolution layers may be three
  • the number of pooled layers may be two
  • the number of fully connected layers may be two
  • the specific network structure may also be as shown in FIG. 2. Shown.
  • the convolutional neural network is composed of three convolutional layers, two pooled layers, and two fully connected layers, and each of the pooled layers may further include a ReLU.
  • the method for segmenting a word proposed in the embodiment of the present application may include two phases, a first phase is a training phase, and a second phase is a prediction phase.
  • the training phase the convolutional neural network is trained by using a large number of sample images to obtain a model with the ability to recognize word spacing.
  • the prediction phase the currently-obtained image to be tested is predicted by the convolutional neural network obtained by the training, and the test image is used for word segmentation by using the prediction result.
  • the processing flow of the training phase may be as shown in step 301, and the processing flow of the prediction phase may be as shown in step 302.
  • Figure 3 In Figure 3:
  • Step 301 Acquire a sample image, where the sample image includes a word interval mark or a non-word interval mark; processing the sample image by using a convolutional neural network to obtain a first feature vector corresponding to the sample image, and a word corresponding to the first feature vector Interval probability values and/or non-word interval probability values.
  • Step 302 Acquire an image to be tested, and process the image to be tested by using the convolutional neural network to obtain a second feature vector corresponding to the image to be tested, and a word interval probability value or a non-word corresponding to the second feature vector.
  • Interval probability value using the currently obtained word interval probability value or non-word interval probability value to perform word segmentation on the image to be tested.
  • the first feature vector and the second feature vector are only used to distinguish the feature vector corresponding to the sample image from the feature vector corresponding to the image to be tested, and the feature vector corresponding to the sample image is referred to as a first feature vector, and will be The feature vector corresponding to the test image is referred to as a second feature vector.
  • step 301 in the process of acquiring the sample image, at least two letter images may be extracted from the font library, and at least two letter images are spliced to obtain a spliced image, and a word interval mark or non-mark is set for the spliced image. Word interval mark.
  • the stitched image may be marked as a sample image; or the stitched image may be subjected to one or any combination of the following operations: a panning operation, a rotation operation, a stretching operation, adding Gaussian noise, and marking the image after the operation as Sample image.
  • a word interval mark or a non-word interval mark needs to be set to indicate whether each area of the sample image is an inter-word area.
  • the sample image may be ab, and there is no space between a and b, such that a non-word interval mark, such as a mark 0, may be set for the sample image.
  • the sample image may be a b and there is a space between a and b, such that a word interval mark can be set for the sample image.
  • At least two letter images may be extracted from the font library, and the at least two letter images are stitched to obtain a stitched image, and a word interval mark or a non-word interval mark is directly set for the stitched image.
  • a word interval mark or a non-word interval mark is directly set for the stitched image.
  • the letter image a and the letter image b can be extracted from the font library, and the two letter images are stitched into the word ab. Assuming that there is no space between the abs, a non-word interval mark is set for the stitched image.
  • the stitched image can be directly marked as a sample image.
  • the stitching image can be randomly translated, rotated, and stretched. And adding a certain degree of Gaussian noise, so that the image after the above operation is as close as possible to the real image, and the image after the above operation can be marked as a sample image, as shown in FIG.
  • the sample image may be converted into a grayscale sample image, and the grayscale sample image is converted into a sample image of size M*N, and the size is
  • M*N sample image is normalized, and the normalized sample image is input to a convolutional neural network, and the sample image is processed by a convolutional neural network.
  • M is a positive integer and N is a positive integer.
  • both M and N may be 32.
  • the sample image acquired in step 301 may be a sample image having a size of 32*32, or the sample image acquired in step 301 is not 32* in size.
  • the sample image acquired in step 301 needs to be converted into a sample image having a size of 32*32.
  • p(i,j) represents one pixel value of the sample image (sample image of gray scale)
  • i represents the abscissa in the sample image
  • j represents the ordinate in the sample image
  • p(i,j) Is the pixel value at the pixel position (i, j).
  • n represents the number of pixels after the sample image is preprocessed.
  • p'(i,j) is the normalized pixel value.
  • the reason for the normalization process is that the range of input data may be particularly large due to the different units of the input data of the sample image, and the result is slow convergence and long training time. Moreover, the input data with a large data range may play a larger role in the pattern classification, while the input data with a small data range may have a smaller role in the pattern classification. Therefore, the input data may be normalized. Map input data to [-128, +128] intervals, or other intervals.
  • the sample image may be processed by using a convolutional neural network to finally obtain a first feature vector corresponding to the sample image, a word interval probability value corresponding to the first feature vector, and/or a non-word interval probability value.
  • a convolutional neural network shown in FIG. 2
  • the sample image is input to the convolution layer, and the sample image is convoluted, and the convolution processing result is output to the pooling layer, and the convolution processing result is pooled.
  • the processing is performed, and the result of the pooling processing is output to the ReLU function, and the result of the pooling processing is mapped by the ReLU function, and the mapping result is output to the convolution layer, and so on, until the full connection processing is performed to obtain the first feature.
  • the vector and the word interval probability value and/or the non-word interval probability value corresponding to the first feature vector are obtained by the convolutional neural network.
  • the value is fine.
  • the process of processing the sample images using the convolutional neural network can be invoked by calling CAFFE (Convolutional Architecture for Fast Feature).
  • CAFFE Convolutional Architecture for Fast Feature
  • Embedding a convolutional neural network framework, processes a sample image on a GPU (Graphics Processing Unit) cluster, which simultaneously processes multiple sample images to speed up processing.
  • a convolutional neural network can be trained, and the current output image to be tested can be predicted by the convolutional neural network, and the test image is subjected to word segmentation.
  • the image to be tested refers to the image that needs to be segmented, and the manner of how to obtain the image to be tested is not described here.
  • the image to be tested may be converted into a grayscale image to be tested, and the image to be tested of the grayscale is converted into a size of M*L.
  • the image is tested, and the image to be tested of size M*L is normalized, and then the normalized image to be tested is input to the convolutional neural network.
  • M is a positive integer and L is a positive integer
  • the ratio of M to L is the same as the aspect ratio of the image to be tested.
  • normalization processing can also be performed, and the formula can be: p(i,j) represents a pixel value of the image to be tested (the image to be tested in grayscale), i represents the abscissa in the image to be tested, j represents the ordinate in the image to be tested, and p(i,j) is The pixel value at the pixel position (i, j). n represents the number of pixels after the image to be tested is preprocessed. p'(i,j) is the normalized pixel value.
  • the reason for the normalization process is that the range of input data may be particularly large due to the different units of the input data of the image to be tested, which results in slow convergence and long training time. Moreover, the input data with a large data range may play a larger role in the pattern classification, while the input data with a small data range may have a smaller role in the pattern classification. Therefore, the input data may be normalized. Map input data to [-128, +128] intervals, or other intervals.
  • the storage type of the image to be tested may be set to a floating point type, so that the pixel value in the image to be tested may include a position after the decimal point.
  • the pixel value is 1.2
  • the pixel value is a real pixel value of 1.2
  • the pixel value may be an unrealistic pixel value of 1.
  • the test image may be processed by using a convolutional neural network, and finally a second feature vector corresponding to the image to be tested, a word interval probability value corresponding to the second feature vector, or a non-word interval probability value may be obtained.
  • a convolutional neural network shown in FIG. 2
  • the image to be tested is input to the convolutional layer, and the test image is convoluted, and the convolution processing result is output to the pooling layer, and the convolution processing result is performed.
  • the pooled processing result is output to the ReLU function, the pooling processing result is mapped by using the ReLU function, and the mapping result is output to the convolution layer, and so on, until the full connection processing is performed to obtain the second feature vector, the first The word interval probability value or the non-word interval probability value corresponding to the two feature vectors.
  • the convolutional neural network it is not described in the embodiment of the present application, as long as the convolutional neural network is used to obtain the second feature vector, the word interval probability value corresponding to the second feature vector, or the non-word interval probability value. can.
  • the convolutional neural network may also be used to obtain the first feature vector. a plurality of first feature vectors matched by the two feature vectors, and a word interval probability value or a non-word interval probability value corresponding to the plurality of first feature vectors as a word interval probability value or a non-word interval corresponding to the second feature vector Probability value.
  • the word interval probability value corresponding to the first feature vector is used as the word interval probability value corresponding to the second feature vector
  • the non-word interval probability value corresponding to the first feature vector is used as the non-word interval corresponding to the second feature vector. Probability value.
  • the process for obtaining the plurality of first feature vectors matched by the second feature vector is determined by the algorithm of the convolutional neural network itself, and is not described in detail in the embodiment of the present application.
  • step 302 the word interval probability value corresponding to the second feature vector is obtained. If the non-word interval probability value corresponding to the first feature vector is obtained and recorded in step 301, then in step 302, the non-word interval probability value corresponding to the second feature vector is obtained. If the word interval probability value and the non-word interval probability value corresponding to the first feature vector are obtained and recorded in step 301, then in step 302, a word interval probability value or a non-word interval probability value corresponding to the second feature vector is obtained.
  • the word interval probability value corresponding to the second feature vector indicates that the image position to be tested corresponding to the second feature vector (eg, the corresponding position of the plurality of first feature vectors corresponding to the second feature vector in the image to be tested) is a word interval.
  • the probability value, the non-word interval probability value corresponding to the second feature vector indicates that the image position to be tested corresponding to the second feature vector (eg, the corresponding position of the plurality of first feature vectors corresponding to the second feature vector in the image to be tested) is not a word The probability value of the interval.
  • the location of e in the word best, assuming that the word interval probability value corresponding to the second feature vector is 5%, indicating that the probability value of position 1 is a word interval is only 5%, assuming the non-word interval probability value corresponding to the second feature vector At 98%, the probability value that position 1 is not a word interval is 98%.
  • the test image is processed by using a convolutional neural network to obtain a corresponding image to be tested.
  • the image to be tested can be shifted K times to the left or K times to the right, and the translation position is different for each translation, and K+1 different images to be tested are obtained, and K is a positive integer.
  • K+1 different images to be tested are processed by convolutional neural network, and K+1 different feature vectors corresponding to the images to be tested are obtained, and K+1 different feature vectors corresponding to the images to be tested are respectively corresponding. Combine together to obtain a second feature vector corresponding to the image to be tested.
  • the pooling process is actually a downsampling (downsampling) process
  • a large number of features of the image to be tested are simplified to a small amount by performing operations such as maximizing, minimizing, and averaging multiple features.
  • Features such that the output features are much less than the number of features of the image to be tested.
  • the image to be tested has 1000 features, but after the image to be tested is processed by the convolutional neural network, only 50 features are included in the obtained second feature vector. In this way, the resulting second feature vector does not fully reflect all the features of the image to be tested.
  • the image to be tested may be shifted K times to the left or K times to the right to obtain K+1 different images to be tested, and K+1 different in the convolutional neural network.
  • K+1 different feature vectors corresponding to the image to be tested are respectively obtained, and the feature vectors corresponding to the K+1 different images to be tested are combined to obtain corresponding images to be tested.
  • the second feature vector such that the number of features is increased, such that the second feature vector can represent all of the features of the image to be tested. For example, when K is 19, then 50*20 features may be included in the second feature vector.
  • the image to be tested is shifted K times to the left, and K is 5 as an example.
  • K is 5 as an example.
  • a 32*30 image to be tested one pixel is shifted to the left to become a 31*30 image.
  • *30 of the image is added to the rightmost side of the image to obtain a 32*30 image to be tested 2, and a column of data added after the rightmost side can be the same as the 31st column of data in the 31*30 image.
  • two pixels are shifted to the left to become a 30*30 image, and two columns of data are added after the rightmost side of the 30*30 image to obtain a 32*30 waiting.
  • Test image 3 the two columns of data added after the far right side can be the same as the 30th column data in the 30*30 image.
  • 32*30 image to be tested 1 three pixels are shifted to the left, and so on, until five pixels are shifted to the left, so that the image to be tested, the image to be tested 2, the image to be tested 3, and the image to be tested are obtained.
  • Test image 4 image to be tested 5, image to be tested 6, using convolutional neural network to test image 1, image to be tested 2, image to be tested 3, image to be tested 4, image to be tested 5, image to be tested 6 Processing, obtaining the feature vectors corresponding to the six different images to be tested, and combining the feature vectors corresponding to the six different images to be tested to obtain the second feature vector.
  • step 302 in the process of performing word segmentation on the test image by using the currently obtained word interval probability value or non-word interval probability value, the first method, the position of the image to be tested corresponding to the second feature vector is taken as the abscissa, The word interval probability value corresponding to the two feature vectors is taken as the ordinate, and the probability value result map is obtained; the maximum value position of the word interval probability value is found from the probability value result map, and the test image is subjected to word segmentation by using the maximum value position.
  • the position of the image to be tested corresponding to the second feature vector is taken as the abscissa, and the non-word interval probability value corresponding to the second feature vector is taken as the ordinate, and the result graph of the probability value is obtained;
  • the position of the word interval probability value is minimized, and the test image is subjected to word segmentation using the minimum value position.
  • the process of determining the position of the image to be tested corresponding to the second feature vector as the abscissa and the word interval probability value/non-word interval probability value corresponding to the second feature vector as the ordinate is actually matching the second feature vector.
  • the plurality of first feature vectors are used as the abscissa at the position corresponding to the image to be tested, and the word interval probability value/non-word interval probability value corresponding to the plurality of first feature vectors is taken as the ordinate.
  • the processing of the first mode and the second mode are similar, and the processing of the first mode is taken as an example for description.
  • the image position to be tested corresponding to the second feature vector may be used as the abscissa, and the word interval probability value corresponding to the second feature vector may be used as the ordinate.
  • the position of the plurality of first feature vectors matching the second feature vector in the image to be tested is taken as the abscissa, and the word interval probability value corresponding to each first feature vector is taken as the ordinate, which is obtained as shown in FIG. 5 .
  • the probability value result graph In FIG. 5, the probability value indicates whether the current position is a probability of a word interval, and the larger the probability value, the more likely the current position is the word interval. Based on this, the maximum value position of the word interval probability value can be found from the probability value result map, and the test image is subjected to word segmentation using the maximum value position.
  • the non-maximum suppression algorithm can be used to find these maxima positions, that is, the position of the vertical line in FIG. 5.
  • the non-maximum suppression algorithm has been widely used in the field of computer vision, and is no longer used here. Narration. After finding these maximum value positions, the image to be tested can be divided at the maximum value position to obtain a small image of a single word, so that the image to be tested is divided into a plurality of small images, and each small image is only Contains one word and does not split a word into multiple small images.
  • the above method is applied to the text recognition task, if the word sentence needs to be translated, the image to be tested is segmented into a plurality of small images by the above method, and each small image can be recognized by the word, and the number of words can be recognized. Add spaces to stitch together and complete the translation of the word sentences.
  • both the sample image and the image to be tested refer to a single line image, and the single line An image containing multiple words.
  • words may include, but are not limited to, English words, Japanese words, German words, Russian units, French words, minority words, and the like.
  • the word segmentation of the image by the convolutional neural network can effectively divide the image including multiple words into several small images, ensuring that each small image contains only one word, and It does not divide a word into multiple small images, so that word segmentation can be accurately performed, the accuracy of word segmentation can be improved, and the speed of word segmentation can be improved, and the text region in the image can be detected quickly and accurately.
  • User experience can effectively divide the image including multiple words into several small images, ensuring that each small image contains only one word, and It does not divide a word into multiple small images, so that word segmentation can be accurately performed, the accuracy of word segmentation can be improved, and the speed of word segmentation can be improved, and the text region in the image can be detected quickly and accurately.
  • the embodiment of the present application provides a word segmentation device, and the word segmentation device is applied to an electronic device.
  • the dividing device of the word can be implemented by software, or can be realized by hardware or a combination of hardware and software.
  • the software implementation as an example, as a logical device, it is formed by reading the corresponding computer program instructions in the non-volatile memory through the processor of the electronic device in which it is located. From the hardware level, as shown in FIG. 6, the hardware structure diagram of the electronic device where the word segmentation device of the present application is located, in addition to the processor and the non-volatile memory shown in FIG.
  • the electronic device can also Including other hardware, such as a forwarding chip, network interface, memory, etc., which is responsible for processing messages; in terms of hardware structure, the electronic device may also be a distributed device, and may include multiple interface cards for packet processing at the hardware level. Extension.
  • a structural diagram of a segmentation device for a word proposed by the present application the device includes:
  • the training module 11 is configured to acquire a sample image, where the sample image includes a word interval mark or a non-word interval mark; processing the sample image by using a convolutional neural network to obtain a first feature vector corresponding to the sample image, a word interval probability value and/or a non-word interval probability value corresponding to the first feature vector;
  • a prediction module 12 configured to acquire an image to be tested, and process the image to be tested by using the convolutional neural network to obtain a second feature vector corresponding to the image to be tested and a word corresponding to the second feature vector An interval probability value or a non-word interval probability value; performing word segmentation on the image to be tested by using a currently obtained word interval probability value or a non-word interval probability value.
  • the training module 11 is specifically configured to: extract at least two letter images from the font library in the process of acquiring the sample image, and splicing the at least two letter images to obtain a stitched image, and The stitched image sets a word interval mark or a non-word interval mark;
  • Marking the stitched image as the sample image or performing one or any combination of the following operations on the stitched image: a panning operation, a rotation operation, a stretching operation, adding Gaussian noise, and performing the operation
  • the image is labeled as the sample image.
  • the training module 11 is further configured to: before processing the sample image by using a convolutional neural network Converting the image into a grayscale sample image, converting the grayscale sample image into a sample image of size M*N, and normalizing the sample image of size M*N;
  • the prediction module 12 is further configured to convert the image to be tested into a grayscale image to be tested before converting the image to be tested by using the convolutional neural network, and convert the grayscale image to be tested into a size of M*L. Testing the image and normalizing the image to be tested of size M*L; wherein M is a positive integer, the N is a positive integer, the L is a positive integer, and the M and The ratio of the L is the same as the aspect ratio of the image to be tested.
  • the prediction module 12 is specifically configured to process the image to be tested by using the convolutional neural network to obtain a second feature vector corresponding to the image to be tested, and to display the image to be tested to the left. Shift K times or K times to the right, the translation position is different for each translation to obtain K+1 different images to be tested, and the K is a positive integer;
  • the K+1 different images to be tested are processed by using a convolutional neural network to obtain K+1 different feature vectors corresponding to the images to be tested, and the K+1 different images to be tested are respectively corresponding.
  • the feature vectors are combined to obtain a second feature vector corresponding to the image to be tested.
  • the prediction module 12 is specifically configured to obtain, by using the convolutional neural network, the second feature vector matching in the process of obtaining a word interval probability value or a non-word interval probability value corresponding to the second feature vector.
  • a first feature vector, and a word interval probability value or a non-word interval probability value corresponding to the plurality of first feature vectors is used as a word interval probability value or a non-word interval probability value corresponding to the second feature vector.
  • the prediction module 12 is specifically configured to: in the process of performing word segmentation on the image to be tested by using a currently obtained word interval probability value or a non-word interval probability value, the image to be tested corresponding to the second feature vector Position as the abscissa, the word interval probability value corresponding to the second feature vector as the ordinate, obtain a probability value result map; find the maximum value position of the word interval probability value from the probability value result map, and use the The maximum value position is used to perform word segmentation on the image to be tested; or the position of the image to be tested corresponding to the second feature vector is taken as the abscissa, and the non-word interval probability value corresponding to the second feature vector is taken as the vertical Coordinates, obtaining a probability value result map; finding a minimum value position of the non-word interval probability value from the probability value result map, and performing word segmentation on the image to be tested by using the minimum value position.
  • the word segmentation of the image by the convolutional neural network can effectively divide the image including multiple words into several small images, ensuring that each small image contains only one word, and It does not divide a word into multiple small images, so that word segmentation can be accurately performed, the accuracy of word segmentation can be improved, and the speed of word segmentation can be improved, and the text region in the image can be detected quickly and accurately.
  • User experience can effectively divide the image including multiple words into several small images, ensuring that each small image contains only one word, and It does not divide a word into multiple small images, so that word segmentation can be accurately performed, the accuracy of word segmentation can be improved, and the speed of word segmentation can be improved, and the text region in the image can be detected quickly and accurately.
  • the modules of the device of the present application may be integrated into one or may be deployed separately.
  • the above modules can be combined into one module, or can be further split into multiple sub-modules.
  • modules in the apparatus in the embodiments may be distributed in the apparatus of the embodiment according to the description of the embodiments, or the corresponding changes may be located in one or more apparatuses different from the embodiment.
  • the modules of the above embodiments may be combined into one module, or may be further split into multiple sub-modules.
  • the serial numbers of the embodiments of the present application are merely for the description, and do not represent the advantages and disadvantages of the embodiments.

Landscapes

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

Abstract

一种单词的分割方法和装置，该方法包括：获取样本图像，所述样本图像包括单词间隔标记或者非单词间隔标记；利用卷积神经网络对所述样本图像进行处理，得到所述样本图像对应的第一特征向量、所述第一特征向量对应的单词间隔概率值和/或非单词间隔概率值（301）；获取待测试图像，并利用所述卷积神经网络对所述待测试图像进行处理，得到所述待测试图像对应的第二特征向量、所述第二特征向量对应的单词间隔概率值或者非单词间隔概率值；利用当前得到的单词间隔概率值或者非单词间隔概率值，对所述待测试图像进行单词分割（302）。通过该技术方案，精准的进行单词分割，使得单词分割的准确率得到提高，提高单词分割的速度，提高用户使用体验。

Description

一种单词的分割方法和装置
本申请要求2016年02月29日递交的申请号为201610112463.6、发明名称为“一种单词的分割方法和装置”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
本申请涉及互联网技术领域，尤其涉及一种单词的分割方法和装置。
背景技术
OCR(Optical Character Recognition，光学字符识别)是指使用电子设备检查纸上打印的字符，并通过检测暗、亮的模式确定其形状，然后使用字符识别方法将形状翻译成计算机文字的过程；即，针对印刷体字符，采用光学的方式将纸质文档中的文字转换成为黑白点阵的图像文件，并通过识别软件将该图像中的文字转换成文本格式，以供文字处理软件进一步进行编辑加工的技术。因此，基于OCR技术，可以将文本图像转换成数字形式。
尽管OCR技术已经较为成熟，但由于图像背景复杂，图像中的文字字体、大小及颜色等变化繁多，再加上拍摄视角的变化，以及不同光照的影响等诸多因素，使得快速、准确地检测出图像中的文字区域的难度较大。
发明内容
本申请提供一种单词的分割方法和装置，以对图像进行单词分割，并快速、准确地检测出图像中的文字区域，提高用户使用体验。技术方案如下：
本申请提供一种单词的分割方法，所述方法包括以下步骤：
获取样本图像，所述样本图像包括单词间隔标记或者非单词间隔标记；利用卷积神经网络对所述样本图像进行处理，得到所述样本图像对应的第一特征向量、所述第一特征向量对应的单词间隔概率值和/或非单词间隔概率值；
获取待测试图像，并利用所述卷积神经网络对所述待测试图像进行处理，得到所述待测试图像对应的第二特征向量、所述第二特征向量对应的单词间隔概率值或者非单词间隔概率值；利用当前得到的单词间隔概率值或者非单词间隔概率值，对所述待测试图像进行单词分割。
所述获取样本图像，包括：从字体库中抽取至少两个字母图像，并对所述至少两个字母图像进行拼接，得到拼接图像，并为所述拼接图像设定单词间隔标记或者非单词间隔标记；将所述拼接图像标记为所述样本图像；或者，对所述拼接图像进行以下操作之一或者任意组合：平移操作、旋转操作、拉伸操作、添加高斯噪声，并将进行所述操作之后的图像标记为所述样本图像。
所述方法进一步包括：在利用卷积神经网络对所述样本图像进行处理之前，将所述样本图像转换成灰度的样本图像，并将所述灰度的样本图像转换成大小为M*N的样本图像，并对所述大小为M*N的样本图像进行归一化处理；在利用所述卷积神经网络对所述待测试图像进行处理之前，将所述待测试图像转换成灰度的待测试图像，并将所述灰度的待测试图像转换成大小为M*L的待测试图像，并对所述大小为M*L的待测试图像进行归一化处理；
其中，所述M为正整数、所述N为正整数、所述L为正整数，且所述M和所述L的比值，与所述待测试图像的长宽比相同。
利用所述卷积神经网络对所述待测试图像进行处理，得到所述待测试图像对应的第二特征向量，包括：
将所述待测试图像向左平移K次或者向右平移K次，每次平移时的平移位置不同，以得到K+1个不同的待测试图像，所述K为正整数；
利用卷积神经网络对K+1个不同的待测试图像进行处理，得到K+1个不同的待测试图像分别对应的特征向量，并将所述K+1个不同的待测试图像分别对应的特征向量组合在一起，得到所述待测试图像对应的第二特征向量。
所述得到所述第二特征向量对应的单词间隔概率值或者非单词间隔概率值的过程，具体包括：
利用所述卷积神经网络获得所述第二特征向量匹配的多个第一特征向量，并将所述多个第一特征向量对应的单词间隔概率值或者非单词间隔概率值，作为所述第二特征向量对应的单词间隔概率值或者非单词间隔概率值。
利用当前得到的单词间隔概率值或者非单词间隔概率值，对所述待测试图像进行单词分割，包括：
将所述第二特征向量对应的待测试图像位置作为横坐标，将所述第二特征向量对应的单词间隔概率值作为纵坐标，得到概率值结果图；从所述概率值结果图中找到单词间隔概率值的极大值位置，并利用所述极大值位置对所述待测试图像进行单词分割；或者， 将所述第二特征向量对应的待测试图像位置作为横坐标，将所述第二特征向量对应的非单词间隔概率值作为纵坐标，得到概率值结果图；从所述概率值结果图中找到非单词间隔概率值的极小值位置，并利用所述极小值位置对所述待测试图像进行单词分割。
本申请提供一种单词的分割装置，所述装置具体包括：
训练模块，用于获取样本图像，所述样本图像包括单词间隔标记或者非单词间隔标记；利用卷积神经网络对所述样本图像进行处理，得到所述样本图像对应的第一特征向量、所述第一特征向量对应的单词间隔概率值和/或非单词间隔概率值；
预测模块，用于获取待测试图像，并利用所述卷积神经网络对所述待测试图像进行处理，得到所述待测试图像对应的第二特征向量、所述第二特征向量对应的单词间隔概率值或者非单词间隔概率值；利用当前得到的单词间隔概率值或者非单词间隔概率值，对所述待测试图像进行单词分割。
所述训练模块，具体用于在获取所述样本图像的过程中，从字体库中抽取至少两个字母图像，并对所述至少两个字母图像进行拼接，得到拼接图像，并为所述拼接图像设定单词间隔标记或者非单词间隔标记；
将所述拼接图像标记为所述样本图像；或者，对所述拼接图像进行以下操作之一或者任意组合：平移操作、旋转操作、拉伸操作、添加高斯噪声，并将进行所述操作之后的图像标记为所述样本图像。
所述训练模块，还用于在利用卷积神经网络对样本图像进行处理之前，将所述样本图像转换成灰度的样本图像，并将所述灰度的样本图像转换成大小为M*N的样本图像，并对所述大小为M*N的样本图像进行归一化处理；
所述预测模块，还用于在利用卷积神经网络对待测试图像进行处理之前，将待测试图像转换成灰度的待测试图像，并将灰度的待测试图像转换成大小为M*L的待测试图像，并对所述大小为M*L的待测试图像进行归一化处理；
其中，所述M为正整数、所述N为正整数、所述L为正整数，且所述M和所述L的比值，与所述待测试图像的长宽比相同。
所述预测模块，具体用于在利用所述卷积神经网络对所述待测试图像进行处理，得到所述待测试图像对应的第二特征向量的过程中，将所述待测试图像向左平移K次或者向右平移K次，每次平移时的平移位置不同，以得到K+1个不同的待测试图像，所述K为正整数；
利用卷积神经网络对K+1个不同的待测试图像进行处理，得到K+1个不同的待测试 图像分别对应的特征向量，并将所述K+1个不同的待测试图像分别对应的特征向量组合在一起，得到所述待测试图像对应的第二特征向量。
所述预测模块，具体用于在得到所述第二特征向量对应的单词间隔概率值或者非单词间隔概率值的过程中，利用所述卷积神经网络获得所述第二特征向量匹配的多个第一特征向量，并将所述多个第一特征向量对应的单词间隔概率值或者非单词间隔概率值，作为所述第二特征向量对应的单词间隔概率值或者非单词间隔概率值。
所述预测模块，具体用于在利用当前得到的单词间隔概率值或者非单词间隔概率值，对所述待测试图像进行单词分割的过程中，将所述第二特征向量对应的待测试图像位置作为横坐标，将所述第二特征向量对应的单词间隔概率值作为纵坐标，得到概率值结果图；从所述概率值结果图中找到单词间隔概率值的极大值位置，并利用所述极大值位置对所述待测试图像进行单词分割；或者，将所述第二特征向量对应的待测试图像位置作为横坐标，将所述第二特征向量对应的非单词间隔概率值作为纵坐标，得到概率值结果图；从所述概率值结果图中找到非单词间隔概率值的极小值位置，并利用所述极小值位置对所述待测试图像进行单词分割。
基于上述技术方案，本申请实施例中，通过卷积神经网络对图像进行单词分割，可以有效地将包含多个单词的图像分割成若干个小图像，保证每个小图像仅包含一个单词，并且不会将一个单词分割到多个小图像中，从而精准的进行单词分割，使得单词分割的准确率得到提高，并且提高单词分割的速度，可以快速、准确地检测出图像中的文字区域，提高用户使用体验。
附图说明
为了更加清楚地说明本申请实施例或者现有技术中的技术方案，下面将对本申请实施例或者现有技术描述中所需要使用的附图作简单地介绍，显而易见地，下面描述中的附图仅仅是本申请中记载的一些实施例，对于本领域普通技术人员来讲，还可以根据这些附图获得其他的附图。
图1和图2是本申请一种实施方式中的卷积神经网络的结构示意图；
图3是本申请一种实施方式中的单词的分割方法的流程图；
图4是本申请一种实施方式中的样本图像的示意图；
图5是本申请一种实施方式中的概率值结果图的示意图；
图6是本申请一种实施方式中的电子设备的硬件结构图；
图7是本申请一种实施方式中的单词的分割装置的结构图。
具体实施方式
在本申请使用的术语仅仅是出于描述特定实施例的目的，而非限制本申请。本申请和权利要求书中所使用的单数形式的“一种”、“所述”和“该”也旨在包括多数形式，除非上下文清楚地表示其它含义。还应当理解，本文中使用的术语“和/或”是指包含一个或多个相关联的列出项目的任何或所有可能组合。
应当理解，尽管在本申请可能采用术语第一、第二、第三等来描述各种信息，但这些信息不应限于这些术语。这些术语仅用来将同一类型的信息彼此区分开。例如，在不脱离本申请范围的情况下，第一信息也可以被称为第二信息，类似地，第二信息也可以被称为第一信息。取决于语境，此外，所使用的词语“如果”可以被解释成为“在……时”或“当……时”或“响应于确定”。
为了从图像中检测出文字区域，如从图像的单行文字中分割出英文单词，在申请人已知的一种技术中，可以采用如下步骤：步骤1、将图像转换为灰度图像。步骤2、对灰度图像进行中值滤波。步骤3、对中值滤波后的图像进行二值化处理，并使用连通分量进行标记，将各联通区作为文字初步候选区域。步骤4、根据图像中文字的先验知识设置剔除判决条件，以文字候选区域的文字相似度、文字的笔画宽度、高度等特征，剔除不是文字的候选区域。步骤5、对步骤4中未被提出的候选区域进行形态学闭处理，并使用卷积神经网络判断该文字区域，剔除不是文字的区域。步骤6、根据文字区域的间隔进行统计，若相邻区域间隔大于人工设定的距离阈值，则将两个区域分开。
在上述方式中，二值化过程对于复杂背景的图像来说，可靠性非常差，很容易失败，导致后续结果错误。而且，该方法引入了很多先验的人工知识和规则，例如，利用文字的先验知识设置判决条件，利用文字笔画的宽度高度来剔除候选文字区域等，这些特征都随着场景的变化而变化，通过人工规则的方式很难适应所有场景。而且，该方法对单词的分割，是基于字符间距离大小设定人工阈值来判断，对于字符间距稍大单词，就容易把其都分割开。
针对上述发现，本申请实施例提出一种单词的分割方法，通过卷积神经网络对图像进行单词分割，可有效地将包含多个单词的图像分割成若干个小图像，保证每个小图像仅包含一个单词，且不会将一个单词分割到多个小图像中，从而精准的进行单词分割，单词分割的准确率得到提高，并提高单词分割的速度，可以快速、准确地检测出图像中 的文字区域，提高用户使用体验。与上述已知技术相比，不需要对图像进行二值化处理，提高算法的可靠性，避免二值化过程导致的结果错误问题。此外，不需要引入很多先验的人工知识和规则，通过给卷积神经网络提供大量的样本图像，可以自动辨别单词与单词之间的间隔，这种方式完全由数据驱动，不需要人工干预。而且，通过卷积神经网络进行加速，在保证识别准确率的前提下，识别速度很快。
其中，本申请实施例中所指的卷积神经网络均是全卷积神经网络。
本申请实施例中，为了完成单词的分割，需要设计有效的卷积神经网络，并使用该卷积神经网络进行单词分割。如图1所示，为卷积神经网络的结构示意图，卷积神经网络的输入是二维图像，卷积神经网络的输出是单词间隔概率值和/或非单词间隔概率值，例如，卷积神经网络的输出可以是两个神经元，一个神经元输出单词间隔概率值，其表示当前图像是单词间隔的概率值，另一个神经元输出非单词间隔概率值，其表示当前图像不是单词间隔的概率值。此外，卷积神经网络的结构单元可以包括卷积层，池化层，全连接层等。
在卷积层中，通过使用卷积核对图像进行卷积运算，可以使图像特征增强，卷积层在空间范围内使用卷积核进行卷积运算，该卷积核可以是一个m*n大小的矩阵，卷积层的输入与卷积核进行卷积，即可得到卷积层的输出。卷积运算实际上是一个滤波的过程，在卷积运算中，是将图像上点(x，y)的灰度值f(x，y)与卷积核w(x，y)进行卷积。例如，提供一个4*4的卷积核，该4*4的卷积核中包含16个数值，这16个数值的大小可以根据实际需要任意配置。按照4*4的大小在图像上依次滑动，得到多个4*4的滑动窗口，将该4*4的卷积核与每个滑动窗口进行卷积，得到多个卷积特征，这些卷积特征，就是卷积层的输出，且被提供给池化层。
在池化层中，实际上是一个降采样过程，通过对多个卷积特征(卷积层的输出)进行取最大、取最小、取平均值等操作，从而减少计算量，并保持特征不变性。在池化层中，利用图像局部相关性的原理，对图像进行子抽样，从而可以减少数据处理量，并保留有用信息。在一个例子中，可以利用如下公式对卷积特征进行池化处理：
Figure PCTCN2017073718-appb-000001
并得到池化处理后的特征。其中，s表示对应的池化处理时的窗口大小(s*s)，m和n均为设定的数值，j和k为卷积层输出的卷积特征，i表示针对第i个图像。
在池化层之后，通常可以使用激活函数(如非线性函数)对池化层输出的特征进行 映射，从而引入非线性因素，使得卷积神经网络通过非线性的组合而增强表达能力。其中，激活函数可以包括但不限于sigmoid(S型)函数、ReLU(Rectified Linear Units，整流线性单元)函数等。以ReLU函数为例进行说明，则该ReLU函数可以将池化层输出的所有特征中，小于0的特征置0，而大于0的特征保持不变。
在全连接层中，全连接层用于将输入给本全连接层的所有特征进行全连接处理，从而得到一个特征向量，且该特征向量中会包括多个特征。进一步的，全连接层还可以采用1*1的卷积层，以构成全卷积的网络。
在图1中，卷积神经网络是以两个卷积层、两个池化层和两个全连接层为例，当然，在实际应用中，卷积层的数量、池化层的数量和全连接层的数量还可以为其它数值。例如，在一种优选的实施方式中，卷积层的数量可以为三个，池化层的数量可以为两个，全连接层的数量可以为两个，具体的网络结构还可以如图2所示。在图2中，卷积神经网络由三个卷积层、两个池化层以及两个全连接层构成，且每个池化层之后还可以包括一个ReLU。
基于上述卷积神经网络，本申请实施例中提出的单词的分割方法，可以包括两个阶段，第一阶段为训练阶段，第二阶段为预测阶段。在训练阶段中，通过使用大量的样本图像，对卷积神经网络进行训练，从而得到一个有识别单词间隔能力的模型。在预测阶段中，通过训练得到的卷积神经网络，对当前输出的待测试图像进行预测，并利用预测结果对待测试图像进行单词分割。
参见图3所示的单词的分割方法的流程图，训练阶段的处理流程可以如步骤301所示，预测阶段的处理流程可以步骤302所示。在图3中：
步骤301，获取样本图像，该样本图像包括单词间隔标记或者非单词间隔标记；利用卷积神经网络对样本图像进行处理，得到该样本图像对应的第一特征向量、该第一特征向量对应的单词间隔概率值和/或非单词间隔概率值。
步骤302，获取待测试图像，并利用该卷积神经网络对该待测试图像进行处理，得到该待测试图像对应的第二特征向量、以及该第二特征向量对应的单词间隔概率值或者非单词间隔概率值；利用当前得到的单词间隔概率值或者非单词间隔概率值，对该待测试图像进行单词分割。
本申请实施例中，第一特征向量和第二特征向量只是为了区分样本图像对应的特征向量和待测试图像对应的特征向量，将样本图像对应的特征向量称为第一特征向量，并将待测试图像对应的特征向量称为第二特征向量。
针对步骤301，在获取样本图像的过程中，可以从字体库中抽取至少两个字母图像，并对至少两个字母图像进行拼接，得到拼接图像，并为该拼接图像设定单词间隔标记或者非单词间隔标记。之后，可以将该拼接图像标记为样本图像；或者，对该拼接图像进行以下操作之一或者任意组合：平移操作、旋转操作、拉伸操作、添加高斯噪声，并将进行操作之后的图像标记为样本图像。
其中，在设计好卷积神经网络之后，为了对卷积神经网络进行训练，需要获取大量的样本图像，每个样本图像的处理过程均相同，即步骤301，因此，以一个样本图像的处理为例进行说明。在样本图像中，需要设定单词间隔标记或者非单词间隔标记，用以表示样本图像的每个区域是否是单词间区域。
例如，针对由单词a和单词b组成的样本图像，该样本图像可以为ab，且a和b之间没有空格，这样，可以为该样本图像设定非单词间隔标记，如标记0。或者，针对由单词a和单词b组成的样本图像，该样本图像可以为a b，且a和b之间有空格，这样，可以为该样本图像设定单词间隔标记。
由于对卷积神经网络进行训练时，需要获取大量的样本图像，如需要上百万的样本图像，针对如此多的样本图像，无法通过人工方式，在每个样本图像中设定单词间隔标记或者非单词间隔标记。因此，本申请实施例中，可以从字体库中抽取至少两个字母图像，并对该至少两个字母图像进行拼接，得到拼接图像，并直接为拼接图像设定单词间隔标记或者非单词间隔标记，从而不需要通过人工方式在每个样本图像中设定单词间隔标记或者非单词间隔标记。例如，可以从字体库中抽取字母图像a和字母图像b，并将这两个字母图像拼接成单词ab，假设ab之间没有空格，则为该拼接成的图像设定非单词间隔标记。
进一步的，可以直接将拼接图像标记为样本图像。或者，考虑到字体库中的图像质量很高，没有任何形变和噪声干扰，与实际采集到的真实图像有一定差距，因此，还可以对拼接图像进行随机的平移操作、旋转操作、拉伸操作，并添加一定程度的高斯噪声，从而使得上述操作之后的图像尽可能地接近真实图像，并可以将进行上述操作之后的图像标记为样本图像，如图4所示。
针对步骤301，在利用卷积神经网络对样本图像进行处理之前，还可以将样本图像转换成灰度的样本图像，并将灰度的样本图像转换成大小为M*N的样本图像，对大小为M*N的样本图像进行归一化处理，并将归一化处理后的样本图像输入到卷积神经网络，利用卷积神经网络对样本图像进行处理。
其中，M为正整数、且N为正整数。在一个例子中，M和N可以均为32，基于此，步骤301中获取的样本图像可以为一个大小为32*32的样本图像，或者，当步骤301中获取的样本图像不是大小为32*32的样本图像时，则需要将步骤301中获取的样本图像转换成大小为32*32的样本图像。
针对大小为32*32的样本图像，还可以进行归一化处理，其公式可以为：
Figure PCTCN2017073718-appb-000002
在上述公式中，p(i,j)表示样本图像(灰度的样本图像)的一个像素值，i表示样本图像中的横坐标，j表示样本图像中的纵坐标，p(i,j)就是在像素位置(i,j)的像素值。n表示样本图像预处理后的像素个数。p’(i,j)为归一化后的像素值。
其中，进行归一化处理的原因是：由于样本图像的输入数据的单位不同，有些输入数据的范围可能特别大，其导致的结果是收敛慢、训练时间长。而且，数据范围大的输入数据在模式分类中的作用可能会偏大，而数据范围小的输入数据在模式分类中的作用可能会偏小，因此，可以通过对输入数据进行归一化处理，将输入数据映射到[-128,+128]区间、或其它区间。
针对步骤301，可以利用卷积神经网络对样本图像进行处理，最终得到该样本图像对应的第一特征向量、该第一特征向量对应的单词间隔概率值和/或非单词间隔概率值。例如，基于图2所示的卷积神经网络，将样本图像输入到卷积层，并对样本图像进行卷积处理，并将卷积处理结果输出给池化层，对卷积处理结果进行池化处理，并将池化处理结果输出给ReLU函数，利用ReLU函数对池化处理结果进行映射，并将映射结果输出给卷积层，以此类推，一直到进行全连接处理，得到第一特征向量以及该第一特征向量对应的单词间隔概率值和/或非单词间隔概率值。针对卷积神经网络内的处理过程，本申请实施例中不再赘述，只要通过卷积神经网络，得到第一特征向量以及该第一特征向量对应的单词间隔概率值和/或非单词间隔概率值即可。
由于卷积神经网络会对大量的样本图像进行处理，如对上百万的样本图像进行处理，因此，利用卷积神经网络对样本图像进行处理的过程，可以通过调用CAFFE(Convolutional Architecture for Fast Feature Embedding，卷积神经网络框架)，在GPU(Graphics Processing Unit，图形处理器)集群上对样本图像进行处理，即同时对多个样本图像进行处理，从而加快处理速度。
基于上述过程，即可以训练得到卷积神经网络，通过该卷积神经网络可以对当前输出的待测试图像进行预测，并对待测试图像进行单词分割。
针对步骤302，在获取待测试图像的过程中，待测试图像是指需要进行单词分割的图像，对于如何获取待测试图像的方式，在此不再赘述。
针对步骤302，在利用卷积神经网络对待测试图像进行处理之前，还可以将该待测试图像转换成灰度的待测试图像，并将灰度的待测试图像转换成大小为M*L的待测试图像，并对大小为M*L的待测试图像进行归一化处理，之后，将归一化处理后的待测试图像输入到卷积神经网络。
其中，M为正整数、L为正整数，M和L的比值与待测试图像的长宽比相同。在一个例子中，M可以为32。假设步骤302中获取的待测试图像的长为64，宽为60，则将待测试图像转换成大小为32*30的待测试图像，即L＝30。
针对大小为M*L的待测试图像，还可以进行归一化处理，其公式可以为：：
Figure PCTCN2017073718-appb-000003
p(i,j)表示待测试图像(灰度的待测试图像)的一个像素值，i表示待测试图像中的横坐标，j表示待测试图像中的纵坐标，p(i,j)就是在像素位置(i,j)的像素值。n表示待测试图像预处理后的像素个数。p’(i,j)为归一化后的像素值。
其中，进行归一化处理的原因是：由于待测试图像的输入数据的单位不同，有些输入数据的范围可能特别大，其导致的结果是收敛慢、训练时间长。而且，数据范围大的输入数据在模式分类中的作用可能会偏大，而数据范围小的输入数据在模式分类中的作用可能会偏小，因此，可以通过对输入数据进行归一化处理，将输入数据映射到[-128,+128]区间、或其它区间。
在实际应用中，针对获取的待测试图像，可以将待测试图像的存储类型设置为浮点型，使得待测试图像中的像素值可以包括小数点后的位置。例如，当像素值为1.2时，使用浮点型存储待测试图像时，像素值是真实的像素值1.2，而使用其它类型存储待测试图像时，像素值可能是不真实的像素值1。
针对步骤302，可以利用卷积神经网络对待测试图像进行处理，最终得到该待测试图像对应的第二特征向量、该第二特征向量对应的单词间隔概率值或者非单词间隔概率值。例如，基于图2所示的卷积神经网络，将待测试图像输入到卷积层，并对待测试图像进行卷积处理，并将卷积处理结果输出给池化层，对卷积处理结果进行池化处理，并 将池化处理结果输出给ReLU函数，利用ReLU函数对池化处理结果进行映射，并将映射结果输出给卷积层，以此类推，一直到进行全连接处理，得到第二特征向量、该第二特征向量对应的单词间隔概率值或者非单词间隔概率值。针对卷积神经网络内的处理过程，本申请实施例中不再赘述，只要通过卷积神经网络，得到第二特征向量、该第二特征向量对应的单词间隔概率值或者非单词间隔概率值即可。
本申请实施例中，针对得到第二特征向量对应的单词间隔概率值或者非单词间隔概率值的过程，在利用卷积神经网络得到第二特征向量后，还可以利用卷积神经网络获得该第二特征向量匹配的多个第一特征向量，并将这多个第一特征向量对应的单词间隔概率值或者非单词间隔概率值，作为该第二特征向量对应的单词间隔概率值或者非单词间隔概率值。具体的，将第一特征向量对应的单词间隔概率值作为该第二特征向量对应的单词间隔概率值，将第一特征向量对应的非单词间隔概率值作为该第二特征向量对应的非单词间隔概率值。
其中，针对获得该第二特征向量匹配的多个第一特征向量的过程，由卷积神经网络自身的算法决定，本申请实施例中不再详加赘述。
其中，如果在步骤301中得到并记录第一特征向量对应的单词间隔概率值，则在步骤302中，得到第二特征向量对应的单词间隔概率值。如果在步骤301中得到并记录第一特征向量对应的非单词间隔概率值，则在步骤302中，得到第二特征向量对应的非单词间隔概率值。如果在步骤301中得到并记录第一特征向量对应的单词间隔概率值和非单词间隔概率值，则在步骤302中，得到第二特征向量对应的单词间隔概率值或者非单词间隔概率值。
其中，第二特征向量对应的单词间隔概率值表示第二特征向量对应的待测试图像位置(如第二特征向量对应的多个第一特征向量在待测试图像中对应的位置)是单词间隔的概率值，第二特征向量对应的非单词间隔概率值表示第二特征向量对应的待测试图像位置(如第二特征向量对应的多个第一特征向量在待测试图像中对应的位置)不是单词间隔的概率值。例如，第二特征向量对应于待测试图像上的位置1(水平方向x＝16的垂直位置，该位置表示一条竖线区域，此处仅以一个位置为例)，位置1是待测试图像上的单词best中的e所在位置，假设第二特征向量对应的单词间隔概率值为5％，表示位置1是单词间隔的概率值仅为5％，假设第二特征向量对应的非单词间隔概率值为98％，表示位置1不是单词间隔的概率值为98％。
针对步骤302，在利用卷积神经网络对待测试图像进行处理，得到待测试图像对应 的第二特征向量的过程中，可以将待测试图像向左平移K次或者向右平移K次，每次平移时的平移位置不同，得到K+1个不同的待测试图像，K为正整数。利用卷积神经网络对K+1个不同的待测试图像进行处理，得到K+1个不同的待测试图像分别对应的特征向量，并将K+1个不同的待测试图像分别对应的特征向量组合在一起，得到待测试图像对应的第二特征向量。
在实际应用中，由于池化处理过程实际上是一个降采样(下采样)过程，通过对多个特征进行取最大、取最小、取平均值等操作，将待测试图像的大量特征简化为少量特征，使得输出特征比待测试图像的特征数量少很多。例如，待测试图像有1000个特征，但是经过卷积神经网络对该待测试图像进行处理后，得到的第二特征向量中只包括50个特征。这样，最终得到的第二特征向量无法完全体现出待测试图像的所有特征。基于此，本申请实施例中，可以将待测试图像向左平移K次或者向右平移K次，得到K+1个不同的待测试图像，在利用卷积神经网络对K+1个不同的待测试图像进行处理后，可以得到K+1个不同的待测试图像分别对应的特征向量，并将K+1个不同的待测试图像分别对应的特征向量组合在一起，得到待测试图像对应的第二特征向量，这样增加了特征数量，使得第二特征向量可以体现出待测试图像的所有特征。例如，当K为19时，则第二特征向量中可以包括50*20个特征。
其中，以将待测试图像向左平移K次，K为5为例进行说明，针对一个32*30的待测试图像1，向左平移一个像素，变成一个31*30的图像，在该31*30的图像最右侧之后添加一列数据，得到一个32*30的待测试图像2，在最右侧之后添加的一列数据可以与31*30的图像中的第31列数据相同。针对该32*30的待测试图像1，向左平移两个像素，变成一个30*30的图像，在该30*30的图像最右侧之后添加两列数据，得到一个32*30的待测试图像3，在最右侧之后添加的两列数据可以与30*30的图像中的第30列数据相同。针对该32*30的待测试图像1，向左平移三个像素，以此类推，一直到向左平移五个像素，这样可以得到待测试图像1、待测试图像2、待测试图像3、待测试图像4、待测试图像5、待测试图像6，利用卷积神经网络对待测试图像1、待测试图像2、待测试图像3、待测试图像4、待测试图像5、待测试图像6分别进行处理，得到6个不同的待测试图像分别对应的特征向量，并将6个不同的待测试图像分别对应的特征向量组合在一起，得到第二特征向量。
由于利用卷积神经网络对待测试图像1、待测试图像2、待测试图像3、待测试图像4、待测试图像5、待测试图像6分别进行处理的过程，共享了大量的卷积操作，从而使 得计算速度得到大大提高，加速了预测过程。
针对步骤302，在利用当前得到的单词间隔概率值或者非单词间隔概率值，对待测试图像进行单词分割的过程中，方式一、将第二特征向量对应的待测试图像位置作为横坐标，将第二特征向量对应的单词间隔概率值作为纵坐标，得到概率值结果图；从概率值结果图中找到单词间隔概率值的极大值位置，并利用极大值位置对待测试图像进行单词分割。或者，方式二、将第二特征向量对应的待测试图像位置作为横坐标，将第二特征向量对应的非单词间隔概率值作为纵坐标，得到概率值结果图；从概率值结果图中找到非单词间隔概率值的极小值位置，并利用极小值位置对待测试图像进行单词分割。
其中，将第二特征向量对应的待测试图像位置作为横坐标，将第二特征向量对应的单词间隔概率值/非单词间隔概率值作为纵坐标的过程，实际上是将第二特征向量匹配的多个第一特征向量在待测试图像对应的位置作为横坐标，并将这多个第一特征向量对应的单词间隔概率值/非单词间隔概率值作为纵坐标。
上述方式一和方式二的处理类似，后续以方式一的处理为例进行说明。
针对方式一，在得到第二特征向量对应的单词间隔概率值之后，可以将第二特征向量对应的待测试图像位置作为横坐标，将第二特征向量对应的单词间隔概率值作为纵坐标，具体的，将第二特征向量匹配的多个第一特征向量在待测试图像中的位置作为横坐标，并将每个第一特征向量对应的单词间隔概率值作为纵坐标，得到如图5所示的概率值结果图。在图5中，概率值表示当前位置是否是单词间隔的概率，概率值越大，则表明当前位置越有可能是单词间隔。基于此，可以从概率值结果图中找到单词间隔概率值的极大值位置，并利用该极大值位置对待测试图像进行单词分割。
在实际应用中，可以使用非极大值抑制算法找到这些极大值位置，即图5中竖线的位置，该非极大值抑制算法在计算机视觉领域中已经被广泛应用，在此不再赘述。在找到这些极大值位置后，可以在极大值位置处将待测试图像分割开，即可得到单个单词的小图像，这样，待测试图像被分割成多个小图像，每个小图像仅包含一个单词，且不会将一个单词分割到多个小图像。
当上述方法应用在文本识别任务时，如果需要对单词句子进行翻译，则采用上述方法将待测试图像分割成多个小图像后，可以对每个小图像进行单词识别，将若干个单词识别结果加上空格拼接在一起，完成单词句子的翻译，
本申请实施例的上述过程中，样本图像和待测试图像均是指单行图像，且该单行中 包含多个单词的图像。此外，单词可以包括但不限于：英文单词、日文单词、德文单词、俄文单位、法文单词、少数民族单词等。
基于上述技术方案，本申请实施例中，通过卷积神经网络对图像进行单词分割，可以有效地将包含多个单词的图像分割成若干个小图像，保证每个小图像仅包含一个单词，并且不会将一个单词分割到多个小图像中，从而精准的进行单词分割，使得单词分割的准确率得到提高，并且提高单词分割的速度，可以快速、准确地检测出图像中的文字区域，提高用户使用体验。
基于与上述方法同样的申请构思，本申请实施例提供一种单词的分割装置，该单词的分割装置应用在电子设备上。该单词的分割装置可通过软件实现，也可通过硬件或者软硬件结合的方式实现。以软件实现为例，作为一个逻辑意义上的装置，是通过其所在的电子设备的处理器，读取非易失性存储器中对应的计算机程序指令形成的。从硬件层面而言，如图6所示，为本申请提出的单词的分割装置所在的电子设备的硬件结构图，除了图6所示的处理器、非易失性存储器外，电子设备还可以包括其他硬件，如负责处理报文的转发芯片、网络接口、内存等；从硬件结构上来讲，该电子设备还可能是分布式设备，可能包括多个接口卡，以便在硬件层面进行报文处理的扩展。
如图7所示，为本申请提出的单词的分割装置的结构图，所述装置包括：
训练模块11，用于获取样本图像，所述样本图像包括单词间隔标记或者非单词间隔标记；利用卷积神经网络对所述样本图像进行处理，得到所述样本图像对应的第一特征向量、所述第一特征向量对应的单词间隔概率值和/或非单词间隔概率值；
预测模块12，用于获取待测试图像，并利用所述卷积神经网络对所述待测试图像进行处理，得到所述待测试图像对应的第二特征向量、所述第二特征向量对应的单词间隔概率值或者非单词间隔概率值；利用当前得到的单词间隔概率值或者非单词间隔概率值，对所述待测试图像进行单词分割。
所述训练模块11，具体用于在获取所述样本图像的过程中，从字体库中抽取至少两个字母图像，并对所述至少两个字母图像进行拼接，得到拼接图像，并为所述拼接图像设定单词间隔标记或者非单词间隔标记；
将所述拼接图像标记为所述样本图像；或者，对所述拼接图像进行以下操作之一或者任意组合：平移操作、旋转操作、拉伸操作、添加高斯噪声，并将进行所述操作之后的图像标记为所述样本图像。
所述训练模块11，还用于在利用卷积神经网络对样本图像进行处理之前，将所述样 本图像转换成灰度的样本图像，并将所述灰度的样本图像转换成大小为M*N的样本图像，并对所述大小为M*N的样本图像进行归一化处理；所述预测模块12，还用于在利用卷积神经网络对待测试图像进行处理之前，将待测试图像转换成灰度的待测试图像，并将灰度的待测试图像转换成大小为M*L的待测试图像，并对所述大小为M*L的待测试图像进行归一化处理；其中，所述M为正整数、所述N为正整数、所述L为正整数，且所述M和所述L的比值，与所述待测试图像的长宽比相同。
所述预测模块12，具体用于在利用所述卷积神经网络对所述待测试图像进行处理，得到所述待测试图像对应的第二特征向量的过程中，将所述待测试图像向左平移K次或者向右平移K次，每次平移时的平移位置不同，以得到K+1个不同的待测试图像，所述K为正整数；
利用卷积神经网络对K+1个不同的待测试图像进行处理，得到K+1个不同的待测试图像分别对应的特征向量，并将所述K+1个不同的待测试图像分别对应的特征向量组合在一起，得到所述待测试图像对应的第二特征向量。
所述预测模块12，具体用于在得到所述第二特征向量对应的单词间隔概率值或者非单词间隔概率值的过程中，利用所述卷积神经网络获得所述第二特征向量匹配的多个第一特征向量，并将所述多个第一特征向量对应的单词间隔概率值或者非单词间隔概率值，作为所述第二特征向量对应的单词间隔概率值或者非单词间隔概率值。
所述预测模块12，具体用于在利用当前得到的单词间隔概率值或者非单词间隔概率值，对所述待测试图像进行单词分割的过程中，将所述第二特征向量对应的待测试图像位置作为横坐标，将所述第二特征向量对应的单词间隔概率值作为纵坐标，得到概率值结果图；从所述概率值结果图中找到单词间隔概率值的极大值位置，并利用所述极大值位置对所述待测试图像进行单词分割；或者，将所述第二特征向量对应的待测试图像位置作为横坐标，将所述第二特征向量对应的非单词间隔概率值作为纵坐标，得到概率值结果图；从所述概率值结果图中找到非单词间隔概率值的极小值位置，并利用所述极小值位置对所述待测试图像进行单词分割。
基于上述技术方案，本申请实施例中，通过卷积神经网络对图像进行单词分割，可以有效地将包含多个单词的图像分割成若干个小图像，保证每个小图像仅包含一个单词，并且不会将一个单词分割到多个小图像中，从而精准的进行单词分割，使得单词分割的准确率得到提高，并且提高单词分割的速度，可以快速、准确地检测出图像中的文字区域，提高用户使用体验。
其中，本申请装置的各个模块可以集成于一体，也可以分离部署。上述模块可以合并为一个模块，也可以进一步拆分成多个子模块。
通过以上的实施方式的描述，本领域的技术人员可以清楚地了解到本申请可借助软件加必需的通用硬件平台的方式来实现，当然也可以通过硬件，但很多情况下前者是更佳的实施方式。基于这样的理解，本申请的技术方案本质上或者说对现有技术做出贡献的部分可以以软件产品的形式体现出来，该计算机软件产品存储在一个存储介质中，包括若干指令用以使得一台计算机设备(可以是个人计算机，服务器，或者网络设备等)执行本申请各个实施例所述的方法。本领域技术人员可以理解附图只是一个优选实施例的示意图，附图中的模块或流程并不一定是实施本申请所必须的。
本领域技术人员可以理解实施例中的装置中的模块可以按照实施例描述进行分布于实施例的装置中，也可以进行相应变化位于不同于本实施例的一个或多个装置中。上述实施例的模块可以合并为一个模块，也可进一步拆分成多个子模块。上述本申请实施例序号仅仅为了描述，不代表实施例的优劣。
以上公开的仅为本申请的几个具体实施例，但是，本申请并非局限于此，任何本领域的技术人员能思之的变化都应落入本申请的保护范围。

Claims (12)

  1. 一种单词的分割方法，其特征在于，所述方法包括以下步骤：
    获取样本图像，所述样本图像包括单词间隔标记或者非单词间隔标记；利用卷积神经网络对所述样本图像进行处理，得到所述样本图像对应的第一特征向量、所述第一特征向量对应的单词间隔概率值和/或非单词间隔概率值；
    获取待测试图像，并利用所述卷积神经网络对所述待测试图像进行处理，得到所述待测试图像对应的第二特征向量、所述第二特征向量对应的单词间隔概率值或者非单词间隔概率值；利用当前得到的单词间隔概率值或者非单词间隔概率值，对所述待测试图像进行单词分割。
  2. 根据权利要求1所述的方法，其特征在于，所述获取样本图像，包括：
    从字体库中抽取至少两个字母图像，并对所述至少两个字母图像进行拼接，得到拼接图像，并为所述拼接图像设定单词间隔标记或者非单词间隔标记；
    将所述拼接图像标记为所述样本图像；或者，对所述拼接图像进行以下操作之一或者任意组合：平移操作、旋转操作、拉伸操作、添加高斯噪声，并将进行所述操作之后的图像标记为所述样本图像。
  3. 根据权利要求1所述的方法，其特征在于，所述方法进一步包括：
    在利用卷积神经网络对所述样本图像进行处理之前，将所述样本图像转换成灰度的样本图像，并将所述灰度的样本图像转换成大小为M*N的样本图像，并对所述大小为M*N的样本图像进行归一化处理；
    在利用所述卷积神经网络对所述待测试图像进行处理之前，将所述待测试图像转换成灰度的待测试图像，并将所述灰度的待测试图像转换成大小为M*L的待测试图像，并对所述大小为M*L的待测试图像进行归一化处理；
    其中，所述M为正整数、所述N为正整数、所述L为正整数，且所述M和所述L的比值，与所述待测试图像的长宽比相同。
  4. 根据权利要求1所述的方法，其特征在于，利用所述卷积神经网络对所述待测试图像进行处理，得到所述待测试图像对应的第二特征向量，包括：
    将所述待测试图像向左平移K次或者向右平移K次，每次平移时的平移位置不同，以得到K+1个不同的待测试图像，所述K为正整数；
    利用卷积神经网络对K+1个不同的待测试图像进行处理，得到K+1个不同的待测试图像分别对应的特征向量，并将所述K+1个不同的待测试图像分别对应的特征向量组 合在一起，得到所述待测试图像对应的第二特征向量。
  5. 根据权利要求1或4所述的方法，其特征在于，所述得到所述第二特征向量对应的单词间隔概率值或者非单词间隔概率值的过程，具体包括：
    利用所述卷积神经网络获得所述第二特征向量匹配的多个第一特征向量，并将所述多个第一特征向量对应的单词间隔概率值或者非单词间隔概率值，作为所述第二特征向量对应的单词间隔概率值或者非单词间隔概率值。
  6. 根据权利要求1或4所述的方法，其特征在于，利用当前得到的单词间隔概率值或者非单词间隔概率值，对所述待测试图像进行单词分割，包括：
    将所述第二特征向量对应的待测试图像位置作为横坐标，将所述第二特征向量对应的单词间隔概率值作为纵坐标，得到概率值结果图；从所述概率值结果图中找到单词间隔概率值的极大值位置，并利用所述极大值位置对所述待测试图像进行单词分割；或者，将所述第二特征向量对应的待测试图像位置作为横坐标，将所述第二特征向量对应的非单词间隔概率值作为纵坐标，得到概率值结果图；从所述概率值结果图中找到非单词间隔概率值的极小值位置，并利用所述极小值位置对所述待测试图像进行单词分割。
  7. 一种单词的分割装置，其特征在于，所述装置具体包括：
    训练模块，用于获取样本图像，所述样本图像包括单词间隔标记或者非单词间隔标记；利用卷积神经网络对所述样本图像进行处理，得到所述样本图像对应的第一特征向量、所述第一特征向量对应的单词间隔概率值和/或非单词间隔概率值；
    预测模块，用于获取待测试图像，并利用所述卷积神经网络对所述待测试图像进行处理，得到所述待测试图像对应的第二特征向量、所述第二特征向量对应的单词间隔概率值或者非单词间隔概率值；利用当前得到的单词间隔概率值或者非单词间隔概率值，对所述待测试图像进行单词分割。
  8. 根据权利要求7所述的装置，其特征在于，
    所述训练模块，具体用于在获取所述样本图像的过程中，从字体库中抽取至少两个字母图像，并对所述至少两个字母图像进行拼接，得到拼接图像，并为所述拼接图像设定单词间隔标记或者非单词间隔标记；
    将所述拼接图像标记为所述样本图像；或者，对所述拼接图像进行以下操作之一或者任意组合：平移操作、旋转操作、拉伸操作、添加高斯噪声，并将进行所述操作之后的图像标记为所述样本图像。
  9. 根据权利要求7所述的装置，其特征在于，
    所述训练模块，还用于在利用卷积神经网络对样本图像进行处理之前，将所述样本图像转换成灰度的样本图像，并将所述灰度的样本图像转换成大小为M*N的样本图像，并对所述大小为M*N的样本图像进行归一化处理；
    所述预测模块，还用于在利用卷积神经网络对待测试图像进行处理之前，将待测试图像转换成灰度的待测试图像，并将灰度的待测试图像转换成大小为M*L的待测试图像，并对所述大小为M*L的待测试图像进行归一化处理；
    其中，所述M为正整数、所述N为正整数、所述L为正整数，且所述M和所述L的比值，与所述待测试图像的长宽比相同。
  10. 根据权利要求7所述的装置，其特征在于，
    所述预测模块，具体用于在利用所述卷积神经网络对所述待测试图像进行处理，得到所述待测试图像对应的第二特征向量的过程中，将所述待测试图像向左平移K次或者向右平移K次，每次平移时的平移位置不同，以得到K+1个不同的待测试图像，所述K为正整数；
    利用卷积神经网络对K+1个不同的待测试图像进行处理，得到K+1个不同的待测试图像分别对应的特征向量，并将所述K+1个不同的待测试图像分别对应的特征向量组合在一起，得到所述待测试图像对应的第二特征向量。
  11. 根据权利要求7或10所述的装置，其特征在于，
    所述预测模块，具体用于在得到所述第二特征向量对应的单词间隔概率值或者非单词间隔概率值的过程中，利用所述卷积神经网络获得所述第二特征向量匹配的多个第一特征向量，并将所述多个第一特征向量对应的单词间隔概率值或者非单词间隔概率值，作为所述第二特征向量对应的单词间隔概率值或者非单词间隔概率值。
  12. 根据权利要求7或10所述的装置，其特征在于，
    所述预测模块，具体用于在利用当前得到的单词间隔概率值或者非单词间隔概率值，对所述待测试图像进行单词分割的过程中，将所述第二特征向量对应的待测试图像位置作为横坐标，将所述第二特征向量对应的单词间隔概率值作为纵坐标，得到概率值结果图；从所述概率值结果图中找到单词间隔概率值的极大值位置，并利用所述极大值位置对所述待测试图像进行单词分割；或者，将所述第二特征向量对应的待测试图像位置作为横坐标，将所述第二特征向量对应的非单词间隔概率值作为纵坐标，得到概率值结果图；从所述概率值结果图中找到非单词间隔概率值的极小值位置，并利用所述极小值位置对所述待测试图像进行单词分割。
PCT/CN2017/073718 2016-02-29 2017-02-16 一种单词的分割方法和装置 Ceased WO2017148265A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
US16/081,011 US10817741B2 (en) 2016-02-29 2017-02-16 Word segmentation system, method and device

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201610112463.6A CN107133622B (zh) 2016-02-29 2016-02-29 一种单词的分割方法和装置
CN201610112463.6 2016-02-29

Publications (1)

Publication Number Publication Date
WO2017148265A1 true WO2017148265A1 (zh) 2017-09-08

Family

ID=59721634

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2017/073718 Ceased WO2017148265A1 (zh) 2016-02-29 2017-02-16 一种单词的分割方法和装置

Country Status (4)

Country Link
US (1) US10817741B2 (zh)
CN (1) CN107133622B (zh)
TW (1) TWI744283B (zh)
WO (1) WO2017148265A1 (zh)

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109376739A (zh) * 2018-10-19 2019-02-22 北京奇艺世纪科技有限公司 一种编组方式确定方法及装置
CN109543777A (zh) * 2018-11-28 2019-03-29 中国科学院自动化研究所 手写汉字书写质量评价方法及系统
CN111062206A (zh) * 2018-10-16 2020-04-24 株式会社理光 一种子词单元的拆分方法、装置及计算机可读存储介质
CN111476067A (zh) * 2019-01-23 2020-07-31 腾讯科技（深圳）有限公司 图像的文字识别方法、装置、电子设备及可读存储介质
CN112528882A (zh) * 2020-12-15 2021-03-19 平安科技（深圳）有限公司 基于ocr识别房产证信息确定方法、装置、设备及介质

Families Citing this family (21)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2018033137A1 (zh) * 2016-08-19 2018-02-22 北京市商汤科技开发有限公司 在视频图像中展示业务对象的方法、装置和电子设备
US11120334B1 (en) * 2017-09-08 2021-09-14 Snap Inc. Multimodal named entity recognition
CN107957989B9 (zh) * 2017-10-23 2021-01-12 创新先进技术有限公司 基于集群的词向量处理方法、装置以及设备
CN108170663A (zh) 2017-11-14 2018-06-15 阿里巴巴集团控股有限公司 基于集群的词向量处理方法、装置以及设备
US10909406B2 (en) * 2018-01-08 2021-02-02 Newgen Software Technologies Limited Image processing system and method
CN110119505A (zh) * 2018-02-05 2019-08-13 阿里巴巴集团控股有限公司 词向量生成方法、装置以及设备
CA3115746C (en) * 2018-10-24 2023-08-29 Fujitsu Frontech Limited Banknote inspection device, banknote inspection method, and banknote inspection program
CN111401099B (zh) * 2018-12-28 2023-04-07 中国电信股份有限公司 文本识别方法、装置以及存储介质
US10616443B1 (en) * 2019-02-11 2020-04-07 Open Text Sa Ulc On-device artificial intelligence systems and methods for document auto-rotation
US10977525B2 (en) * 2019-03-29 2021-04-13 Fuji Xerox Co., Ltd. Indoor localization using real-time context fusion of visual information from static and dynamic cameras
CN110472227B (zh) * 2019-07-05 2023-08-18 平安国际智慧城市科技股份有限公司 一种词汇掌握程度评估方法、装置、终端设备及存储介质
CN110703075B (zh) * 2019-09-10 2022-03-15 国网浙江省电力有限公司嘉兴供电公司 一种基于大数据技术的高压断路器快速检修方法
CN110751234B (zh) * 2019-10-09 2024-04-16 科大讯飞股份有限公司 Ocr识别纠错方法、装置及设备
CN111104484B (zh) * 2019-12-19 2021-09-03 南京中孚信息技术有限公司 文本相似度检测方法、装置及电子设备
CN113282779B (zh) * 2020-02-19 2025-02-25 阿里巴巴集团控股有限公司 图像搜索方法、装置、设备
CN111597809B (zh) * 2020-06-09 2023-08-08 腾讯科技（深圳）有限公司 训练样本获取方法、模型训练方法、装置以及设备
CN112305506B (zh) * 2020-10-26 2023-02-14 哈尔滨工程大学 一种基于cnn的lpi雷达信号检测方法
CN114764765A (zh) * 2020-12-30 2022-07-19 富泰华工业（深圳）有限公司 图像瑕疵检测方法、装置、电子设备及存储介质
CN113780265B (zh) * 2021-09-16 2023-12-15 平安科技（深圳）有限公司 英文单词的空格识别方法、装置、存储介质及计算机设备
US12417505B2 (en) * 2022-09-14 2025-09-16 Capital One Services, Llc Detecting reliability across the internet after scraping
US20250124153A1 (en) * 2023-10-17 2025-04-17 Capital One Services, Llc Verifying digital files against rules

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20050027511A1 (en) * 2003-07-31 2005-02-03 Yoshihisa Ohguro Language recognition method, system and software
CN102254157A (zh) * 2011-07-07 2011-11-23 北京文通图像识别技术研究中心有限公司 一种寻找左右字符的字符切分位置评价方法
CN102822846A (zh) * 2010-03-30 2012-12-12 微软公司 检测文本行图像中的断字位置
CN103530600A (zh) * 2013-06-06 2014-01-22 东软集团股份有限公司 复杂光照下的车牌识别方法及系统
CN104050471A (zh) * 2014-05-27 2014-09-17 华中科技大学 一种自然场景文字检测方法及系统

Family Cites Families (24)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US5402504A (en) 1989-12-08 1995-03-28 Xerox Corporation Segmentation of text styles
US6108444A (en) 1997-09-29 2000-08-22 Xerox Corporation Method of grouping handwritten word segments in handwritten document images
FR2810765B1 (fr) 2000-06-27 2002-08-23 Mannesmann Dematic Postal Automation Sa Segmentation d'une image numerique d'un objet postal par la transformation de hough
GB2364416B (en) 2000-06-30 2004-10-27 Post Office Image processing for clustering related text objects
US7082219B2 (en) 2002-02-04 2006-07-25 The United States Of America As Represented By The Secretary Of The Air Force Method and apparatus for separating text from images
DE60325934D1 (de) 2002-11-22 2009-03-12 Oce Tech Bv Segmentierung eines Bildes mittels kürzester Zyklen
US7376272B2 (en) 2004-06-14 2008-05-20 Xerox Corporation Method for image segmentation to identify regions with constant foreground color
US8098936B2 (en) 2007-01-12 2012-01-17 Seiko Epson Corporation Method and apparatus for detecting objects in an image
US8009928B1 (en) 2008-01-23 2011-08-30 A9.Com, Inc. Method and system for detecting and recognizing text in images
US7471826B1 (en) 2008-03-31 2008-12-30 International Business Machines Corporation Character segmentation by slices
US8649600B2 (en) 2009-07-10 2014-02-11 Palo Alto Research Center Incorporated System and method for segmenting text lines in documents
US8509534B2 (en) 2010-03-10 2013-08-13 Microsoft Corporation Document page segmentation in optical character recognition
US9391716B2 (en) 2010-04-05 2016-07-12 Microsoft Technology Licensing, Llc Data center using wireless communication
CN102426649B (zh) * 2011-10-13 2013-08-21 石家庄开发区冀科双实科技有限公司 一种简单的高准确率的钢印数字自动识别方法
US20150055866A1 (en) 2012-05-25 2015-02-26 Mark Joseph Cummins Optical character recognition by iterative re-segmentation of text images using high-level cues
US8965127B2 (en) 2013-03-14 2015-02-24 Konica Minolta Laboratory U.S.A., Inc. Method for segmenting text words in document images
US9058517B1 (en) * 2014-04-16 2015-06-16 I.R.I.S. Pattern recognition system and method using Gabor functions
CN104463241A (zh) * 2014-10-31 2015-03-25 北京理工大学 一种智能交通监控系统中的车辆类型识别方法
US9430703B2 (en) 2014-12-19 2016-08-30 Konica Minolta Laboratory U.S.A., Inc. Method for segmenting text words in document images using vertical projections of center zones of characters
US9384409B1 (en) 2015-01-29 2016-07-05 Konica Minolta Laboratory U.S.A., Inc. Word segmentation for document image using recursive segmentation
WO2016197381A1 (en) 2015-06-12 2016-12-15 Sensetime Group Limited Methods and apparatus for recognizing text in an image
CN104966097B (zh) * 2015-06-12 2019-01-18 成都数联铭品科技有限公司 一种基于深度学习的复杂文字识别方法
CN104899571B (zh) * 2015-06-12 2018-12-04 成都数联铭品科技有限公司 一种用于复杂文字识别的随机样本产生方法
CN105469053A (zh) 2015-11-25 2016-04-06 成都数联铭品科技有限公司 一种基于贝叶斯优化的图像表格文字切分方法

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20050027511A1 (en) * 2003-07-31 2005-02-03 Yoshihisa Ohguro Language recognition method, system and software
CN102822846A (zh) * 2010-03-30 2012-12-12 微软公司 检测文本行图像中的断字位置
CN102254157A (zh) * 2011-07-07 2011-11-23 北京文通图像识别技术研究中心有限公司 一种寻找左右字符的字符切分位置评价方法
CN103530600A (zh) * 2013-06-06 2014-01-22 东软集团股份有限公司 复杂光照下的车牌识别方法及系统
CN104050471A (zh) * 2014-05-27 2014-09-17 华中科技大学 一种自然场景文字检测方法及系统

Cited By (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111062206A (zh) * 2018-10-16 2020-04-24 株式会社理光 一种子词单元的拆分方法、装置及计算机可读存储介质
CN111062206B (zh) * 2018-10-16 2023-11-21 株式会社理光 一种子词单元的拆分方法、装置及计算机可读存储介质
CN109376739A (zh) * 2018-10-19 2019-02-22 北京奇艺世纪科技有限公司 一种编组方式确定方法及装置
CN109376739B (zh) * 2018-10-19 2021-03-26 北京奇艺世纪科技有限公司 一种编组方式确定方法及装置
CN109543777A (zh) * 2018-11-28 2019-03-29 中国科学院自动化研究所 手写汉字书写质量评价方法及系统
CN111476067A (zh) * 2019-01-23 2020-07-31 腾讯科技（深圳）有限公司 图像的文字识别方法、装置、电子设备及可读存储介质
CN111476067B (zh) * 2019-01-23 2023-04-07 腾讯科技（深圳）有限公司 图像的文字识别方法、装置、电子设备及可读存储介质
CN112528882A (zh) * 2020-12-15 2021-03-19 平安科技（深圳）有限公司 基于ocr识别房产证信息确定方法、装置、设备及介质
CN112528882B (zh) * 2020-12-15 2024-05-10 平安科技（深圳）有限公司 基于ocr识别房产证信息确定方法、装置、设备及介质

Also Published As

Publication number Publication date
CN107133622B (zh) 2022-08-26
CN107133622A (zh) 2017-09-05
US20190019055A1 (en) 2019-01-17
TWI744283B (zh) 2021-11-01
US10817741B2 (en) 2020-10-27
TW201732651A (zh) 2017-09-16

Similar Documents

Publication Publication Date Title
TWI744283B (zh) 一種單詞的分割方法和裝置
CN109344701B (zh) 一种基于Kinect的动态手势识别方法
Saha et al. Multi-lingual scene text detection and language identification
CN112364873A (zh) 弯曲文本图像的文字识别方法、装置及计算机设备
CN110766020A (zh) 一种面向多语种自然场景文本检测与识别的系统及方法
Shivakumara et al. New gradient-spatial-structural features for video script identification
Maheshwari et al. Unmasking embedded text: A deep dive into scene image analysis
CN107766864B (zh) 提取特征的方法和装置、物体识别的方法和装置
CN116071294A (zh) 一种光纤表面缺陷检测方法和装置
CN111666842A (zh) 一种基于双流空洞卷积神经元网络的阴影检测方法
CN118135203A (zh) 一种基于深度学习预提取RoI的AprilTag检测方法
CN114386413A (zh) 处理数字化的手写
CN114882204A (zh) 船名自动识别方法
CN117218481A (zh) 鱼类识别方法、装置、设备及存储介质
CN109977834B (zh) 从深度图像中分割人手与交互物体的方法和装置
Zhang et al. Artifact detection in endoscopic video with deep convolutional neural networks
Gui et al. A fast caption detection method for low quality video images
CN115035390B (zh) 一种基于gan和特征增强的航拍图检测方法
Qin et al. Robust and accurate text stroke segmentation
CN120298859A (zh) 一种遮挡场景下小目标轮廓识别方法及系统
Bao et al. A new method for optical steel rope non-destructive damage detection
Suryani et al. Advancing Parking Systems: A Performance Comparison of MobileNet and Canny in License Plate Detection
CN117237921A (zh) 一种基于深度学习提高的交通标志小目标检测的方法
Thomas et al. Real Time Face Mask Detection and Recognition using Python
Shetty et al. Automated Identity Document Recognition and Classification (AIDRAC)-A Review

Legal Events

Date Code Title Description
NENP Non-entry into the national phase

Ref country code: DE

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

Ref document number: 17759116

Country of ref document: EP

Kind code of ref document: A1

122 Ep: pct application non-entry in european phase

Ref document number: 17759116

Country of ref document: EP

Kind code of ref document: A1