CA3144405A1 - Text information recognizing method, extracting method, devices and system - Google Patents

Abstract

Pertaining to the field of text information processing technology, the present invention discloses a text information recognizing method, an extracting method, and corresponding device and system. The recognizing method comprises: performing text detection on an image, and obtaining candidate boxes and corresponding original confidence levels;
calculating a loss parameter of the second candidate box according to size of the intersection area and size of a closure area between the first candidate box and the second candidate box;
calculating an original Intersection-over-Union between the first candidate box and the second candidate box, correcting the original Intersection-over-Union and obtaining corrected Intersection-over-Union according to the loss parameter of the second candidate box; calculating corrected confidence level of the second candidate box according to the corrected Intersection-over-Union and the original confidence level of the second candidate box; judging whether the corrected confidence level of the second candidate box satisfies a confidence level condition.

Description

TEXT INFORMATION RECOGNIZING METHOD, EXTRACTING METHOD, DEVICES AND SYSTEM
BACKGROUND OF THE INVENTION
Technical Field [0001] The present invention relates to the field of text information processing technology, and more particularly to a text information recognizing method, an extracting method, and corresponding devices and system.
Description of Related Art

[0002] Commodity marketing images mainly indicate images that are presented on the pages of an e-commerce operation platform to propagandize and popularize the commodities. In order to attract consumers to enhance sales rate, merchants would do everything possible to present ever more commodity information and sales promotional activity information in commodity marketing images, so the spacings between text lines in the commodity marketing images are usually not fixed, and this increases difficulties to persons skilled in the art in the process of recognizing text information in the commodity marketing images; moreover, since commodity sales promotional activity information changes frequently, the mode of human recognition cannot keep up with the change in the text information, so there is an urgent need to invent an automatic text information recognition technique applicable to complex images.
SUMMARY OF THE INVENTION

[0003] In order to solve the problems pending in the state of the art, embodiments of the present invention provide a text information recognizing method, an extracting method, and Date Recue/Date Received 2021-12-30 corresponding devices and system. The technical solutions are as follows.

[0004] According to the first aspect, there is provided a text information recognizing method, and the method comprises:

[0005] performing text detection on an image, and obtaining candidate boxes identifying text line locations in the image, and original confidence levels to which the various candidate boxes correspond;

[0006] selecting the candidate box with the greatest original confidence level from the candidate boxes sharing intersection areas to serve as a first candidate box, selecting any other candidate box to serve as a second candidate box, and calculating a loss parameter of the second candidate box according to a size of the intersection area between the first candidate box and the second candidate box and a size of a closure area;

[0007] calculating an original Intersection-over-Union between the first candidate box and the second candidate box, correcting the original Intersection-over-Union according to the loss parameter of the second candidate box, and obtaining a corrected Intersection-over-Union;

[0008] calculating a corrected confidence level of the second candidate box according to the corrected Intersection-over-Union and the original confidence level of the second candidate box;

[0009] judging whether the corrected confidence level of the second candidate box satisfies a confidence level condition, if yes, taking the first candidate box and the second candidate box as text boxes to be recognized; and

[0010] recognizing text information in the text boxes to be recognized.

[0011] Further, the step of calculating a loss parameter of the second candidate box according to a size of the intersection area between the first candidate box and the second candidate box and a size of a closure area includes:

[0012] obtaining a width and a height of the intersection area, and obtaining a width and a height of the closure area; and Date Recue/Date Received 2021-12-30

[0013] calculating the loss parameter of the second candidate box according to a height ratio of the intersection area to the closure area and a width ratio of the intersection area to the closure area.

[0014] Further, the step of calculating a corrected confidence level of the second candidate box according to the corrected Intersection-over-Union and the original confidence level of the second candidate box includes:

[0015] calculating an attenuation weight of the second candidate box according to the corrected Intersection-over-Union; and

[0016] employing the attenuation weight of the second candidate box to correct the original confidence level of the second candidate box, and obtaining the corrected confidence level of the second candidate box.

[0017] Further, the step of recognizing text information in the text boxes to be recognized includes:

[0018] using a neutral network model to recognize text information in the text boxes to be recognized, the neutral network model including a convolutional layer and a pooling layer;
wherein

[0019] the convolutional layer includes alternately connected standard convolutional kernels and dilated convolutional kernels, and a width of a receptive field of the dilated convolutional kernel is larger relative to a width of a receptive field of the standard convolutional kernel;
and

[0020] wherein the pooling layer has rectangular block windows, and adopts weighting mix pooling of standard maximum pooling and average pool, and a pooling weight coefficient is calculated and determined according to a global maximum value and mean value of block images.

[0021] According to the second aspect, there is provided a text information extracting method, and the method comprises:

Date Recue/Date Received 2021-12-30

[0022] employing the method according to anyone of the first aspect to recognize text information in a commodity image;

[0023] matching the text information with a pre-created extended warranty synonym dictionary, the extended warranty synonym dictionary containing extended warranty keywords and synonyms;

[0024] performing synonym replacement to the extended warranty keywords contained in the text information after matching has succeeded; and

[0025] extracting the text information having been performed with the synonym replacement.

[0026] Further, the method further comprises:

[0027] determining a first extended warranty time limit according to the extended warranty text information, comparing the first extended warranty time limit with a second extended warranty time limit, and obtaining a time limit difference;

[0028] comparing the time limit difference with a first time limit threshold condition, determining a corresponding commodity code for the extended warranty text information that satisfies the first time limit threshold condition, and determining customer communication information according to the commodity code; and

[0029] sending to a customer extended warranty pushing information corresponding to a comparison result of the time limit difference according to the customer communication information.

[0030] Further, the method further comprises:

[0031] determining a first extended warranty time limit according to the extended warranty text information, comparing the first extended warranty time limit with a second extended warranty time limit, and obtaining a time limit difference; and

[0032] comparing the time limit difference with a second time limit threshold condition, and determining a display location of extended warranty text information, with which the second extended warranty time limit associates, on a webpage according to a comparison result.

Date Recue/Date Received 2021-12-30

[0033] According to the third aspect, there is provided a text information recognizing device, and the device comprises:

[0034] a detecting module, for performing text detection on an image, and obtaining candidate boxes identifying text line locations in the image, and original confidence levels to which the various candidate boxes correspond;

[0035] a loss parameter calculating module, for selecting the candidate box with the greatest original confidence level from the candidate boxes sharing intersection areas to serve as a first candidate box, selecting any other candidate box to serve as a second candidate box, and calculating a loss parameter of the second candidate box according to a size of the intersection area between the first candidate box and the second candidate box and a size of a closure area;

[0036] an Intersection-over-Union correcting module, for calculating an original Intersection-over-Union between the first candidate box and the second candidate box, correcting the original Intersection-over-Union according to the loss parameter of the second candidate box, and obtaining a corrected Intersection-over-Union;

[0037] a confidence level correcting module, for calculating a corrected confidence level of the second candidate box according to the corrected Intersection-over-Union and the original confidence level of the second candidate box;

[0038] a to-be-recognized text box obtaining module, for judging whether the corrected confidence level of the second candidate box satisfies a confidence level condition, if yes, taking the first candidate box and the second candidate box as text boxes to be recognized;
and

[0039] a recognizing module, for recognizing text information in the text boxes to be recognized.

[0040] Further, the loss parameter calculating module is specifically employed for:

[0041] obtaining a width and a height of the intersection area, and obtaining a width and a height of the closure area; and

[0042] calculating the loss parameter of the second candidate box according to a height ratio of the intersection area to the closure area and a width ratio of the intersection area to the Date Recue/Date Received 2021-12-30 closure area.

[0043] Further, the confidence level correcting module is specifically employed for:

[0044] calculating an attenuation weight of the second candidate box according to the corrected Intersection-over-Union; and

[0045] employing the attenuation weight of the second candidate box to correct the original confidence level of the second candidate box, and obtaining the corrected confidence level of the second candidate box.

[0046] Further, the recognizing module is specifically employed for:

[0047] using a neutral network model to recognize text information in the text boxes to be recognized,

[0048] the neutral network model used by the recognizing module includes a convolutional layer and a pooling layer; wherein

[0049] the convolutional layer includes alternately connected standard convolutional kernels and dilated convolutional kernels, and a width of a receptive field of the dilated convolutional kernel is larger relative to a width of a receptive field of the standard convolutional kernel;
and

[0050] wherein the pooling layer has rectangular block windows, and adopts weighting mix pooling of standard maximum pooling and average pool, and a pooling weight coefficient is calculated and determined according to a global maximum value and mean value of block images.

[0051] According to the fourth aspect, there is provided a text information extracting device, and the device comprises:

[0052] a text information recognizing module, for executing the method according to any of the second aspect to recognize text information;

[0053] a matching module, for matching the text information with a pre-created extended warranty synonym dictionary, the extended warranty synonym dictionary containing Date Recue/Date Received 2021-12-30 extended warranty keywords and synonyms;

[0054] a filtering module, for performing synonym replacement to the extended warranty keywords contained in the text information after matching has succeeded; and

[0055] an extracting module, for extracting the text information having been performed with synonym replacement.

[0056] Further, the text information extracting device further comprises:

[0057] an extended warranty synonym dictionary updating module, for updating the extended warranty synonym dictionary, and specifically for:

[0058] determining extended warranty keywords;

[0059] extracting sample text information associated with the extended warranty keywords from a sample image, and employing a word segmentation tool to perform word segmentation on the sample text information;

[0060] judging whether the sample text information contains any invalid word, if yes, deleting the sample text information, if not, performing similarity calculation on the sample text information and vocabulary in an extended warranty synonym database; and

[0061] determining according to a similarity calculation result whether vocabulary in the sample text information is synonymous with the extended warranty keywords, if yes, adding the synonyms to the extended warranty synonym dictionary.

[0062] Further, the text information extracting device further comprises:

[0063] an information pushing determining module, for:

[0064] determining a first extended warranty time limit according to the extended warranty text information, comparing the first extended warranty time limit with a second extended warranty time limit, and obtaining a time limit difference;

[0065] comparing the time limit difference with a first time limit threshold condition, determining a corresponding commodity code for the extended warranty text information that satisfies the first time limit threshold condition, and determining customer communication information according to the commodity code; and Date Recue/Date Received 2021-12-30

[0066] sending to a customer extended warranty pushing information corresponding to a comparison result of the time limit difference according to the customer communication information.

[0067] Further, the text information extracting device further comprises:

[0068] an information display determining module, for:

[0069] determining a first extended warranty time limit according to the extended warranty text information, comparing the first extended warranty time limit with a second extended warranty time limit, and obtaining a time limit difference; and

[0070] comparing the time limit difference with a second time limit threshold condition, and determining a display location of extended warranty text information, with which the second extended warranty time limit associates, on a webpage according to a comparison result of the time limit difference.

[0071] According to the fifth aspect, there is provided a computer system that comprises:

[0072] one or more processor(s); and

[0073] a memory, associated with the one or more processor(s), wherein the memory is employed to store a program instruction, and the program instruction executes the method according to any of the first aspect when it is read and executed by the one or more processor(s).

[0074] The technical solutions provided by the embodiments of the present invention bring about the following advantageous effects.
1. In the text information recognizing method and device disclosed by the embodiments of the present invention, calculation of sizes of the intersection area of the candidate box with the greatest original confidence level and another candidate box and the closure area is introduced to obtain a loss parameter, the Intersection-over-Union between the candidate box with the greatest original confidence level and the other candidate box is corrected, the corrected Date Recue/Date Received 2021-12-30 Intersection-over-Union is utilized to calculate an attenuation weight coefficient, and the attenuation weight coefficient is utilized to correct the original confidence level, so the method and device are applicable to the extraction of text information in images with relatively small text line spacing, and effectively prevent missed detection of text lines.
2. The text information extracting method and device disclosed by the embodiments of the present invention effectively and precisely recognize extended warranty text information, and make use of the extended warranty text information to precisely push extended warranty information and determine its display location on the page, whereby the effect for marketing products with extended warranty is enhanced.
BRIEF DESCRIPTION OF THE DRAWINGS

[0075] To more clearly describe the technical solutions in the embodiments of the present invention, drawings required to illustrate the embodiments will be briefly introduced below. Apparently, the drawings introduced below are merely directed to some embodiments of the present invention, while persons ordinarily skilled in the art may further acquire other drawings on the basis of these drawings without spending creative effort in the process.

[0076] Fig. 1 is a view illustrating the effect of currently available text detection provided by an embodiment of the present invention;

[0077] Fig. 2 is a flowchart illustrating the text information recognizing method provided by an embodiment of the present invention;

[0078] Fig. 3 is a view schematically illustrating an interaction area and a closure area provided by an embodiment of the present invention;

[0079] Fig. 4 is a view illustrating the effect of text detection disclosed in the present invention Date Recue/Date Received 2021-12-30 provided by an embodiment of the present invention;

[0080] Fig. 5 is a flowchart illustrating the text information extracting method provided by an embodiment of the present invention;

[0081] Fig. 6 is a view illustrating the process of determining extended warranty pushing information comprised in the text information extracting method provided by an embodiment of the present invention;

[0082] Fig. 7 is a view illustrating the process of determining extended warranty information display comprised in the text information extracting method provided by an embodiment of the present invention;

[0083] Fig. 8 is a view schematically illustrating the structure of the text information recognizing device provided by an embodiment of the present invention;

[0084] Fig. 9 is a view schematically illustrating the structure of the text information extracting device provided by an embodiment of the present invention; and

[0085] Fig. 10 is a view schematically illustrating the structure of the computer system provided by an embodiment of the present invention.
DETAILED DESCRIPTION OF THE INVENTION

[0086] To make more lucid and clear the objectives, technical solutions and advantages of the present invention, the technical solutions in the embodiments of the present invention will be clearly and comprehensively described below with reference to the accompanying drawings in the embodiments of the present invention. Apparently, the embodiments as described are merely partial, rather than the entire, embodiments of the present invention.
Any other embodiments makeable by persons ordinarily skilled in the art on the basis of the embodiments in the present invention without creative effort shall all fall within the protection scope of the present invention.

[0087] Extended warranty, namely guarantee over an extended period, means an elongation of guarantee time or extension of product service range provided by an extended warranty Date Recue/Date Received 2021-12-30 provider in addition to the quality guarantee period and service range provided by the manufacturer of the purchased product. Currently, electric appliance marketers and e-commerce service providers both serve as extended warranty providers to provide extended warranty services, and it is therefore possible to facilitate precise pushing of extended warranty activities of e-commerce service providers if extended warranty information of electric appliance marketers can be automatically recognized.

[0088] Since bonus extended warranty is only one type of promotional means of electric appliance marketers, the information of other promotional means such as the free provision of special-offer coupons, gifts and maintenance services will be displayed together with the bonus extended warranty information in electric appliance images in the sales pages of electric appliances. As a result, in order that the electric appliance marketers display promotional activity information more through electric appliance images, the electric appliance images are usually presented in a way that the information is very complicated and variegated, and the spacing between lines of the text information is relatively small.

[0089] It is modus operandi in prior-art technology for recognizing text information in an image to firstly perform text detection, obtain text boxes identifying text line locations and confidence levels to which the text boxes correspond, subsequently employ a non-maximum suppression algorithm to remove other text boxes overlapping relatively greatly in area with the text box having the greatest confidence level, and finally recognize the text information in the text boxes that remain. As shown in Fig.
1, the aforementioned text information detecting method is applied to such a recognition scenario with relatively small spacing between text lines as an electric appliance image, since the spacing between text lines is relatively small, during the process of screening text boxes by means of the non-maximum suppression algorithm, the text box that identifies text lines is easily deleted as a text box overlapping relatively greatly in area with the text box having the greatest confidence level, whereby text information is Date Recue/Date Received 2021-12-30 incompletely recognized.

[0090] In order to overcome the aforementioned problems, the present invention provides a text information recognizing method and a text information recognizing device, and the specific technical solutions are described below.

[0091] As shown in Fig. 2, the text information recognizing method comprises the following steps.

[0092] S21 - performing text detection on an image, and obtaining candidate boxes identifying text line locations in the image, and original confidence levels to which the various candidate boxes correspond.

[0093] The text detection is mainly directed to finding out the locations where the text lines in the image locate, and the CTPN model algorithm can be employed in the present invention therefor.

[0094] In one embodiment, use of the CTPN model algorithm mainly includes the following process.

[0095] S211 ¨ preparing a model training sample set.

[0096] In order to adapt to the text detection of commodity images, commodity images uploaded by plural e-commerce services can be used in selecting training samples, such as the images in the TaoBao commodity network image data set of Ali Tianchi ICP2018, and images in the Suning Tesco image data set. Since images coming from different origins have different data formats, some of which are of four coordinate points starting from the upper left corner as the original point while others of which are of two coordinate points at the upper left and the lower right, it is therefore required to process the images having Date Recue/Date Received 2021-12-30 different data formats into Bbox labels with a unified data format. The images and the labels are thereafter scaled to the same specification.

[0097] Due to the anchor mechanism of CTPN, it is therefore required to transform the Bbox labels into anchor labels, while foreground and background classification labels, longitudinal coordinates and height of the anchor central point, and horizontal offsets serve as labels in the training sample set.

[0098] S212 ¨ constructing a text detection model, and employing the training sample set to train the text detection model.

[0099] The text detection model is embodied as a neutral network model, in which the CNN
layer employs VGG16 to extract spatial features; the intermediate layer thereafter transforms a sliding local block extracted from conv5 of VGG16 into the input of LSTM;
the RNN layer employs BilSTM to extract sequence features; a fully connected layer is finally connected, including classification of foreground and background, positioning of longitudinal coordinates and height of the anchor central point, and boundary-optimized multi-task loss functions.

[0100] The aforementioned training sample set is used to train the neutral network model, and obtain a text detection model capable of obtaining candidate boxes and the corresponding confidence levels.

[0101] S22 - selecting the candidate box with the greatest original confidence level from the candidate boxes sharing intersection areas to serve as a first candidate box, selecting any other candidate box to serve as a second candidate box, and calculating a loss parameter of the second candidate box according to a size of the intersection area between the first candidate box and the second candidate box and a size of a closure area.

Date Recue/Date Received 2021-12-30

[0102] In one embodiment, step S22 includes the following.

[0103] S221 - obtaining a width and a height of the intersection area, and a width and a height of the closure area.

[0104] As shown in Fig. 3, area a in Fig. 3 is an intersection area, and areas b indicated by dotted lines are closure areas.

[0105] S222 - calculating the loss parameter according to a height ratio of the intersection area to the closure area and a width ratio of the intersection area to the closure area.

[0106] Specifically, the calculation formula of loss parameter C is as follows:
h, w c C = (k1 X ¨ + k2 , .., ¨) 2 hi, wu where hc and wc are respectively the height and the width of the intersection area, wc and Iv, are respectively the height and the width of the closure area, ki and k2 are respectively weight coefficients, preferably, k1 = 0.7, k2 = 0.3.

[0107] When the two candidate boxes respectively identify two text line locations, the ratio of the height of their closure area to the height of their intersection area is smaller than the ratio of the height of a closure area to the height of an intersection area of two candidate boxes that identify one text line location, so the current loss parameter corrects the original confidence level mainly from the height, and so the weight coefficient of the height ratio is greater than the weight coefficient of the width ratio.

[0108] In the calculation formula of the loss parameter are included the height ratio and the width ratio between the intersection area and the closure area.

[0109] S23 - calculating an original Intersection-over-Union between the first candidate box and Date Recue/Date Received 2021-12-30 the second candidate box, correcting the original Intersection-over-Union according to the loss parameter of the second candidate box, and obtaining a corrected Intersection-over-Union.

[0110] The loss parameter calculated in the above step S22 is mainly used to correct the original Intersection-over-Union between the first candidate box and the second candidate box.

[0111] Accordingly, in one embodiment, step S23 includes the following.

[0112] S231 - calculating an original Intersection-over-Union between the first candidate box and the second candidate box, wherein the original Intersection-over-Union is a ratio between intersection area and union area of the first candidate box and the second candidate box.

[0113] The calculation formula of the original Intersection-over-Union IOU is as follows:
¨ area(M) n area(Si) IOU
area(M)U area(Si) where M indicates the first candidate box, Si indicates the second candidate box, area(M) n area(Si) is the intersection area of the two candidate boxes, and area(M)U area(Si) is the union area of the two candidate boxes.

[0114] The original Intersection-over-Union IOU reflects the overlapping circumstance of the two candidate boxes.

[0115] S232 ¨ employing the loss parameter of the second candidate box to correct the original Intersection-over-Union, and obtaining a corrected Intersection-over-Union.

[0116] The corrected Intersection-over-Union is a difference between the original Intersection-over-Union and the loss parameter, and the calculation formula of the corrected Date Recue/Date Received 2021-12-30 Intersection-over-Union IOUnew is specifically as follows:
10U,,,,, = IOU ¨ C
where IOU is the original Intersection-over-Union, and C is the loss parameter.

[0117] The loss parameter to which two candidate boxes identifying two text line locations correspond is smaller than the loss parameter to which two candidate boxes identifying one text line location correspond, so the corrected Intersection-over-Union of the two candidate boxes identifying two text line locations is larger.

[0118] S24 - calculating a corrected confidence level of the second candidate box according to the corrected Intersection-over-Union and the original confidence level of the second candidate box.

[0119] S241 - employing the attenuation weight of the second candidate box to correct the original confidence level, and obtaining the corrected confidence level of the second candidate box.

[0120] Original confidence levels are confidence levels to which the various candidate boxes correspond as obtained by using the text information detection model in step S21.
Through the text information detection model can be obtained more than one candidate box, so it is required to screen the candidate boxes, usually by employing the NMS non maximum suppression algorithm. When the IOU between the first candidate box having the greatest original confidence level and the second candidate box is greater than a preset IOU threshold, the traditional non maximum suppression algorithm lowers the original confidence level of this candidate box to 0, so that this candidate box is deleted when screening is made according to the original confidence levels; the traditional non maximum suppression algorithm is as shown below:

Date Recue/Date Received 2021-12-30 newscore = scorei, /0 U(M, S1)< threshold 0, /0U(M, Si) threshold where newscore, is the original confidence level after screening by the traditional non maximum suppression algorithm, score, is an original confidence level, M
indicates the first candidate box, S, indicates the second candidate box, and threshold is an IOU
threshold.

[0121] Use of the traditional non maximum suppression algorithm would lead to missed detection of other candidate boxes identifying new text line locations and overlapping relatively greatly in area with the first candidate box, so the present invention provides a method of correcting the original confidence level that makes use of Gaussian weight coefficients and adds loss parameters to further increase punishment in terms of height, with the specific calculation formula as follows:
newscorei = scorei x f (10Une,(M, Si)) where newscore, is the original confidence level after screening by the traditional non maximum suppression algorithm, score, is an original confidence level, and f (10Une,(M,Si)) is an attenuation weight coefficient relevant to the corrected iounew2 Intersection-over-Union, specifically, f (10Une,(M, Si)) = e 0-0¨c) , where a is a standard deviation parameter of a Gaussian function.

[0122] The attenuation weight coefficient to which two candidate boxes identifying two text line locations correspond is greater than the attenuation weight coefficient to which two candidate boxes identifying one text line location correspond, so the corrected confidence level obtained thereby is also relatively great, and would not be deleted.

[0123] S25 -judging whether the corrected confidence level of the second candidate box satisfies a confidence level condition, if yes, taking the first candidate box and the second candidate box as text boxes to be recognized.

Date Recue/Date Received 2021-12-30

[0124] Finally, the text boxes to be recognized as obtained are as shown in Fig. 4. In comparison of Fig. 4 with Fig. 1, the candidate box identifying the text line location below is retained in Fig. 4, and there would be no missed detection.

[0125] In summary, the aforementioned steps S21 to S25 provide a text information detecting method, in which a loss parameter with the dominant height ratio between the intersection area and the closure area of the first candidate box and another candidate box is introduced to correct the original Intersection-over-Union, employ the corrected Intersection-over-Union as obtained and the loss parameter to calculate the attenuation weight coefficient, and employ the attenuation weight coefficient to correct the original confidence level; in comparison with the traditional practice of reducing to 0 the confidence level of the candidate box that does not satisfy the confidence level threshold condition, the claimed method is more flexible, and does not engender the circumstance of missed detection of any candidate box.

[0126] S26 - recognizing text information in the text boxes to be recognized.

[0127] In one embodiment, a CRNN model is employed in the present application to recognize text information in the text box to be recognized, specifically including the following.

[0128] S261 ¨ preparing a training set.

[0129] The training set can specifically be selected from common Chinese words in a Chinese corpus, English letters, digits and punctuations, and the image data set of the current enterprise. The data set is transformed into lmdb format. An image is performed with grayscale and scaling normalization processes prior to being input in the model, each pixel is transformed into a digit of [-1,11, and word labels to which the image corresponds are coded and transformed into digits.

Date Recue/Date Received 2021-12-30

[0130] S262 ¨ constructing a text recognition model.

[0131] The text recognition model is embodied as a CRNN model with a CNN+RNN+CTC
framework. The traditional CNN network employs VGG to extract spatial features, which are subsequently input in the RNN layer to employ BilSTM to extract sequence features and predict label distribution, and CTC-Loss is finally used to solve the problem of aligning the variable-length sequence. After the model has been well trained, the prediction result is decoded and output to obtain corresponding words.

[0132] In order to adapt to the characteristics of lesser height and wider length of text lines in the image, the present invention makes improvement over the CNN network, in which the convolutional layer includes alternately connected standard convolutional kernels and dilated convolutional kernels, and a width of a receptive field of the dilated convolutional kernel is larger relative to a width of a receptive field of the standard convolutional kernel.
The block window of the pooling layer is modified from a square into a rectangle with unequal lengths and widths, and is mixedly pooled by standard maximum pooling and average pool weighting, and a pooling weight coefficient is calculated and determined according to a global maximum value and a global mean value of a block image.

[0133] Specifically, the calculation formula of the mixed value mixed (x) of mixed pooling is as follows:
mixed (x) = w x max (x) + (1 ¨ w)mean(x) where w is the pooling weight coefficient, max(x) is the global maximum value, and mean(x) is the mean value; specifically, the calculation formula of w is as follows:
abs(max(x)) W =
abs(max(x)) + abs(mean(x))

[0134] As shown in Fig. 5, the present invention further provides a text information extracting method employed mainly to extract extended warranty text information in a commodity Date Recue/Date Received 2021-12-30 image, and the specific technical solution is described below.

[0135] S51 - employing the text information recognizing method disclosed above to recognize text information in a commodity image.

[0136] The commodity image mainly indicates an image that contains a commodity and commodity promotional text, or an image that only contains commodity promotional text.

[0137] S52 - matching the text information with a pre-created extended warranty synonym dictionary, the extended warranty synonym dictionary containing extended warranty keywords and synonyms.

[0138] In one embodiment, an embodiment of the present invention further discloses a method of updating an extended warranty synonym dictionary, and the method comprises the following.

[0139] S521 ¨ determining extended warranty keywords.

[0140] The extended warranty keywords can be: guarantee, extended warranty, etc.

[0141] S522 - extracting sample text information associated with the extended warranty keywords from a sample image, and employing a word segmentation tool to perform word segmentation on the sample text information.

[0142] The sample image can be collected and obtained from multiple sources.

[0143] S523 - judging whether the sample text information contains any invalid word, if yes, deleting the sample text information, if not, performing similarity calculation on the sample text information and vocabulary in an extended warranty synonym database.
Date Recue/Date Received 2021-12-30

[0144] Invalid words are invalid vocabulary designated by persons skilled in the art, such as the extended warranty vocabulary that is no longer in use, and so on. This step is directed to a filtering process made on the sample text information in advance. The extended warranty synonym database usually makes use of a large-scale synonym dictionary, such as the Expanded Version of Synonym Dictionary; the database is a tree-shaped classification system consisting of 5 layers altogether, and this database is employed to perform similarity calculation, taking for example the similarity calculation of expression A "free repair" and expression B "repair", the code of expression A in the database is Hd04B03#, and the code of expression B in the database is Hd04B01=, then the similarity Sim (A, B) calculation formula is as follows:
n x n n¨k+ 1 Sim(A, B) = a x cos (80) x _______________________________ n where n is the number of codes starting with Hd04B, k is a phase difference of the fifth layer, and a are parameters of various layer branches with a valuation range of [0,1], valuation is empirically made, usually speaking, the deeper the layer is, the greater the a will be, then the similarity of the two expressions "free repair" and "repair" as calculated is 79.56%.

[0145] S524 - determining according to a similarity calculation result whether vocabulary in the sample text information is synonymous with the extended warranty keywords, if yes, adding the synonyms to the extended warranty synonym dictionary.

[0146] The calculated similarity is compared with a similarity threshold condition, if the similarity threshold condition is satisfied, it is determined that the two are synonyms.

[0147] S53 - performing synonym replacement to the extended warranty keywords contained in the text information after matching has succeeded, and deleting text information that contains no extended warranty keyword.

Date Recue/Date Received 2021-12-30

[0148] S54 - extracting the text information having been performed with synonym replacement.

[0149] In one embodiment, in order to make full use of the extended warranty text information to precisely push extended warranty services, as shown in Fig. 6, the text information extracting method disclosed by the present invention further comprises an extended warranty pushing information determining method that comprises the following steps.

[0150] S61 - determining a first extended warranty time limit according to the extended warranty text information, comparing the first extended warranty time limit with a second extended warranty time limit, and obtaining a time limit difference.

[0151] The first extended warranty time limit mainly indicates an extended warranty time limit provided by the commodity producer, and the second extended warranty time limit mainly indicates an extended warranty time limit provided by the e-commerce platform.

[0152] S62 - comparing the time limit difference with a first time limit threshold condition, determining a corresponding commodity code for the extended warranty text information that satisfies the first time limit threshold condition, and determining customer communication information according to the commodity code.

[0153] The time limit difference is mainly used to determine customers having further extended warranty service requirement. When the second extended warranty time limit is greater than the first extended warranty time limit and exceeds the first extended warranty time limit by a certain limit, it is then determined that the customer who purchased the commodity in point has further extended warranty service requirement.

[0154] The use of the extended warranty text information to determine the commodity code can be done according to the commodity name in the extended warranty information, or Date Recue/Date Received 2021-12-30 according to the commodity image to which the extended warranty text information corresponds. The customer communication information can be the mobile phone number of the customer, or any other contact ID.

[0155] S63 - sending to a customer extended warranty pushing information corresponding to a comparison result of the time limit difference according to the customer communication information.

[0156] The extended warranty pushing information corresponds to the comparison result of the time limit difference and the first time limit threshold condition, for instance:

[0157] when the time limit difference is not smaller than the first time limit threshold, this indicates that the extended warranty time limit provided by the commodity producer is longer than or equal to the extended warranty time limit provided by the e-commerce platform, in which case the corresponding extended warranty pushing information can be to push extended warranty and supplementary purchasing information of other commodities;

[0158] when the time limit difference is smaller than the first time limit threshold, this indicates that the extended warranty time limit provided by the commodity producer is longer than the extended warranty time limit provided by the e-commerce platform, in which case the corresponding extended warranty pushing information can be to push extended warranty and supplementary purchasing information of the commodity in point; and

[0159] with respect to a commodity that does not have the first extended warranty time limit, this indicates that the commodity producer does not provide any extended warranty service, in which case the corresponding extended warranty pushing information can be to push extended warranty and supplementary purchasing information of the commodity in point.

[0160] In one embodiment, in order to make full use of the extended warranty text information to clearly display the extended warranty service on the commodity sales page, as shown Date Recue/Date Received 2021-12-30 in Fig. 7, the information extracting method disclosed by the present invention further comprises a method of determining a display location of extended warranty information, and the method comprises the following steps.

[0161] S71 - determining a first extended warranty time limit according to the extended warranty text information, comparing the first extended warranty time limit with a second extended warranty time limit, and obtaining a time limit difference.

[0162] The first extended warranty time limit mainly indicates an extended warranty time limit provided by the commodity producer, and the second extended warranty time limit mainly indicates an extended warranty time limit provided by the e-commerce platform.

[0163] S72 - comparing the time limit difference with a second time limit threshold condition, and determining a display location of extended warranty text information, with which the second extended warranty time limit associates, on a webpage according to a comparison result of the time limit difference.

[0164] The second time limit threshold condition can include two thresholds, namely threshold 1 and threshold 2. The display location mainly indicates the sequential location in a page, for instance:

[0165] when the time limit difference is not smaller than threshold 1, this indicates that the extended warranty time limit provided by the commodity producer is by far longer than and exceeds the extended warranty time limit provided by the e-commerce platform, in which case the corresponding extended warranty text information is not displayed on the page;

[0166] when the time limit difference is smaller than threshold 1 and not smaller than threshold 2, this indicates that the extended warranty time limit provided by the commodity producer is slightly longer than, or does not exceed, the extended warranty time limit provided by the e-commerce platform, in which case the corresponding extended Date Recue/Date Received 2021-12-30 warranty text information is displayed on the page in the rear;

[0167] when the time limit difference is not greater than threshold 2, this indicates that the extended warranty time limit provided by the commodity producer is slightly shorter than the extended warranty time limit provided by the e-commerce platform, in which case the corresponding extended warranty text information is displayed on the page in the front;
and

[0168] with respect to a commodity that does not have the first extended warranty time limit, this indicates that the commodity producer does not provide any extended warranty service, in which case the corresponding extended warranty text information is normally emphatically displayed on the page according to the preset display location.

[0169] Steps S61 to S63 and steps S71 to S72 in the aforementioned two embodiments are all directed to methods of propagandizing extended warranty information by means of a time limit difference between the first extended warranty time limit and the second extended warranty time limit, so as to enable the extended warranty information to be precisely pushed.

[0170] As shown in Fig. 8, on the basis of the text information recognizing method provided by the embodiment of the present invention, an embodiment of the present invention further provides a text information recognizing device that comprises the following modules.

[0171] A detecting module 801 is employed for performing text detection on an image, and obtaining candidate boxes identifying text line locations in the image, and original confidence levels to which the various candidate boxes correspond.

[0172] The detecting module 801 performs text detection with a well-trained text detection model, and the text detection model is embodied as a neutral network model, in which the CNN layer employs VGG16 to extract spatial features; the intermediate layer thereafter transforms a sliding local block extracted from conv5 of VGG16 into the input Date Recue/Date Received 2021-12-30 of LSTM; the RNN layer employs BilSTM to extract sequence features; a fully connected layer is finally connected, including classification of foreground and background, positioning of longitudinal coordinates and height of the anchor central point, and boundary-optimized multi-task loss functions.

[0173] A loss parameter calculating module 802 is employed for selecting the candidate box with the greatest original confidence level from the candidate boxes sharing intersection areas to serve as a first candidate box, selecting any other candidate box to serve as a second candidate box, and calculating a loss parameter of the second candidate box according to a size of the intersection area and a size of a closure area between the first candidate box and the second candidate box.

[0174] In one embodiment, the loss parameter calculating module 802 is specifically employed for:

[0175] obtaining a width and a height of the intersection area, and obtaining a width and a height of the closure area; and

[0176] calculating the loss parameter according to a height ratio of the intersection area to the closure area and a width ratio of the intersection area to the closure area.

[0177] Specifically, the calculation formula of loss parameter C is as follows:
h c wc \ 2 C = (k1 X ¨ + k2 , .., ¨) wu where hc and wc are respectively the height and the width of the intersection area, wc and wi, are respectively the height and the width of the closure area, ki and k2 are respectively weight coefficients, preferably, ki = 0.7, k2 = 0.3.

[0178] An Intersection-over-Union correcting module 803 is employed for calculating an original Intersection-over-Union between the first candidate box and the second candidate box, correcting the original Intersection-over-Union according to the loss Date Recue/Date Received 2021-12-30 parameter of the second candidate box, and obtaining a corrected Intersection-over-Union.

[0179] In one embodiment, the Intersection-over-Union correcting module 803 is specifically employed for:

[0180] calculating an original Intersection-over-Union between the first candidate box and the second candidate box, wherein the original Intersection-over-Union is a ratio between intersection area and union area of the first candidate box and the second candidate box.

[0181] The calculation formula of the original Intersection-over-Union IOU is as follows:
¨ area(M) n area(Si) IOU
area(M) U area(Si) where M indicates the first candidate box, S, indicates the second candidate box, area(M) n area(Si) is the intersection area of the two candidate boxes, and area(M) U area(Si) is the union area of the two candidate boxes.

[0182] The loss parameter is employed to correct the original Intersection-over-Union, and a corrected Intersection-over-Union is obtained.

[0183] The corrected Intersection-over-Union is a difference between the original Intersection-over-Union and the loss parameter, and the calculation formula of the corrected Intersection-over-Union IOUnew is specifically as follows:
10Une, = IOU ¨ C
where IOU is the original Intersection-over-Union, and C is the loss parameter.

[0184] A confidence level correcting module 804 is employed for calculating a corrected confidence level of the second candidate box according to the corrected Intersection-over-Union and the original confidence level of the second candidate box.

Date Recue/Date Received 2021-12-30

[0185] In one embodiment, the confidence level correcting module 804 is specifically employed for:

[0186] employing the attenuation weight of the second candidate box calculated and obtained according to the corrected Intersection-over-Union to correct the original confidence level of the second candidate box, and obtaining the corrected confidence level of the second candidate box.

[0187] The correction of the original confidence level makes use of Gaussian weight coefficients and adds loss parameters to further increase punishment in terms of height, with the specific calculation formula as follows:
newscorei = scorei x where newscore, is the original confidence level after screening by the traditional non maximum suppression algorithm, score, is an original confidence level, and f(10Une,(M,Si)) is an attenuation weight coefficient relevant to the corrected iounew2 Intersection-over-Union, specifically, f(10Une,(M,Si)) = e 0-0¨c) , where a is a standard deviation parameter of a Gaussian function.

[0188] A to-be-recognized text box obtaining module 805 is employed for judging whether the corrected confidence level of the second candidate box satisfies a confidence level condition, if yes, taking the first candidate box and the second candidate box as text boxes to be recognized.

[0189] A recognizing module 806 is employed for recognizing text information in the text boxes to be recognized.

[0190] The recognizing module 806 employs a CRNN model with a CNN+RNN+CTC
framework. The convolutional layer of the CNN network includes alternately connected standard convolutional kernels and dilated convolutional kernels, and a width of a receptive field of the dilated convolutional kernel is larger relative to a width of a Date Recue/Date Received 2021-12-30 receptive field of the standard convolutional kernel. The block window of the pooling layer is modified from a square into a rectangle with unequal lengths and widths, and is mixedly pooled by standard maximum pooling and average pool weighting, and a pooling weight coefficient is calculated and determined according to a global maximum value and a global mean value of a block image.

[0191] Specifically, the calculation formula of the mixed value mixed (x) of mixed pooling is as follows:
mixed (x) = w x max (x) + (1 ¨ w)mean(x) where w is the pooling weight coefficient, max(x) is the global maximum value, and mean(x) is the mean value; specifically, the calculation formula of w is as follows:
abs(max(x)) W =
abs(max(x)) + abs(mean(x))

[0192] As shown in Fig. 9, on the basis of the text information extracting method provided by the embodiment of the present invention, an embodiment of the present invention further provides a text information extracting device that comprises:

[0193] a text information recognizing module 901, for recognizing text information by means of the text information recognizing method disclosed in the above embodiment;

[0194] a matching module 902, for matching the text information with a pre-created extended warranty synonym dictionary, the extended warranty synonym dictionary containing extended warranty keywords and synonyms;

[0195] a filtering module 903, for performing synonym replacement to the extended warranty keywords contained in the text information after matching has succeeded, and deleting text information containing no extended warranty keyword; and

[0196] an extracting module 904, for extracting the text information having been performed with synonym replacement.

[0197] In one embodiment, the text information extracting device disclosed by the embodiment of the present invention further comprises:

Date Recue/Date Received 2021-12-30

[0198] an extended warranty synonym dictionary updating module, for updating the extended warranty synonym dictionary, and specifically for:

[0199] determining extended warranty keywords;

[0200] extracting sample text information associated with the extended warranty keywords from a sample image, and employing a word segmentation tool to perform word segmentation on the sample text information;

[0201] judging whether the sample text information contains any invalid word, if yes, deleting the sample text information, if not, performing similarity calculation on the sample text information and vocabulary in an extended warranty synonym database; and

[0202] determining according to a similarity calculation result whether vocabulary in the sample text information is synonymous with the extended warranty keywords, if yes, adding the synonyms to the extended warranty synonym dictionary.

[0203] In one embodiment, the text information extracting device disclosed by the embodiment of the present invention further comprises:

[0204] an information pushing determining module, for:

[0205] determining a first extended warranty time limit according to the extended warranty text information, comparing the first extended warranty time limit with a second extended warranty time limit, and obtaining a time limit difference; comparing the time limit difference with a first time limit threshold condition, determining a corresponding commodity code for the extended warranty text information that satisfies the first time limit threshold condition, and determining customer communication information according to the commodity code; and

[0206] sending to a customer extended warranty pushing information corresponding to a comparison result of the time limit difference according to the customer communication information.

[0207] In one embodiment, the text information extracting device disclosed by the embodiment of the present invention further comprises:
Date Recue/Date Received 2021-12-30

[0208] an information display determining module, for:

[0209] determining a first extended warranty time limit according to the extended warranty text information, comparing the first extended warranty time limit with a second extended warranty time limit, and obtaining a time limit difference; and

[0210] comparing the time limit difference with a second time limit threshold condition, and determining a display location of extended warranty text information, with which the second extended warranty time limit associates, on a webpage according to a comparison result of the time limit difference.

[0211] On the basis of the aforementioned text information recognizing method, the present invention further provides a computer system that comprises:

[0212] one or more processor(s); and

[0213] a memory, associated with the one or more processor(s), wherein the memory is employed to store a program instruction, and the program instruction executes the aforementioned text information recognizing method when it is read and executed by the one or more processor(s).

[0214] Fig. 10 exemplarily illustrates the framework of the computer system that can specifically include a processor 1010, a video display adapter 1011, a magnetic disk driver 1012, an input/output interface 1013, a network interface 1014, and a memory 1020. The processor 1010, the video display adapter 1011, the magnetic disk driver 1012, the input/output interface 1013, the network interface 1014, and the memory 1020 can be communicably connected with one another via a communication bus 1030.

[0215] The processor 1010 can be embodied as a general CPU (Central Processing Unit), a microprocessor, an ASIC (Application Specific Integrated Circuit), or one or more integrated circuit(s) for executing relevant program(s) to realize the technical solutions provided by the present application.

Date Recue/Date Received 2021-12-30

[0216] The memory 1020 can be embodied in such a form as an ROM (Read Only Memory), an RAM (Random Access Memory), a static storage device, or a dynamic storage device.
The memory 1020 can store an operating system 1021 for controlling the running of an electronic equipment 1000, and a basic input/output system 1022 (BIOS) for controlling lower-level operations of the electronic equipment 1000. In addition, the memory 1020 can also store a web browser 1023, a data storage management system 1024, and an equipment identification information processing system 1025, etc. The equipment identification information processing system 1025 can be an application program that specifically realizes the aforementioned various step operations in the embodiments of the present application. To sum it up, when the technical solutions provided by the present application are to be realized via software or firmware, the relevant program codes are stored in the memory 1020, and invoked and executed by the processor 1010.

[0217] The input/output interface 1013 is employed to connect with an input/output module to realize input and output of information. The input/output module can be equipped in the device as a component part (not shown in the drawings), and can also be externally connected with the device to provide corresponding functions. The input means can include a keyboard, a mouse, a touch screen, a microphone, and various sensors etc., and the output means can include a display screen, a loudspeaker, a vibrator, an indicator light etc.

[0218] The network interface 1014 is employed to connect to a communication module (not shown in the drawings) to realize intercommunication between the current device and other devices. The communication module can realize communication in a wired mode (via USB, network cable, for example) or in a wireless mode (via mobile network, WIFI, Bluetooth, etc.).

[0219] The bus 1030 includes a passageway transmitting information between various component parts of the device (such as the processor 1010, the video display adapter 1011, Date Recue/Date Received 2021-12-30 the magnetic disk driver 1012, the input/output interface 1013, the network interface 1014, and the memory 1020).

[0220] Additionally, the electronic equipment 1000 may further obtain information of specific collection conditions from a virtual resource object collection condition information database for judgment on conditions, and so on.

[0221] As should be noted, although merely the processor 1010, the video display adapter 1011, the magnetic disk driver 1012, the input/output interface 1013, the network interface 1014, the memory 1020, and the bus 1030 are illustrated for the aforementioned device, the device may further include other component parts prerequisite for realizing normal running during specific implementation. In addition, as can be understood by persons skilled in the art, the aforementioned device may as well only include component parts necessary for realizing the solutions of the present application, without including the entire component parts as illustrated.

[0222] As can be known through the description to the aforementioned embodiments, it is clearly learnt by person skilled in the art that the present application can be realized through software plus a general hardware platform. Based on such understanding, the technical solutions of the present application, or the contributions made thereby over the state of the art, can be essentially embodied in the form of a software product, and such a computer software product can be stored in a storage medium, such as an ROM/RAM, a magnetic disk, an optical disk etc., and includes plural instructions enabling a computer equipment (such as a personal computer, a server, or a network device etc.) to execute the methods described in various embodiments or some sections of the embodiments of the present application.

[0223] The various embodiments are progressively described in the Description, identical or similar sections among the various embodiments can be inferred from one another, and Date Recue/Date Received 2021-12-30 each embodiment stresses what is different from other embodiments.
Particularly, with respect to the system or system embodiment, since it is essentially similar to the method embodiment, its description is relatively simple, and the relevant sections thereof can be inferred from the corresponding sections of the method embodiment. The system or system embodiment as described above is merely exemplary in nature, units therein described as separate parts can be or may not be physically separate, parts displayed as units can be or may not be physical units, that is to say, they can be located in a single site, or distributed over a plurality of network units. It is possible to base on practical requirements to select partial modules or the entire modules to realize the objectives of the embodied solutions. It is understandable and implementable by persons ordinarily skilled in the art without spending creative effort in the process.

[0224] Technical solutions provided by the embodiments of the present invention bring about the following advantageous effects.
1. In the text information recognizing method and device disclosed by the embodiments of the present invention, calculation of sizes of the intersection area and the closure area of the candidate box with the greatest original confidence level and another candidate box is introduced to obtain a loss parameter, the Intersection-over-Union between the candidate box with the greatest original confidence level and the other candidate box is corrected, the corrected Intersection-over-Union is utilized to calculate an attenuation weight coefficient, and the attenuation weight coefficient is utilized to correct the original confidence level, so the method and device are applicable to the extraction of text information in images with relatively small text line spacing, and effectively prevent missed detection of text lines.
2. The text information extracting method and device disclosed by the embodiments of the present invention effectively and precisely recognize extended warranty Date Recue/Date Received 2021-12-30 text information, and make use of the extended warranty text information to precisely push extended warranty information and determine its display location on the page, whereby the effect for marketing products with extended warranty is enhanced.

[0225] All the above optional technical solutions can be randomly combined to form optional embodiments of the present invention, to which no repetition is made thereto in this context.

[0226] What is described above is merely directed to preferred embodiments of the present invention, and is not meant to restrict the present invention. Any amendment, equivalent replacement and improvement makeable within the spirit and principle of the present invention shall all fall within the protection scope of the present invention.
Date Recue/Date Received 2021-12-30

Claims (10)

What is claimed is:

1. A text information recognizing method, characterized in comprising:
performing text detection on an image, and obtaining candidate boxes identifying text line locations in the image, and original confidence levels to which the various candidate boxes correspond;
selecting the candidate box with the greatest original confidence level from the candidate boxes sharing intersection areas to serve as a first candidate box, selecting any other candidate box to serve as a second candidate box, and calculating a loss parameter of the second candidate box according to a size of the intersection area between the first candidate box and the second candidate box and a size of a closure area;
calculating an original Intersection-over-Union between the first candidate box and the second candidate box, correcting the original Intersection-over-Union according to the loss parameter of the second candidate box, and obtaining a corrected Intersection-over-Union;
calculating a corrected confidence level of the second candidate box according to the corrected Intersection-over-Union and the original confidence level of the second candidate box;
judging whether the corrected confidence level of the second candidate box satisfies a confidence level condition, if yes, taking the first candidate box and the second candidate box as text boxes to be recognized; and recognizing text information in the text boxes to be recognized.

2. The method according to Claim 1, characterized in that the step of calculating a loss parameter of the second candidate box according to a size of the intersection area between the first candidate box and the second candidate box and a size of a closure area includes:
obtaining a width and a height of the intersection area, and obtaining a width and a height of the closure area; and calculating the loss parameter of the second candidate box according to a height ratio of the intersection area to the closure area and a width ratio of the intersection area to the closure area.

3. The method according to Claim 1, characterized in that the step of calculating a corrected Date Recue/Date Received 2021-12-30 confidence level of the second candidate box according to the corrected Intersection-over-Union and the original confidence level of the second candidate box includes:
calculating an attenuation weight of the second candidate box according to the corrected Intersection-over-Union; and employing the attenuation weight of the second candidate box to correct the original confidence level of the second candidate box, and obtaining the corrected confidence level of the second candidate box.

4. The method according to any of Claims 1 to 3, characterized in that the step of recognizing text information in the text boxes to be recognized includes:
using a neutral network model to recognize text information in the text boxes to be recognized, the neutral network model including a convolutional layer and a pooling layer;
wherein the convolutional layer includes alternately connected standard convolutional kernels and dilated convolutional kernels, and a width of a receptive field of the dilated convolutional kernel is larger relative to a width of a receptive field of the standard convolutional kernel;
wherein the pooling layer has rectangular block windows, and adopts weighting mix pooling of standard maximum pooling and average pool, and a pooling weight coefficient is calculated and determined according to a global maximum value and mean value of block images.

5. A text information extracting method, characterized in comprising:
employing the method according to any of Claims 1 to 4 to recognize text information in a commodity image;
matching the text information with a pre-created extended warranty synonym dictionary, the extended warranty synonym dictionary containing extended warranty keywords and synonyms;
performing synonym replacement to the extended warranty keywords contained in the text information after matching has succeeded; and extracting the text information having been performed with the synonym replacement.

6. The method according to Claim 5, characterized in further comprising:
determining a first extended warranty time limit according to the extended warranty text information, comparing the first extended warranty time limit with a second extended warranty time limit, and obtaining a time limit difference;
comparing the time limit difference with a first time limit threshold condition, determining a Date Recue/Date Received 2021-12-30 corresponding commodity code for the extended warranty text information that satisfies the first time limit threshold condition, and determining customer communication information according to the commodity code; and sending to a customer extended warranty pushing information corresponding to a comparison result of the time limit difference according to the customer communication information.

7. The method according to Claim 5, characterized in further comprising:
determining a first extended warranty time limit according to the extended warranty text information, comparing the first extended warranty time limit with a second extended warranty time limit, and obtaining a time limit difference;
comparing the time limit difference with a second time limit threshold condition, and determining a display location of extended warranty text information, with which the second extended warranty time limit associates, on a webpage according to a comparison result of the time limit difference.

8. A text information recognizing device, characterized in comprising:
a detecting module, for performing text detection on an image, and obtaining candidate boxes identifying text line locations in the image, and original confidence levels to which the various candidate boxes correspond;
a loss parameter calculating module, for selecting the candidate box with the greatest original confidence level from the candidate boxes sharing intersection areas to serve as a first candidate box, selecting any other candidate box to serve as a second candidate box, and calculating a loss parameter of the second candidate box according to a size of the intersection area between the first candidate box and the second candidate box and a size of a closure area;
an Intersection-over-Union correcting module, for calculating an original Intersection-over-Union between the first candidate box and the second candidate box, correcting the original Intersection-over-Union according to the loss parameter of the second candidate box, and obtaining a corrected Intersection-over-Union;
a confidence level correcting module, for calculating a corrected confidence level of the second candidate box according to the corrected Intersection-over-Union and the original confidence level of the second candidate box;
a to-be-recognized text box obtaining module, for judging whether the corrected confidence level of the second candidate box satisfies a confidence level condition, if yes, taking the first Date Recue/Date Received 2021-12-30 candidate box and the second candidate box as text boxes to be recognized; and a recognizing module, for recognizing text information in the text boxes to be recognized.

9. A text information extracting device, characterized in comprising:
a text information recognizing module, for executing the method according to any of Claims 1 to 4 to recognize text information;
a matching module, for matching the text information with a pre-created extended warranty synonym dictionary, the extended warranty synonym dictionary containing extended warranty keywords and synonyms;
a filtering module, for performing synonym replacement to the extended warranty keywords contained in the text information after matching has succeeded; and an extracting module, for extracting the text information having been performed with synonym replacement.

10. A computer system, characterized in comprising:
one or more processor(s); and a memory, associated with the one or more processor(s), wherein the memory is employed to store a program instruction, and the program instruction executes the method according to any of Claims 1 to 4 when it is read and executed by the one or more processor(s).

Date Recue/Date Received 2021-12-30