Animal information visual retrieval method and system based on machine learning
By establishing an event information text database and species picture database, and using natural language processing and deep learning models to correlate it, the problem of classification and retrieval of massive animal information is solved, and efficient visual retrieval of animal information is achieved.
Patent Information
- Application Number
- CN202411959422.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-30
- Publication Date
- 2025-04-25
AI Technical Summary
It is difficult for the prior art to efficiently classify and retrieve massive animal information, especially when looking for information that is valuable for species identification in picture resources, there are difficulties.
By establishing an event information text database and species image database, and using natural language processing and deep learning models to correlate it, visual retrieval of animal information is achieved.
It realizes intelligent retrieval of animal event information, can accurately and efficiently identify and track species information, reducing the pressure of artificial classification.
Smart Images
Figure CN119884359A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of data processing technology, relates to machine learning technology, and specifically to an animal information visualization retrieval method and system based on machine learning. Background Art
[0002] At present, the classification and management of animal information mainly relies on manual classification. With the rapid growth of the amount of information and the increase in the amount of image information, the daily workload of manual work is relatively fixed, and manual classification alone is insufficient. With the development of the digital age, a large number of symbolic resources such as images, audio, text, and video can be obtained through the Internet with the help of various information means, so as to carry out multimodal image recognition research. However, network resources are often scattered, chaotic, and unclear, making it difficult to find information with reference value for species identification in massive image resources. Therefore, the development of computer automatic classification retrieval is both accurate and efficient and can reduce manual pressure, but there is currently no suitable feasible solution. Summary of the invention
[0003] To solve the above problems, the present invention discloses an animal information visualization retrieval method and system based on machine learning, which associates the event-related information text database with the species image database, and uses computer recognition algorithms to meet the dual needs of species identification and information tracking.
[0004] To achieve the above object, the technical solution of the present invention is as follows:
[0005] The animal information visualization retrieval method based on machine learning includes the following steps:
[0006] Establish a database of event information and use Tableau to visualize the data; use pyhanlp entity recognition to refine the documents and visualize the number of animal events;
[0007] Step 2, collect data;
[0008] Step 3, using natural language processing technology (NLP), taking the animal event information text as the original corpus, purifying the text to make it a standardized text; selecting a suitable stem extractor to search for vocabulary, classifying the text, extracting feature values, and then performing machine learning on the classified text, and finally applying the classifier to the unclassified text to obtain classification tags, thereby constructing an event information text database;
[0009] Step 4: Collect and organize the animal pictures involved, mark the key feature points of the species in the pictures, use the deep learning model to train computer learning, and classify and identify the unclassified pictures, so as to build a species picture database;
[0010] Step 5: Associate the event information text database with the species image database.
[0011] Furthermore, the steps 1-4 specifically include the following contents:
[0012] Step 1: Create a database
[0013] Step 1-1, establish an animal category database
[0014] Create tables for the class, order, family, and species to which the animals belong:
[0015] Animal species table: primary key - Ename, Cname, foreign key - KeE, protection level;
[0016] Animal family table: primary key - KeE, KeC, foreign key - MuE;
[0017] Animal table: primary key - MuE, MuC, foreign key - GangE, GangC;
[0018] Animal class table: primary key - GangE, GangC, foreign key - MenE, MenC;
[0019] According to the above data table, it is converted into a relationship model between entities as shown below:
[0020] The animal species table includes: Ename, Cname, protection level, KeE (FK);
[0021] The animal family table includes: KeE, KeC, MuE(FK);
[0022] The animal list includes: MuE, GangE (FK), GangC, MuC;
[0023] The animal class list includes: GangE, GangC, MenE (FK), MenC;
[0024] Step 1-2, create an animal event information table
[0025] Prepare a summary table of event information based on animal event information, including time, unit, geographic coordinates, species involved, number of individuals and classification;
[0026] Step 2: Collect data
[0027] Collect and organize event information, text data, pictures of species involved, basic geographic data of the event location, and information collection;
[0028] Step 3: Event information processing
[0029] Step 3-1, data extraction
[0030] To extract event information according to the format of the event data source, it is necessary to first use pyhanlp entity recognition for precision, and regular expression matching is adopted overall;
[0031] Step 3-2, data processing
[0032] Select the required qualitative names, processing results and species content, make predictions through named entity recognition and multi-label classification, and use ALBERT to extract features from the text; finally, use the word cloud map to filter out a large amount of text information and highlight the keywords that appear frequently in the text; specifically:
[0033] Step 3-2-1, perform length screening, deduplication and digital conversion on the content returned by the regular expression, remove unreasonable content with too long characters in the extraction results; remove duplicate information in the extraction results; convert Chinese characters into Arabic numerals;
[0034] Step 3-2-2, model prediction
[0035] Select the required qualitative names, processing results and species content, make predictions through named entity recognition and multi-label text classification, use ALBERT to extract features from the text, and make three modifications to BERT:
[0036] (1) Factorization of word embedding vector parameters
[0037] Use ALBERT to factorize the word embedding parameters and decompose them into two small matrices. Instead of directly mapping one-hot vectors to a latent space of size H, they are first mapped to a low-dimensional word embedding space E and then mapped to the latent space.
[0038] (2) Cross-layer parameter sharing
[0039] Contribute all parameters of all layers;
[0040] (3) Paragraph continuity task
[0041] Use paragraph continuity tasks;
[0042] Establishing event information database;
[0043] Step 4: Image recognition
[0044] Step 4-1, image preprocessing
[0045] Preprocess and optimize the collected biological images; uniformly change the image format in the data set to jpg format, and crop the image size to the specified format; mark the key feature points of the species in the image;
[0046] Step 4-2, image recognition
[0047] Using the prepared sample images, the reinforcement transfer learning algorithm is used directly in the pycharm software for recognition training, unclassified images are intelligently identified according to known classification labels, the images are classified, and a species image database is constructed.
[0048] Furthermore, the step 1 also includes:
[0049] Steps 1-3, data table visualization
[0050] Use Tableau to connect to Microsoft SQL Server, connect the animal level, class, order, family, and species, and display the information visually; through Tableau custom variables and metrics, you can select the number of events in different years according to retrieval needs and visualize the fields in the data table.
[0051] Furthermore, in step 3-2-1, the entries with missing data are only used as samples for the total event data statistics and are not included in the statistical data when performing data analysis on each species later.
[0052] Furthermore, it also includes:
[0053] Step 6: Use PGIS to visualize event text information
[0054] The geographic coordinate marking method can be used to visualize event text information. PGIS technology can be used to visualize and trace the locations and species distribution involved in the animal event information intelligent retrieval system, evaluate the event information relationship network, supervise animal event information, and handle new event information in a timely manner.
[0055] The present invention also provides an animal information visualization retrieval system based on machine learning, including a database, a data acquisition module, an event information text database construction module, a species picture database construction module, and a retrieval module.
[0056] The beneficial effects of the present invention are:
[0057] Determine the characteristic values of information text and species pictures to successfully build the model, select machine learning model training sample data, test the classification effect of the classification model, establish associations between text information, species pictures, and GIS, and realize visual intelligent retrieval of animal event information. Based on the scheme of the present invention, it is possible to visually trace the locations and species distribution involved in the event, evaluate the event information relationship network, supervise animal event information, and process new event information in a timely manner. BRIEF DESCRIPTION OF THE DRAWINGS
[0058] Figure 1 A schematic diagram of the main process flow of the animal information visualization retrieval method based on machine learning provided by the present invention.
[0059] Figure 2 This is a list of animal species (excerpt).
[0060] Figure 3 This is a table of animal families (excerpt).
[0061] Figure 4 This is an animal list (excerpt).
[0062] Figure 5 This is an animal class table (excerpt).
[0063] Figure 6 This is a relationship diagram for the animal data table.
[0064] Figure 7 This is a summary table of animal event information (excerpt).
[0065] Figure 8 This is a diagram for connecting to the Tableau data source.
[0066] Fig. 9 This is a Tableau data visualization demonstration diagram.
[0067] Fig.10 Extract model graph for regular expressions.
[0068] Fig.11 This is a distribution map of the number of animal incidents from 2011 to 2020.
[0069] Fig.12 This is a schematic diagram showing the distribution of the number of animal documents in each province and city from 2011 to 2020.
[0070] Fig.13 Flowchart for bird image recognition.
[0071] Fig.14 Implement code for the reinforcement transfer learning algorithm.
[0072] Fig.15 A programming language for image recognition algorithms (excerpt).
[0073] Fig.16 Run the program. DETAILED DESCRIPTION
[0074] The technical solution provided by the present invention will be described in detail below in conjunction with specific embodiments. It should be understood that the following specific implementation methods are only used to illustrate the present invention and are not used to limit the scope of the present invention.
[0075] The present invention provides a method for visual retrieval of animal information based on machine learning, the main process of which is as follows: Figure 1 As shown, the following steps are included:
[0076] Step 1: Create a database
[0077] The construction of the database can visualize the relevant locations and species distribution in the event, evaluate the event information network, supervise the animal-related event information, and process the newly added event information in a timely manner. Therefore, the present invention establishes the following database:
[0078] Step 1-1, establish an animal category database.
[0079] According to the event information in the database, we obtain the species involved and create tables for the class, order, family, and species to which the animals belong, such as Figure 2-Figure 5 As shown:
[0080] Animal species table: primary key - Ename (animal-Latin name), Cname (animal-Chinese name), foreign key - KeE (family-Latin name), protection level.
[0081] Animal family table: primary key - KeE (family - Latin name), KeC (family - Chinese name), foreign key - MuE (order - Latin name).
[0082] Animal order table: primary key - MuE (order-Latin name), MuC (order-Chinese name), foreign key - GangE (class-Latin name), GangC (class-Chinese name).
[0083] Animal class table: primary key - GangE (class - Latin name), GangC (class - Chinese name), foreign key - MenE (phylum - Latin name), MenC (phylum - Chinese name).
[0084] According to the above data table, it is converted into Figure 6 The relationship model between entities is as follows:
[0085] 1. Animal species table (Ename, Cname, level, KeE(FK))
[0086] 2. Animal family table (KeE, KeC, MuE(FK))
[0087] 3. Animal list (MuE, GangE (FK), GangC, MuC)
[0088] 4. Animal class table (GangE, GangC, MenE (FK), MenC)
[0089] Step 1-2, create an animal event information table
[0090] Prepare an event information summary table based on animal event information, including time, unit, geographic coordinates, species involved, number of individuals and classification. For details, see Figure 7 .
[0091] In this example, the database is created using Microsoft SQL Server.
[0092] Steps 1-3, data table visualization
[0093] Use Tableau to connect to Microsoft SQL Server, connect animal classification, class, order, family, and species, and reduce data redundancy through information visualization. Figure 8 shown.
[0094] Through Tableau custom variables and metrics, you can select the number of event records in different years according to the search needs and visualize the fields in the data table, such as Fig. 9 shown.
[0095] Step 2: Collect data
[0096] Collect and organize event information, text data, pictures of species involved, basic geographic data of locations involved, and information collection.
[0097] Step 3: Event information processing and visualization
[0098] The event data used in the embodiment of the present invention comes from online records and documents. Full-text search was performed with specific animals as keywords, and a total of 20,548 information items were obtained. The text information was converted into txt format through word, and after data cleaning, 5,000 data were used. According to the 9:1 ratio of the ten-fold cross-validation method, the data used for the training set contained 4,500 documents, and the data utilization rate reached 21.9%.
[0099] Step 3-1, data extraction
[0100] According to the semi-structured format of the document, we extract the information about the personnel, time, processing agency, region, qualitative name, processing results, species involved, etc. Because the textual expression of the document usually has a fixed pattern, it needs to be refined through pyhanlp entity recognition first, and the overall regular expression matching is mainly adopted. The training set ran 10 rounds, and the overall accuracy range was 60%-72%. The regular expression is shown in the following table:
[0101] Target entity and regular expression
[0102] entity expression Personnel Name Requestee / Requester (.*?) time 20. *year. *month. *day Processing Agency Area (? <=document).*People's Court qualitative name (.*?)name Processing results Prefix (.*?) Animals (.*?)(^[0-9]*$) only / item (equivalent word)
[0103] Step 3-2, data processing
[0104] Select the required qualitative names, processing results and species content, make predictions through named entity recognition and multi-label classification, and use ALBERT to extract features from the text. Finally, use the word cloud map to filter out a large amount of text information and highlight the keywords that appear frequently in the text.
[0105] Step 3-2-1, information extraction
[0106] In order to improve the extraction accuracy and obtain better data formats, the return content of the regular expression needs to be processed by length screening, deduplication and digital conversion. Length screening is to remove unreasonable content with too long characters in the extraction results; deduplication removes repeated information in the extraction results; digital conversion converts Chinese characters into Arabic numerals. The overall structure of the extraction model is as follows Fig.10 shown.
[0107] Because each document involves different specific events and has different processing procedures, some events cannot be extracted in time; some documents do not clearly name the species of animals involved, but only generally classify them, such as crocodiles and parrots, and it is impossible to extract detailed information about the animals involved. Such entries with missing data are only used as samples for the total event data statistics and are not included in the statistical data when the data of each species is analyzed later.
[0108] Step 3-2-2, model prediction
[0109] Select the required qualitative names, processing results and species content, make predictions through named entity recognition and multi-label text classification, and use ALBERT to extract features from the text. The maximum text length is 200. The ALBERT model is an improved version of BERT. Unlike other state of the art models, it is suitable for pre-training small models, with better results and fewer parameters. It has made three modifications to BERT:
[0110] (1) Factorization of word embedding vector parameters
[0111] In BERT and subsequent XLNet and RoBERTa, the WordPiece word embedding size E and the hidden layer size H are equal, that is, E≡H. Due to modeling and practical use, it has two disadvantages:
[0112] ① From a modeling perspective, the goal of WordPiece embedding is to learn context-independent representations, while the goal of hidden layer embedding is to learn context-dependent representations. Through context-dependent experiments, a large part of BERT's representational ability comes from using context to provide context-dependent representation signals for the learning process. Therefore, separating the WordPiece embedding size E from the hidden layer size H can more efficiently utilize the overall model parameters, where H is much larger than E.
[0113] ② From a practical perspective, the dictionary size V used in natural language processing is very large. If E is always equal to H, then increasing H will directly increase the size of the embedding matrix, and this increase will also be amplified by V.
[0114] Therefore, ALBERT is used to factorize the word embedding parameters and decompose them into two small matrices. Instead of directly mapping the one-hot vectors to the hidden space of size H, they are first mapped to a low-dimensional word embedding space E and then mapped to the hidden space. Through this decomposition, the word embedding parameters can be reduced from O(V×H) to O(V×E+E×H), which is a very significant reduction in the number of parameters when H is much larger than E.
[0115] Taking ALBert_xxlarge as an example, V=30000, H=4096, E=128; the original parameters were V×H=30000×4096=123 million parameters, and now they are V×E+E×H=30000×128+128×4096=3.84 million+520,000=4.36 million. The parameters related to word embedding are 28 times larger before the change.
[0116] (2) Cross-layer parameter sharing
[0117] Parameter sharing can significantly reduce the number of parameters. There are many ways to share parameters, which can be divided into parameter sharing of fully connected layers and attention layers. The parameters of the attention layer have a smaller impact on the weakening effect. ALBERT uses all parameters of all layers to contribute.
[0118] (3) Paragraph continuity task
[0119] Use paragraph continuity tasks. For positive examples, use two consecutive text paragraphs from a document; for negative examples, use two consecutive text paragraphs from a document, but the positions are swapped. Avoid using the original NSP task, which contains overly simple tasks such as implicit topic prediction.
[0120] Step 3-2-3, word segmentation and word cloud
[0121] Word segmentation is a basic operation for text processing. Chinese word segmentation is much more complicated than English word segmentation. In Python, you can use the third-party jieba library for Chinese word segmentation. The jieba library can separate a Chinese text into a sequence of Chinese words. The word segmentation principle of the jieba library relies on a Chinese word library, which uses a Chinese word library to determine the association probability between Chinese characters. The Chinese characters with a high probability form a phrase to form the word segmentation result. In addition to word segmentation, users can also add custom phrases.
[0122] Word cloud is an emerging tool for data visualization on the Internet. Word cloud filters out a large amount of text information and arranges the keywords that appear frequently in the text in two-dimensional space. Different keywords use different colors and font sizes. By forming keyword rendering, the viewer can understand the main idea of the text with just one glance. Word Cloud library is an excellent third-party library for word cloud display. Word cloud uses words as the basic unit to display text more intuitively and artistically. The usual operation steps are: first, the text is segmented, then concatenated with spaces, and finally the Word Cloud library function is called for processing.
[0123] Currently, most semantic word cloud work relies on natural language processing methods to represent words and uses various dimensionality reduction algorithms to project words into two-dimensional space. The word cloud layout completed according to this idea may have local word neighbor relationships that are difficult to explain, such as words that originally have little semantic correlation but happen to be placed in close positions.
[0124] Step 3-3, result display
[0125] Using "region, handling agency, requested person, event name, time, qualitative name, handling result, amount, and animals involved" as keywords, the key information of the document is extracted according to the above technical process. The information extracted from the document is relatively complete, and all kinds of keywords can be accurately identified and extracted.
[0126] Statistics on the total number of documents published each year, such as Fig.11 By displaying the documents by year, it can be found that the number of animal-related incidents has increased year by year, with a significant increase starting in 2014, and especially in the past two years.
[0127] By visualizing the number of animal-related incidents in provinces across the country on a map, it can be found that areas with high incidence of animal-related incidents in my country are concentrated in East China, South China and Southwest China, and basically show a trend of decreasing from southeast to northwest, and from coastal and border provinces to inland areas.
[0128] Statistics on the number of documents per province and city from 2011 to 2020 Fig.12 Over the past 10 years, the number of incidents in each province and city has generally increased year by year.
[0129] The extracted event information is used for model training based on "judgment-duration|fine-amount content".
[0130] By comparing the prediction results obtained by the model with the actual trial situation, it is found that a small sample size will reduce the prediction accuracy, as shown in the following table:
[0131] Prediction results display
[0132]
[0133] By counting the species and numbers involved and displaying the species names in a word cloud, we can know which species appear in the largest number in animal events. The species with the largest number are called hotspot species, which will be discussed in detail in the following sections.
[0134] By counting the qualitative names in statistical documents and displaying them in word clouds, we can know the main qualitative types.
[0135] Through statistical analysis of event information, common problems in events can be discovered, which is more direct than standardized research methods.
[0136] Step 4: Image recognition
[0137] Step 4-1, image preprocessing
[0138] Image recognition, as an important method of biometric information identification, has very important application value in the field of information security and is a hot topic in the field of pattern recognition and computer vision. At present, biological image recognition cannot reach the general cognitive level of humans, but a trained system can classify and identify specific images to a certain extent.
[0139] Since transfer learning can only accept a few fixed sizes when accepting input images, and the sizes of the original image data set are different, the images must be preprocessed and optimized before inputting. The main methods of image preprocessing are: image cropping, size change, filter denoising, grayscale processing, enhancement, segmentation, etc.
[0140] First, we organized the bird pictures, uniformly changed the image format in the dataset to jpg format, and cropped the image size to 200 × 200. We performed preprocessing such as scaling, denoising, and segmentation on the images to further enhance the quality of the dataset and improve the diversity of the data, making it suitable for machine learning algorithms.
[0141] The key feature points of the species in the picture are marked, and the recognition training is performed using the reinforcement transfer learning algorithm to obtain the morphological characteristics of each bird, as follows:
[0142] Visual variables representing elements of species characteristic symbols
[0143] Visual variables Representative elements Body Parts Head, neck, wing spots, back, abdomen, tarsus, tail, toes shape Point, line, surface Color Code #000000-#FFFFFF
[0144] Using pycharm software as the platform, the algorithm is continuously improved by calling the reinforcement transfer learning algorithm data package, and unclassified images are intelligently identified based on known classification tags (which can be manually classified in advance), such as Fig.13 shown.
[0145] Step 4-2, image recognition
[0146] Some birds have very similar morphological features, but their genomes are very different. This recognition algorithm includes the recognition function of birds with similar appearance, which can accurately identify similar birds.
[0147] Using the prepared sample images, we directly performed the recognition operation in the pycharm software, performed 500,000 recognition calculations on the same target (bird), compared the obtained recognition results with the actual data, and selected the results with a recognition rate of more than 95% as successful recognition, and the percentage of successful recognition was the recognition success rate. After evaluating the recognition calculation of the target species, the algorithm can finally achieve an 83.77% recognition rate for the birds that need to be identified (the program code and running process are as follows Figure 14-16 shown).
[0148] During the experiment, by continuously proposing a series of identifiable bird images with increasing difficulty, the recognition algorithm can be gradually improved to improve the algorithm accuracy and recognition success rate, and gradually meet more accurate recognition requirements. By analyzing the recognition accuracy of the model and the characteristics of the sample images of the dataset, it is found that the discriminant features between species in the dataset are mainly color features, followed by morphological features. When the color and texture features of a certain category of dataset are relatively simple, the recognition accuracy is higher (such as black-necked swans); when the color features within the dataset sample class are more diverse, morphological features are the main basis for species discrimination (such as [tree] sparrows).
[0149] At present, image classification and recognition technology is a very important part of pattern recognition research. Reinforcement transfer learning also brings new ideas to image classification and recognition, and combines feature extraction and classification in traditional image classification technology. The research results show that the use of reinforcement transfer learning is better than traditional machine learning models and improves the high-precision classification and recognition of bird species.
[0150] After image recognition, a database of images of the species involved is constructed.
[0151] Step 5, linking the event information text database with the species image database by building an animal event information intelligent retrieval system;
[0152] Step 6: Use PGIS to visualize event text information.
[0153] PGIS is an extension of GIS technology. It uses geographic coordinate marking to visualize event text information. PGIS technology can be used to visualize the locations and species distribution involved in the animal event information intelligent retrieval system, evaluate the event information relationship network, supervise animal event information, and promptly process new event information, ultimately realizing the construction of an event emergency management intelligence system.
[0154] Based on the same inventive concept, the present invention provides an animal information visualization retrieval system based on machine learning, including a database, a data acquisition module, an event information text database construction module, a species picture database construction module, and a retrieval module, wherein the database is the database mentioned in step 1 in the aforementioned animal information visualization retrieval method based on machine learning, the data acquisition module implements the content of step 2 in the aforementioned method, the event information text database construction module implements the content of step 3 in the aforementioned method, and the retrieval module is used to associate the event information text database and the species picture database for retrieval.
[0155] It should be noted that the above content only illustrates the technical idea of the present invention and cannot be used to limit the protection scope of the present invention. For ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications all fall within the protection scope of the claims of the present invention.
Claims
1. A method for visual retrieval of animal information based on machine learning, characterized in that: The steps include: Step 1: Create a database Establish a database of event information and use Tableau to visualize the data; use pyhanlp entity recognition to refine the documents and visualize the number of animal events; Step 2, collect data; Step 3, using natural language processing technology (NLP), taking the animal event information text as the original corpus, purifying the text to make it a standardized text; selecting a suitable stem extractor to search for vocabulary, classifying the text, extracting feature values, and then performing machine learning on the classified text, and finally applying the classifier to the unclassified text to obtain classification tags, thereby constructing an event information text database; Step 4: Collect and organize the animal pictures involved, mark the key feature points of the species in the pictures, use the deep learning model to train the computer to learn, and classify and identify the unclassified pictures, so as to build a species picture database; Step 5: Associate the event information text database with the species image database.
2. The method for visual retrieval of animal information based on machine learning according to claim 1, characterized in that: The steps 1-4 specifically include the following contents: Step 1: Create a database Step 1-1, establish an animal category database Create tables for the class, order, family, and species to which the animals belong: Animal species table: primary key - Ename, Cname, foreign key - KeE, protection level; Animal family table: primary key - KeE, KeC, foreign key - MuE; Animal table: primary key - MuE, MuC, foreign key - GangE, GangC; Animal class table: primary key - GangE, GangC, foreign key - MenE, MenC; According to the above data table, it is converted into a relationship model between entities as shown below: The animal species table includes: Ename, Cname, protection level, KeE (FK); The animal family table includes: KeE, KeC, MuE(FK); The animal list includes: MuE, GangE (FK), GangC, MuC; The animal class list includes: GangE, GangC, MenE (FK), MenC; Step 1-2, create an animal event information table Prepare a summary table of event information based on animal event information, including time, unit, geographic coordinates, species involved, number of individuals and classification; Step 2: Collect data Collect and organize event information, text data, pictures of species involved, basic geographic data of the event location, and information collection; Step 3: Event information processing Step 3-1, data extraction To extract event information according to the format of the event data source, it is necessary to first use pyhanlp entity recognition for precision, and regular expression matching is adopted overall; Step 3-2, data processing Select the required qualitative names, processing results and species content, make predictions through named entity recognition and multi-label classification, and use ALBERT to extract features from the text; finally, use the word cloud map to filter out a large amount of text information and highlight the keywords that appear frequently in the text; specifically: Step 3-2-1, filter the length, remove duplicates and convert numbers of the content returned by the regular expression, remove unreasonable content with too long characters in the extraction results; remove duplicate information in the extraction results; convert Chinese characters to Arabic numerals; Step 3-2-2, model prediction Select the required qualitative names, processing results and species content, make predictions through named entity recognition and multi-label text classification, use ALBERT to extract features from the text, and make three modifications to BERT: (1) Factorization of word embedding vector parameters Use ALBERT to factorize the word embedding parameters and decompose them into two small matrices. Instead of directly mapping one-hot vectors to a latent space of size H, they are first mapped to a low-dimensional word embedding space E and then mapped to the latent space. (2) Cross-layer parameter sharing Contribute all parameters of all layers; (3) Paragraph continuity task Use paragraph continuity tasks; Establishing event information database; Step 4: Image recognition Step 4-1, image preprocessing Preprocess and optimize the collected biological images; uniformly change the image format in the data set to jpg format, and crop the image size to the specified format; mark the key feature points of the species in the image; Step 4-2, image recognition Using the prepared sample images, the reinforcement transfer learning algorithm is used directly in the pycharm software for recognition training, unclassified images are intelligently identified according to known classification labels, the images are classified, and a species image database is constructed.
3. The method for visual retrieval of animal information based on machine learning according to claim 2, characterized in that: The step 1 also includes: Steps 1-3, data table visualization Use Tableau to connect to Microsoft SQL Server, connect the animal level, class, order, family, and species, and display the information visually; through Tableau custom variables and metrics, you can select the number of events in different years according to retrieval needs and visualize the fields in the data table.
4. The method for visual retrieval of animal information based on machine learning according to claim 2, characterized in that: In step 3-2-1, the entries with missing data are only used as samples for the total event data statistics and are not included in the statistical data when performing data analysis on each species later.
5. The method for visual retrieval of animal information based on machine learning according to claim 1, characterized in that: Also includes: Step 6: Use PGIS to visualize event text information The geographic coordinate marking method can be used to visualize event text information. PGIS technology can be used to visualize and trace the locations and species distribution involved in the animal event information intelligent retrieval system, evaluate the event information relationship network, supervise animal event information, and handle new event information in a timely manner.
6. The animal information visualization retrieval system based on machine learning is characterized by: A method for visually retrieving animal information based on machine learning for implementing any one of claims 1 to 5, comprising a database, a data acquisition module, an event information text database construction module, a species picture database construction module, and a retrieval module.