A method for automatically cataloging books

Abstract

The present application relates to the technical field of book cataloging, and particularly relates to a book automatic cataloging method, comprising: generating a multi-row cataloging text box image of a book cover image through improved YOLO; generating name text and description text through an OCR model from the cataloging text box image; classifying the description text to generate title text and synopsis text; if the text length of the title text is greater than the lower limit of the text length, then generating a book classification label through a title classification model based on an improved clustering algorithm by clustering calculation of the title text and the classified title in a book management library; if the text length of the title text is less than or equal to the lower limit of the text length, then generating a book classification label through a text classification model based on an improved neural network from the synopsis text; and inputting the text and the book classification label into a library automation system. The present application avoids the interference of background patterns or decorative text on OCR, makes the cataloging and classification result of the book more accurate, and realizes accurate cataloging and management of the book.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of book cataloging, and particularly relates to a book automatic cataloging method. BACKGROUND

[0002] Book cataloging is the process of entering book information into a system, which is a crucial link in the automated book management system of a library, and the quality of the entered information directly determines the accuracy and completeness of the library's collection information. Therefore, most libraries currently use optical character recognition (OCR) and natural language processing (NLP) technologies to intelligently extract and classify the entered information, in order to achieve the efficiency and accuracy of book automatic cataloging.

[0003] However, there are the following problems in the implementation process of book automatic cataloging: the background pattern of the book cover is relatively complex and rich, and there are decorative texts, which leads to inaccurate OCR recognition of the title text, or the decorative text is recognized as the title, resulting in incorrect entered information; the length of the title text of the book is too short, which leads to inaccurate classification, for example, the book with the title "Memory", NLP classifies the semantic classification by vectorizing the title text, and since the feature vector dimension generated by less than six characters is too low, the NLP designed for large dimension vector data cannot accurately classify it.

[0004] For example, the patent document with application number 202311551618.2 discloses a book automatic cataloging method, device and medium for a cataloging robot, but the cataloging text is not intelligently recognized and extracted in the automatic cataloging process, and the title of the short text cannot be classified.

[0005] Therefore, how to realize intelligent recognition and extraction of cataloging text and classification of short text titles to achieve high accuracy of book automatic cataloging is a technical problem to be solved at present. SUMMARY

[0006] Therefore, the present application provides a book automatic cataloging method, which improves the text detection model of YOLO to realize accurate positioning and recognition of the cataloging text area, avoids the interference of background patterns or decorative texts on OCR, and also makes the position information of the text area available for classification and cataloging judgment of the text. The clustering algorithm is used to cluster and calculate the longer title text, or the improved neural network is used to recognize the semantic of the introduction text, so that the classification and cataloging result of the book is more accurate, and accurate cataloging management of the book is realized.

[0007] To achieve the above purpose, the present application provides a book automatic cataloging method, comprising:

[0008] The book cover image is converted into a multi-line catalog text box image by an improved YOLO-based text detection model, wherein the text detection model is embedded with an improved ASPP module;

[0009] The catalog text box image is converted into name text and description text by an OCR model;

[0010] According to the positional relationship of the name text and the description text, as well as the name text area of the name text and the title text area of the description text, title text and introduction text are classified from the description text;

[0011] If the text length of the title text is greater than the lower limit of the text length, the title text and the classified title of the book management library are clustered by a title classification model based on an improved clustering algorithm to generate a book classification label;

[0012] If the text length of the title text is less than or equal to the lower limit of the text length, the introduction text is converted into the book classification label by a text classification model based on an improved neural network;

[0013] The name text, the introduction text, the title text and the book classification label are input into the library automation system.

[0014] Further, the branch of the improved ASPP module is embedded with a coordinate attention mechanism, wherein the ASPP is used to blur the catalog text box image positioning of the image features obtained at different sampling rates through the coordinate attention mechanism.

[0015] Further, the improved ASPP module has a first branch to a fifth branch, a feature splicing layer and a feature convolution output layer, wherein the first branch is provided with a first size convolution layer and the coordinate attention mechanism, the second branch to the fourth branch are respectively provided with a second size convolution layer and the coordinate attention mechanism, the fifth branch is provided with a pooling layer and a deconvolution layer, the convolution kernel of the first size convolution layer is smaller than the convolution kernel of the second size convolution layer, the outputs of the first branch to the fifth branch are connected with the feature splicing layer, and the output of the feature splicing layer is connected with the feature convolution output layer;

[0016] The image features obtained at different sampling rates are positioned in the blurred catalog text box image by the first branch to the fourth branch;

[0017] The positioning is spliced into a positioning splicing feature by the feature splicing layer;

[0018] The positioning splicing feature is extracted and output by the feature convolution output layer.

[0019] Further, the text detection model comprises a smooth label sub-model, an improved loss function and a decoupling head, so as to perform inclined frame selection on the catalog text box diagram, and the process of outputting the catalog text box diagram by the decoupling head comprises:

[0020] The decoupling head classifies the rotation angle of the catalog information box by the smooth label sub-model to generate an angle label value;

[0021] The decoupling head learns the inclined angle of the real frame by the improved loss function, wherein the improved loss function is constructed based on the angle label value and a cross-entropy function.

[0022] In the above scheme, the ASPP module embedded with the coordinate attention mechanism makes the YOLO model pay more attention to the accurate position information of different sampling rates of the image, and the ASPP module and the improved loss function make the YOLO model capable of detecting the catalog text with an inclined design and generating an inclined frame diagram, thereby avoiding the OCR from recognizing the inclined text as different lines of text.

[0023] Further, the clustering algorithm is a Kmeans clustering algorithm, and the process of generating a book classification label by clustering calculation of the title text and the classified title based on the improved clustering algorithm title classification model comprises:

[0024] The title text and the classified title are generated into a clustering cluster comprising all text vectors by a word embedding model;

[0025] The clustering cluster is divided into two sub-clustering clusters according to the average value of all text vectors in the clustering cluster, the sub-clustering cluster that reduces the total error is retained, and the sub-clustering cluster that cannot reduce the total error is removed;

[0026] The retained sub-clustering cluster is subjected to a binary division of average value calculation until the number of clustering clusters reaches a set value, and the book classification label is output.

[0027] Further, if the text vector corresponding to the classified title is less than the product of the average value and a weight value, the classified title is removed.

[0028] In the above scheme, the Kmeans clustering algorithm with binary division is used to perform clustering calculation on the longer title text, which overcomes the problem that the Kmeans clustering algorithm is prone to converge to a global minimum value due to insufficient semantic information of the title text itself, and realizes accurate classification of books based on the semantic information of the title text.

[0029] Further, the process of generating a book classification label by the introduction text based on the improved neural network text classification model comprises:

[0030] extracting keywords from the synopsis text;

[0031] generating keyword vectors from the keywords through a word embedding model;

[0032] generating the book classification label based on a neural network module and a bidirectional long short-term memory network module from the keyword vectors.

[0033] Further, the process of generating a book classification label based on a neural network module and a bidirectional long short-term memory network module from keyword vectors includes:

[0034] generating first and second vectors from the keyword vectors through the neural network module and the bidirectional long short-term memory network module, respectively;

[0035] performing weighted summation of the first and second vectors to generate a comprehensive feature vector;

[0036] generating the book classification label from the comprehensive feature vector through a fully connected layer and an activation function in sequence.

[0037] In the above scheme, the correlation between the words before and after the synopsis text is extracted and learned through the neural network module and the bidirectional long short-term memory network module, and better text classification effect of the synopsis text is achieved.

[0038] Further, the process of generating a title text and a synopsis text from a description text based on the positional relationship between the name text and the description text and the area of the name text and the area of the title text of the description text includes:

[0039] dividing the description text into multiple line texts and calculating the height ratio between the multiple line texts;

[0040] if the height ratio is greater than a ratio threshold, determining the line text with the highest height as the title text and the remaining line texts as the synopsis text;

[0041] if the height ratio is less than or equal to the ratio threshold, determining the title text and the synopsis text according to the weighted calculation result of the height ratio, the center point distance between the name text and the description text.

[0042] Further, the catalog text box diagram includes a name text box diagram, a description text box diagram, a price box diagram, a publisher box diagram, a publication date box diagram, a library location box diagram, and a call number box diagram.

[0043] The catalog text box diagram is further converted by an OCR model to generate other information texts.

[0044] Extract keywords from the other information text to generate description text, price, publisher, publication date, library location and call number;

[0045] Extract keywords from the name text to generate author name, translator name and editor name.

[0046] In the above scheme, the book cover text is accurately classified according to position information and size information.

[0047] Compared with the prior art, the beneficial effects of the present application are that,

[0048] 1. By improving the text detection model of YOLO, accurate positioning and recognition of the cataloging text area are realized, the interference of background patterns or decorative text on OCR is avoided, and the position information of the text area can be used for classification and cataloging judgment of the text. Through clustering algorithm, long title text is clustered and calculated, or through improved neural network, semantic recognition of the introduction text is performed, so that the classification and cataloging result of the book is more accurate, and accurate cataloging management of the book is realized.

[0049] 2. By embedding the ASPP module of the coordinate attention mechanism, the YOLO model pays more attention to the accurate position information of different sampling rates of the image. Through the ASPP module and the improved loss function, the YOLO model can detect the cataloging text designed to be inclined, generate an inclined frame diagram, and thus avoid the recognition of the inclined text as different lines of text by OCR.

[0050] 3. The Kmeans clustering algorithm is used to cluster and calculate the long title text, which overcomes the problem that the semantic information of the title text itself is not rich enough, leading to the Kmeans clustering algorithm converging to the global minimum value, and realizes accurate classification of the book based on the semantic information of the title text.

[0051] 4. The correlation between the front and rear words of the introduction text is extracted and learned through the neural network module and the bidirectional long short-term memory network module, so as to achieve better text classification effect of the introduction text. BRIEF DESCRIPTION OF DRAWINGS

[0052] Figure 1 The flowchart of the book automatic cataloging method of the embodiment of the present application;

[0053] Figure 2 The structure diagram of the improved YOLO text detection model of the book automatic cataloging method of the embodiment of the present application;

[0054] Figure 3 The structure diagram of the improved ASPP module of the improved YOLO text detection model of the book automatic cataloging method of the embodiment of the present application;

[0055] Figure 4 The structural schematic diagram of the text classification model of the improved neural network of the book automatic cataloging method is shown. DETAILED DESCRIPTION

[0056] In order to make the objects and advantages of the present application clearer, the present application will be further described below in conjunction with embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and should not be used to limit the present application.

[0057] The preferred embodiments of the present application will be described below with reference to the accompanying drawings. It should be understood by those skilled in the art that these embodiments are only used to explain the technical principles of the present application and are not used to limit the protection scope of the present application.

[0058] It should be noted that, in the description of the present application, the terms of direction or position relationship such as 'upper', 'lower', 'left', 'right', 'inner', 'outer' and the like are based on the direction or position relationship shown in the drawings, which is only for the convenience of description and does not indicate or imply that the device or element must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present application.

[0059] In addition, it should also be noted that, in the description of the present application, unless otherwise explicitly specified and limited, the terms of'mounting', 'connecting', 'connection' should be understood in a broad sense, for example, it can be fixed connection, or detachable connection, or integral connection; it can be mechanical connection, or electrical connection; it can be direct connection, or indirect connection through an intermediate medium, or the internal communication of two elements. Those skilled in the art can understand the specific meaning of the above terms in the present application according to the specific circumstances.

[0060] As shown in Figures 1 to 4 The present application provides a book automatic cataloging method, which realizes accurate positioning and recognition of the cataloging text area by improving the text detection model of YOLO, avoids the interference of background patterns or decorative text on OCR, and also makes the position information of the text area available for the classification and cataloging judgment of the text. The clustering algorithm is used to cluster and calculate the long title text, or the improved neural network is used to perform semantic recognition on the introduction text, so that the classification and cataloging result of the book is more accurate, and accurate cataloging management of the book is realized.

[0061] As shown in Figures 1 to 4 The present embodiment provides a book automatic cataloging method, which comprises:

[0062] generating a multi-line catalog text box diagram from a book cover image through a text detection model based on an improved YOLO (You Only Look Once), wherein the text detection model is embedded with an improved ASPP (Atrous Spatial Pyramid Pooling) module for fuzzy catalog text box diagram positioning;

[0063] converting the catalog text box diagram through an OCR (Optical Character Recognition) model to generate name text and description text;

[0064] classifying the description text according to the positional relationship of the name text and the description text, as well as the name text area of the name text and the title text area of the description text to generate title text and synopsis text;

[0065] if the text length of the title text is greater than the lower limit of the text length, then the title text and the classified title of the book management library are clustered through a title classification model based on an improved clustering algorithm to generate a book classification label;

[0066] if the text length of the title text is less than or equal to the lower limit of the text length, then the synopsis text is input into a text classification model based on an improved neural network to generate the book classification label;

[0067] inputting the name text, the synopsis text, the title text, and the book classification label into a library automation system.

[0068] It can be understood that the YOLO with the improved ASPP module realizes the processing of fuzzy and damaged book information, ensures the integrity of the catalog information, and can learn and output the text distribution characteristics of the book cover to generate the title text and the synopsis text according to the text distribution characteristics. The description text is longer than other information on the book cover, including the title text and the synopsis text, the title text includes the book title and the sub-title, and the synopsis text is the content of the book. Compared with other information on the book cover, such as price, author, translator, editor, publisher, publication date, library location, and call number, it is difficult to classify and extract through characteristic words such as "author" and "translator", so the frame diagram generated by the YOLO model is used for auxiliary classification to improve the accuracy of the classification.

[0069] Further, as shown in Figure 3 the branch of the improved ASPP module is embedded with a coordinate attention mechanism, wherein the ASPP is used to make image features obtained at different sampling rates pass through the coordinate attention mechanism for fuzzy catalog text box diagram positioning.

[0070] It can be understood that the YOLO is preferably a YOLOv5 model as shown in Figure 2 Although the YOLOv5 model has high performance for target detection, the bounding box output by the YOLOv5 model itself is a rectangular box without a rotation angle and parallel to each other, which results in low accuracy when the bounding box output by the YOLOv5 model has an inclined angle. Therefore, the ASPP module can accurately locate the fuzzy and damaged text box image and output a high-accuracy inclined-angle bounding box.

[0071] It can be understood that the coordinate attention module (Coord i nate Attent ion) uses global average pooling to process the input feature map in width, height and rotation angle, and encodes each channel. This method can make the model pay more attention to the accurate position information of the image, and can obtain attention in the width, height and rotation angle of the image. Therefore, the coordinate attention module can improve the accuracy of the model.

[0072] Further, as shown in Figure 3 The improved ASPP module has a first branch to a fifth branch, a feature splicing layer and a feature convolution output layer, wherein the first branch is provided with a first size convolution layer and the coordinate attention mechanism, the second branch to the fourth branch are respectively provided with a second size convolution layer and the coordinate attention mechanism, the fifth branch is provided with a pooling layer and a deconvolution layer, the convolution kernel of the first size convolution layer is smaller than the convolution kernel of the second size convolution layer, the outputs of the first branch to the fifth branch are connected with the feature splicing layer, and the output of the feature splicing layer is connected with the feature convolution output layer.

[0073] The first branch to the fourth branch are used to locate the fuzzy catalog text box image by the image features obtained by different sampling rates; the feature splicing layer is used to splice the location to generate a location splicing feature; and the feature convolution output layer is used to extract and output the location splicing feature.

[0074] Specifically, as shown in Figure 3 The convolution kernel of the first size convolution layer is 1x1, the convolution kernel of the second size convolution layer is 3x3, the dilation rates of the convolution layers of the first branch to the fourth branch are 1, 6, 12 and 18 in sequence, and the convolution kernel of the deconvolution layer of the fifth branch and the convolution kernel of the feature convolution output layer are 1x1.

[0075] Further, as shown in Figure 2 The text detection model includes a smooth label sub-model, an improved loss function and a decoupling head, which are used to select an inclined box for the catalog text box image, and the process of the decoupling head for outputting the catalog text box image includes:

[0076] The decoupling head classifies the rotation angle of the catalog information box through the smooth label sub-model to generate an angle label value;

[0077] The decoupling head learns the tilt angle of the real box through an improved loss function, wherein the improved loss function is constructed based on the angle label value and a cross-entropy function.

[0078] Specifically, the smooth label sub-model is preferably a circular smooth label (CSL), which converts the tilt angle regression problem into a classification problem for processing.

[0079] Specifically, the improved loss function is:

[0080]

[0081] In the formula, loss(z, y) represents the improved loss function, y i,j represents the true label value of the i-th sample at the j-th angle in the circular smooth label, δ represents a sigmoid activation function, z i,j represents the predicted label value of the i-th sample at the j-th angle in the circular smooth label. The formula sums the true label value and the predicted label value of each angle after logarithmic calculation to generate the loss function.

[0082] Preferably, the YOLOv5 model has multiple decoupling heads, and the improved loss function is only used in one of the decoupling heads, and the other decoupling heads can output normal box diagrams.

[0083] In the above scheme, the ASPP module embedded with the coordinate attention mechanism makes the YOLO model pay more attention to the accurate position information of different sampling rates of the image. Through the ASPP module and the improved loss function, the YOLO model can detect the tilt designed catalog text and generate a tilt box diagram, thereby avoiding the OCR from recognizing the tilt text as different lines of text.

[0084] Further, the clustering algorithm is a Kmeans clustering algorithm, and the process of generating a book classification label by clustering calculation of the title text and the classified title of the library management library through a title classification model based on the improved clustering algorithm includes:

[0085] Generating a clustering cluster including all text vectors from the title text and the classified title through a word embedding model;

[0086] Dividing the clustering cluster into two sub-clustering clusters according to the average value of all text vectors in the clustering cluster, retaining the sub-clustering cluster that reduces the total error, and removing the sub-clustering cluster that cannot reduce the total error;

[0087] The reserved sub-cluster is subjected to a two-division average value calculation until the number of cluster clusters reaches a set value, and the book classification label is output.

[0088] Specifically, the word embedding model is preferably a Word2Vec model, which is used to map words to a high-dimensional vector space. The Kmeans clustering algorithm sets the number of cluster clusters K to be 8, i.e. eight categories of literature, science, history and geography, philosophy, economics and politics, medicine, education, and industry and technology, then randomly selects 8 initial points as cluster centers, assigns the title text and the classified title to the cluster center closest to the cluster center, then updates the cluster center of each cluster to the average value of all points in the cluster, and performs two-division, thereby overcoming the problem that the kmeans algorithm easily converges to a global minimum value.

[0089] The total error before the two-division is compared with the errors of the two sub-cluster clusters after the two-division to determine whether the sub-cluster clusters reduce the total error.

[0090] Further, if the text vector corresponding to the classified title is less than the product of the average value and the weight value, the classified title is removed, which is calculated by the following formula:

[0091] d(p i )<ηMed(P)

[0092] In the formula, d(p i ) is the text vector corresponding to the classified title i, η is the weight value, preferably 1.7, and Med(P) is the average value. The text vector is the distance from the classified title i to the cluster center, and Med(P) is the average of the distances from all classified titles i in the cluster to the cluster center. Therefore, the classified title that is too far away from the cluster center and is not conducive to classification can be removed, thereby making the Kmeans clustering algorithm have better clustering and classification effect on short text title texts.

[0093] In the above scheme, the Kmeans clustering algorithm based on two-division is used to cluster and calculate longer title texts, which overcomes the problem that the semantic information of the title text itself is not rich enough to cause the Kmeans clustering algorithm to easily converge to a global minimum value, and realizes accurate classification of books based on the semantic information of the title text.

[0094] Further, as shown in Figure 4 the process of generating a book classification label from the brief text by the text classification model based on the improved neural network includes:

[0095] extracting keywords from the brief text;

[0096] generating a keyword vector through a word embedding model;

[0097] generating the book classification label based on a neural network module and a bidirectional long short-term memory (BiLSTM) module.

[0098] Preferably, the keywords with commonalities are extracted from the synopsis text through TextRank, and the word embedding model is a Word2Vec model.

[0099] Further, as shown in Figure 4 the process of generating a book classification label based on a keyword vector through a neural network module and a bidirectional long short-term memory (BiLSTM) module includes:

[0100] generating a first vector and a second vector through the neural network module and the bidirectional long short-term memory module, respectively, from the keyword vector;

[0101] performing weighted summation on the first vector and the second vector to generate a comprehensive feature vector;

[0102] generating the book classification label through a fully connected layer and an activation function in sequence from the comprehensive feature vector.

[0103] Preferably, the activation function is Softmax.

[0104] Specifically, since the synopsis text may involve information of other labels, the neural network module may extract features of other labels, thereby affecting the classification result. According to the different degrees of dependence of different texts on local semantic features, a gating mechanism is introduced to assign weights to the significant features captured by the neural network module and the context features captured by the bidirectional long short-term memory module, and then fuse the two kinds of features according to the weights to improve the accuracy of text classification. The calculation formula of the gating mechanism is:

[0105] a1 = δ (W·T cnn +b)

[0106] T text = a1T cnn +(1-a1)T Bi

[0107] In the formula, a1 is a weight coefficient, δ represents a sigmoid activation function, W represents a weight matrix, b is a bias term, T text , T cnn , T Bi are the comprehensive feature vector, the first vector and the second vector, respectively.

[0108] In the scheme, the correlation between the front and rear words of the brief introduction text is extracted and learned through the neural network module and the bidirectional long short-term memory network module, and better text classification effect of the brief introduction text is achieved.

[0109] Further, the process of classifying and generating the title text and the brief introduction text from the position relationship between the name text and the description text and the name text area of the name text and the title text area of the description text includes:

[0110] The description text is divided into multiple line texts, and the height ratio between the multiple line texts is calculated;

[0111] If the height ratio is greater than the ratio threshold, the line text with the highest height is determined as the title text, and the remaining line texts are determined as the brief introduction text.

[0112] Specifically, the ratio threshold is preferably 0.5, which conforms to the font size relationship of the title and the brief introduction.

[0113] If the height ratio is less than or equal to the ratio threshold, the title text and the brief introduction text are determined according to the weighted calculation result of the height ratio, the center point distance between the name text and the description text, specifically:

[0114]

[0115] In the formula, f represents the score result of the weighted calculation, 0.3 and 0.7 are the preferred weighted coefficients, p is the height ratio, d1 and d2 are the center point distances of the first part and the second part of the name text and the description text respectively. When the score result is greater than 1.3, the first part is the title text and the second part is the brief introduction text, otherwise, the opposite is true.

[0116] It can be understood that generally the author's name is closer to the title text and farther away from the brief introduction text, so the calculation result of the above formula can distinguish the title text and the brief introduction text.

[0117] Further, the catalog text box diagram includes a name text box diagram, a description text box diagram, a price box diagram, a publisher box diagram, a publication date box diagram, a library location box diagram, and a call number box diagram.

[0118] The catalog text box diagram is converted by the OCR model to further generate other information texts;

[0119] The key words of the other information texts are extracted to generate the description text, the price, the publisher, the publication date, the library location, and the call number.

[0120] The key words of the name text are extracted to generate the author's name, the translator's name, and the editor's name.

[0121] Specifically, the classification can be realized by matching the feature words of other information texts and name texts through TextRank, and the classification is input into the library automation system for management.

[0122] In the above scheme, the accurate classification of the book cover texts according to the position information and size information is realized.

[0123] In the embodiment, the accurate positioning and recognition of the cataloging text region are realized by improving the text detection model of YOLO, the interference of background patterns or decorative texts on OCR is avoided, and the position information of the text region can be used for the classification judgment of the text. The clustering algorithm is used for clustering calculation of the long title text, or the semantic recognition of the introduction text is performed through the improved neural network, so that the classification result of the book is more accurate, the accurate classification and management of the book are realized. Through the ASPP module embedded with the coordinate attention mechanism, the YOLO model pays more attention to the accurate position information of different sampling rates of the image. Through the ASPP module and the improved loss function, the YOLO model can detect the cataloging text designed to be inclined, generate an inclined frame diagram, and thus avoid the recognition of the inclined text as different lines of text by OCR. The Kmeans clustering algorithm is used for clustering calculation of the long title text, which overcomes the problem that the Kmeans clustering algorithm is easy to converge to the global minimum value due to the insufficient semantic information of the title text itself, and realizes the accurate classification of the book according to the semantic information of the title text. The correlation between the front and rear words of the introduction text is extracted and learned through the neural network module and the bidirectional long short-term memory network module, and better text classification effect of the introduction text is achieved.

[0124] So far, the technical solutions of the present application have been described in combination with the preferred embodiments shown in the drawings, but those skilled in the art can easily understand that the protection scope of the present application is obviously not limited to these specific embodiments. Those skilled in the art can make equivalent changes or replacements to the related technical features without departing from the principles of the present application, and the technical solutions after the changes or replacements will fall within the protection scope of the present application.

[0125] The above description is only the preferred embodiments of the present application and is not used to limit the present application; for those skilled in the art, the present application can have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A method of automatically cataloging books, characterized by, The method comprises the following steps: generating a multi-line catalog text box image of a book cover image through a text detection model based on improved YOLO, wherein the text detection model is embedded with an improved ASPP module; converting the catalog text box image through an OCR model to generate name text and description text; classifying the description text according to the positional relationship of the name text and the description text, and the name text area of the name text and the title text area of the description text to generate title text and introduction text; if the text length of the title text is greater than the lower limit of the text length, then the title text and the classified title of the book management library are clustered through a title classification model based on an improved clustering algorithm to generate a book classification label; if the text length of the title text is less than or equal to the lower limit of the text length, then the introduction text is input into a text classification model based on an improved neural network to generate the book classification label; inputting the name text, the introduction text, the title text and the book classification label into a library automation system; wherein the process of generating title text and introduction text comprises: dividing the description text into multiple line texts and calculating the height ratio between the multiple line texts; if the height ratio is greater than the ratio threshold, then the line text with the highest height is determined as the title text, and the remaining line texts are determined as the introduction text; if the height ratio is less than or equal to the ratio threshold, then the title text and the introduction text are determined according to the weighted calculation result of the height ratio and the center point distance of the name text and the description text.

2. The method of claim 1, wherein, The branch of the improved ASPP module is embedded with a coordinate attention mechanism, wherein the ASPP is used to blur the catalog text box image through the coordinate attention mechanism for positioning.

3. The method of claim 2, wherein, The improved ASPP module has a first branch to a fifth branch, a feature splicing layer and a feature convolution output layer, wherein the first branch is provided with a first size convolution layer and the coordinate attention mechanism, the second branch to the fourth branch are respectively provided with a second size convolution layer and the coordinate attention mechanism, the fifth branch is provided with a pooling layer and a deconvolution layer, the convolution kernel of the first size convolution layer is smaller than the convolution kernel of the second size convolution layer, the outputs of the first branch to the fifth branch are connected with the feature splicing layer, and the output of the feature splicing layer is connected with the feature convolution output layer; the first branch to the fourth branch are used to blur the catalog text box image for positioning through the image features obtained at different sampling rates; the feature splicing layer is used to perform feature splicing on the positioning to generate a positioning splicing feature; the feature convolution output layer is used to perform feature extraction and output on the positioning splicing feature.

4. The method of claim 2, wherein, The text detection model comprises a smoothing label sub-model, an improved loss function and a decoupling head, which are used to perform inclined frame selection on the catalog text box image, and the process of outputting the catalog text box image by the decoupling head comprises: the decoupling head classifies the rotation angle of the catalog text box image through the smoothing label sub-model to generate an angle label value; The decoupling head learns the tilt angle of the real box by improving the loss function, wherein the improved loss function is constructed based on the angle label value and a cross-entropy function.

5. The method of claim 1, wherein, The clustering algorithm is a Kmeans clustering algorithm, and the process of generating a book classification label by clustering the title text and the classified title of the book management library through a title classification model based on the improved clustering algorithm includes: Generating a clustering cluster including all text vectors by the title text and the classified title through a word embedding model; Dividing the clustering cluster into two sub-clustering clusters according to the average value of all text vectors in the clustering cluster, retaining the sub-clustering cluster that reduces the total error, and removing the sub-clustering cluster that cannot reduce the total error; Cyclically taking the average value of the retained sub-clustering cluster for two-partitioning until the number of clustering clusters reaches a set value, and outputting the book classification label.

6. The method of claim 5, wherein, If the text vector corresponding to the classified title is less than the product of the average value and the weight value, the classified title is removed.

7. The method of claim 1, wherein, The process of generating a book classification label by the synopsis text through a text classification model based on an improved neural network includes: Extracting keywords of the synopsis text; Generating a keyword vector by the keywords through a word embedding model; Generating the book classification label based on a neural network module and a bidirectional long short-term memory network module.

8. The method of claim 7, wherein, The process of generating a book classification label by a keyword vector based on a neural network module and a bidirectional long short-term memory network module includes: Generating a first vector and a second vector by the keyword vector through the neural network module and the bidirectional long short-term memory network module, respectively; Generating a comprehensive feature vector by weighted summation of the first vector and the second vector; Generating the book classification label by the comprehensive feature vector through a full connection layer and an activation function in sequence.

9. The method of claim 1 to 8, wherein, The cataloging text box diagram includes a name text box diagram, a description text box diagram, a price box diagram, a publisher box diagram, a publication date box diagram, a library location box diagram, and a call number box diagram; The cataloging text box diagram is converted by an OCR model to further generate other information text; Extracting keywords from the other information text to generate description text, price, publisher, publication date, library location, and call number; Extracting keywords from the name text to generate author name, translator name, and editor name.

Images