Content-adaptive and aware web map intelligent recommendation system

By employing automatic sampling methods based on information content and approximation, and image clustering driven by color and texture, an intelligent retrieval and recommendation platform for network maps is built. This solves the problems of poor search experience and low efficiency in existing map services, and realizes map service recommendation based on image content, thereby improving user experience and discovery efficiency.

CN114238772BActive Publication Date: 2024-11-12韩效遥
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202111603724.1
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-12-24
Publication Date
2024-11-12
Estimated Expiration
2041-12-24

AI Technical Summary

Technical Problem

Existing technologies cannot effectively manage massive amounts of online geographic information resources, cannot quickly match user needs, lack image content-based map service recommendation functions, resulting in poor map service search experience and low discovery efficiency, and text-based image retrieval suffers from labeling ambiguity and insufficient accuracy.

Method used

An automatic sampling method for web map services based on information content and approximation is adopted, combined with color and texture-driven image clustering, to build an intelligent retrieval and recommendation platform for web maps. By compiling web crawlers to crawl map service metadata, a keyword retrieval library and a feature library are established to achieve recommendations based on image content.

Benefits of technology

It improves the search experience and discovery efficiency of map services, accurately identifies map content, recommends similar layers with a high degree of matching with actual needs, makes up for the shortcomings of keyword-based search, and provides a more intelligent map retrieval and recommendation method.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114238772B_ABST
    Figure CN114238772B_ABST
Patent Text Reader

Abstract

This application integrates image content-based clustering and image recommendation into the retrieval and recommendation of network map services, which greatly improves the experience of map service search in the existing technology and improves the efficiency of service discovery. It proposes a complete set of technical processes from data collection and processing, thumbnail generation, image clustering, and platform construction, realizes automatic sampling of network map services based on information volume and proximity, and recommends color and texture-driven image clustering maps. On this basis, a map service retrieval and recommendation platform on the Web is designed and implemented, which has accurate grasp and recognition of map content and can make intelligent recommendations for network maps based on content adaptive perception. It also develops a map service retrieval platform based on keyword search and image content recommendation, and the retrieved map content has a high degree of match with actual needs, which verifies the rationality and effectiveness of the method of this application and has high practical value.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to a network map intelligent recommendation system, and in particular to a content-adaptive-aware network map intelligent recommendation system, belonging to the technical field of network map content intelligent recommendation. Background Art

[0002] Web map service is a new way of organizing maps in the mobile Internet era. It integrates multi-scale coupling and high interoperability on the basis of basic map visualization. In the era of big data, map service has become the best knowledge expression and presentation service with low resource requirements, intuitive visualization and seamless cross-platform connection. In recent years, the quantity and quality of map services have been continuously improved, and the public's demand for map services has also been gradually increasing. How to effectively manage these map services and quickly match user needs has become a hot spot in research and development and application.

[0003] Currently, a large amount of geographic information data, geographic information networks, geographic information portals, geographic information documents, etc. are scattered on heterogeneous network nodes, which constitute a massive amount of network geographic information resources. At the same time, units and individuals will obtain various geographic information resources for auxiliary decision-making, application management, and auxiliary scientific research. Among them, network map services represented by OGC WMS are widely used in various industries. How to quickly retrieve and find map services that meet the needs is a problem that needs to be solved urgently.

[0004] Compared with traditional keyword-based text retrieval, image-based search has a better visual experience and more accurate search results. Baidu, Google, etc. have launched image-based search functions. If these two Internet ideas are introduced into map services, the development of image-based map service recommendation functions can effectively assist keyword-based search functions, improve the existing map service search experience, and improve service discovery efficiency.

[0005] The content-aware map recommendation system is an important and effective supplement to the existing text-based service search. The recommendation function based on image content makes up for the existing single keyword search-based geographic information service discovery method. It plays an important role and has great application value in enhancing user experience and improving the efficiency of map service discovery.

[0006] Image retrieval initially mainly retrieved images through text information, such as the shooting date, author, category and size of the image. This type of image retrieval method is a text-based image retrieval technology. After more than ten years of development, image retrieval has developed into content-based retrieval, which mainly retrieves images through image features such as color, texture, and shape. Its purpose is to avoid using text descriptions, but to retrieve similar images based on visual similarity through images provided by users or image features specified by users.

[0007] At present, text-based image retrieval is very mature, but it also has its shortcomings and deficiencies: first, the text and keyword annotation of images needs to be completed manually; second, manual annotation has problems such as different observation angles and subjective differences, which will cause annotation ambiguity, that is, it is difficult for manual annotation to fully express an image that contains multiple targets or even covers some emotional colors. Text-based image retrieval still has shortcomings. With the improvement and deepening of market demand and application, image retrieval applications based on image content such as Baidu Image Recognition and Google Image Recognition have emerged one after another. Although the accuracy needs to be improved, it can automatically extract image features and avoid ambiguity.

[0008] In general, the existing web map recommendation system has obvious shortcomings, and its main defects and design difficulties include:

[0009] First, in recent years, the quantity and quality of map services have been continuously improving, and the public's demand for map services has also been gradually increasing. Existing technologies cannot effectively manage these map services and quickly match user needs. Currently, a large amount of geographic information data, geographic information networks, geographic information portals, geographic information documents, etc. are scattered on heterogeneous network nodes, which constitute a massive amount of network geographic information resources. Organizations and individuals will obtain various geographic information resources for auxiliary decision-making, application management, and auxiliary scientific research. Existing technologies cannot quickly retrieve and discover map services that meet the needs, and lack a more intelligent content-based map retrieval recommendation method.

[0010] Second, compared with traditional keyword-text based retrieval, image content based search has a better visual experience and more accurate search results. The existing technology has not introduced this Internet idea into map services, and lacks the map service recommendation function based on image content, resulting in poor search experience and low service discovery efficiency of map services in the existing technology. Since content-aware map recommendation is an effective supplement to the text-based service search in the existing technology, image content-based recommendation can make up for the single keyword search geographic information service discovery method in the existing technology, and plays an important role and has great application value in enhancing user experience and improving map service discovery efficiency. However, it is difficult to implement the existing technology of intelligent recommendation of network maps based on image content.

[0011] Third, the existing text-based map image retrieval technology is very mature, but it also has its shortcomings and deficiencies: first, the text and keyword annotation of the image needs to be completed manually; second, there are problems such as different observation angles and subjective differences in manual annotation, which will cause annotation ambiguity, that is, it is difficult for manual annotation to fully express an image containing multiple targets or even covering some emotional colors. Text-based image retrieval still has shortcomings. With the improvement and deepening of market demand and application, image retrieval applications based on image content have emerged one after another, but they cannot be applied to maps, or there are problems such as low accuracy, inability to automatically extract image features, and too much ambiguity;

[0012] Fourth, there are several difficulties in implementing content-aware web map recommendations. First, there is a lack of feasible automatic sampling methods for web map services based on information volume and proximity, and a lack of map service image sampling methods under information volume and proximity constraints. It is impossible to generate multiple sampling maps for each layer, and it is impossible to use these sampling maps to establish an image library; second, there is a lack of feasible image clustering map recommendation methods, a lack of image group-based clustering methods to classify layers based on the features of multiple sampling maps of each layer, and a lack of similar layer recommendation methods that take into account the features of layer image groups; third, there is a lack of feasible web map service retrieval and recommendation platform design and development methods, and a lack of high-quality and efficient web map intelligent retrieval and recommendation platforms. Summary of the invention

[0013] In response to the shortcomings of the existing technology, this application integrates image content-based clustering and image recommendation into the retrieval and recommendation of network map services, which greatly improves the experience of map service search in the existing technology and improves the efficiency of service discovery. It proposes a complete set of technical processes from data collection and processing, thumbnail generation, image clustering, and platform construction, and realizes automatic sampling of network map services based on information volume and proximity, and color and texture driven image clustering map recommendation. On this basis, a map service retrieval and recommendation platform on the Web is designed and implemented, which has accurate grasp and identification of map content and can make intelligent recommendations for network maps based on content adaptive perception. It also develops a map service retrieval platform based on keyword search and image content recommendation. The retrieved map content has a high degree of match with actual needs, which verifies the rationality and effectiveness of the method of this application and has high practical value.

[0014] In order to achieve the above technical effects, the technical solutions adopted in this application are as follows:

[0015] The content-adaptive network map intelligent recommendation system integrates image content-based clustering and image recommendation into the retrieval and recommendation of network map services. From data collection and processing, thumbnail generation, image clustering, to platform construction, a complete set of technical processes are designed and implemented on this basis. The web-based map service retrieval and recommendation platform mainly includes:

[0016] First, an automatic sampling method for web map services based on information volume and similarity includes: first, calculating and evaluating the information volume of map tiles, and second, calculating the similarity of cross-scale map tiles; using information volume combined with a quadtree structure to locate content-rich areas in the map, and the information volume is approximately expressed by image complexity; the similarity is calculated by a color quantized unified histogram to screen the scale ratio that has undergone mutations. The two methods realize the orderly traversal of map services from the plane position dimension and the scale ratio dimension respectively, providing effective data for clustering recommendations based on image content;

[0017] Second, color and texture driven image clustering map recommendation, including: first, map feature collaborative representation of integrated color and texture features, second, map content image clustering, third, clustering method based on image group, fourth, map service recommendation based on image group; color features and texture features are used to comprehensively represent image features. Color features are represented by color quantization unified histogram, and texture features are represented by grayscale co-occurrence matrix. The inclusion relationship and color similarity of image groups are considered. The image group recommendation method reduces the proximity difference caused by layer positioning by assigning different weight values ​​to different proximity groups. A multi-feature representation method of integrated color and texture is proposed to realize map service recommendation based on image groups.

[0018] Third, build a network map intelligent retrieval recommendation platform, including: first, the map intelligent retrieval recommendation platform architecture, second, network map data collection and processing, third, map keyword retrieval library, fourth, map image preprocessing and image retrieval, and fifth, network map intelligent retrieval recommendation platform access; crawl a large number of map services by compiling web crawlers, parse their metadata and store them in the database, build a keyword retrieval library based on Apache Solr service and publish keyword retrieval services, generate thumbnails for each service and store them, calculate the color and texture features of each thumbnail and store them in the feature library, and store the clustering results in the category library. When recommending similar images, first find the category they belong to and then find the most similar image under this category and recommend it to the user.

[0019] Content-adaptive-aware web map intelligent recommendation system, further, content-based adaptive sampling method for map service preview: using some web map tiles as preview images, in tile map mode, the map service is concisely expressed as selecting a number of tiles that can represent the map service content from a large number of map tiles, which is further decomposed into two sub-modules: tile selection in the plane position dimension and tile selection in the scaling dimension;

[0020] (1) Tile selection in the plane position dimension: Based on a specific scaling ratio, the representativeness of the map tile is quantified. The complexity factor of the map tile is used to approximate its information content and is used as a quantitative indicator of tile representativeness.

[0021] (2) Tile selection in the scaling dimension: The key is to find the tile where mutation occurs. Image approximation is used to quantitatively evaluate the similarity between two scale tiles, and based on this, the scaling ratio where significant changes occur is determined.

[0022] Based on the two quantitative indicators of complexity and approximation, two rounds of screening are used to achieve full coverage of the plane position dimension and automatic selection of multi-scale tiles in the scaling dimension:

[0023] In the first round of screening, starting from the smallest scale, only the most complex tiles are selected at each level, and the geographical range of the tile is used as a constraint to process the next scale. The first round of screening filters out most tiles;

[0024] Second round of screening: Compare the similarity of tiles between scaling ratios, further remove redundancy, and define the sampling set as R. The specific steps are as follows:

[0025] Step 1: Starting from level L0, add the only map tile in level L0 to R and use it as the seed tile;

[0026] Step 2: Request four higher-level map tiles with the same geographical range as the seed tile, and use the tile Ringo with the richest information as the new seed tile;

[0027] Step 3: Combine the four tiles to get Rpvda. If the similarity between Rpvda and the seed tile is greater than the critical value, add Rpvda to R, otherwise do not add it.

[0028] Repeat steps 2 and 3 until all scaling ratios are traversed to obtain the final set of sampling images;

[0029] Flexible adjustment of multiple variables: The number of tiles with adjacent scaling ratios does not necessarily satisfy the 4-fold relationship, which affects the number of tiles to be stitched together. The first r (r≥1) tiles with the most information are used as seed tiles to increase the probability of obtaining representative tiles.

[0030] Content-adaptive-aware network map intelligent recommendation system, further, map tile information volume calculation evaluation: based on the map information volume, the first round of screening is carried out, and the map tile with the richest information volume is selected as the seed tile at each zoom ratio, and the geographic space range covered by the seed tile is used as a constraint to heuristically traverse the next zoom ratio. Except for L0 and L1 levels, it is not guaranteed that the seed tiles of other levels have the maximum information volume, but a higher hit rate is obtained at a lower calculation cost. After rough selection, no more than 4 map tiles are involved in the comparison at each level;

[0031] In terms of the calculation and evaluation of map information volume, the data volume of the compressed map tiles is used as the representation indicator of its information volume, and the data volume of the map tiles in PNG format is used as the information volume indicator.

[0032] The content-adaptive network map intelligent recommendation system further calculates the similarity of cross-scale map tiles: compares the similarity of map tiles between two adjacent zoom ratios, and judges sudden or slow changes based on this, and only retains map tiles that have changed significantly; analyzes the similarity of map tiles based on color features, uses color quantization unified histogram to quantify the color features of the map, and on this basis measures the similarity of map tiles based on cosine distance weights;

[0033] This application combines map expression features to make two improvements to the histogram generation method:

[0034] Improvement 1, for color quantization: First, quantize the colors of the map tiles. First, combine the two tiles to be compared together, unify all the colors of the two tiles, and quantize to obtain a significant and unique color set;

[0035] Improvement 2, calculate the frequency of color occurrence: first divide the map into several continuous color blocks, and use the perimeter of these color blocks to calculate the frequency of occurrence of the corresponding color. The division standard of continuous color blocks is that pixels of the same color are adjacent in four neighborhoods. The perimeter calculation adopts Moore neighborhood boundary tracking method;

[0036] In the RGB color space, color quantization uses the minimum variance method to specify the number of retained colors. Based on the quantized color set, the frequency of occurrence of each color in the two tiles is counted to obtain a unified color quantization histogram. Finally, the cosine distance is used to calculate the similarity between the histograms. The weight value of the color with the largest proportion in the histogram is set to 0 to suppress its influence on the approximation.

[0037] Content-adaptive-aware web map intelligent recommendation system. Further, color and texture-driven image clustering map recommendation: map images are classified according to image features. Before layer recommendation, the layers are clustered to narrow the search range of similar layers. The color quantization unified histogram and gray-level co-occurrence matrix are used to represent the color and texture features of the image. On the basis of generating 3 sampling images for each layer, image group clustering recommendation is performed based on the web map service features.

[0038] Collaborative representation of map features based on color and texture features: For a color map, its color features and texture features are denoted as G s and G r , for two given images D and E, the color feature difference between them is recorded as A s , the texture feature difference is A r , k1 and k2 are the weight values ​​of the distance between the two features, then the comprehensive distance between D and E is:

[0039] A(D, E) = k1*As (GS D , G.S. E )+k2*Ar(Gr D , Gr E )

[0040] This is used as a similarity comparison method for map images and retrieval.

[0041] Content-adaptive-aware web map intelligent recommendation system, further, map content image clustering: at the beginning of clustering, all objects in the map data set are used as cluster centers. In each iteration, each cluster center is updated through the message passing mode. After each iteration, the object that failed in the competition selects a cluster center as its center. Through continuous iteration and selection, multiple high-quality clustering results are finally obtained;

[0042] Clustering is performed based on the approximation between M map data points. This approximation constitutes an M×M similarity matrix. All map data points are regarded as potential cluster centers Z. The larger the value c(w, w) on the diagonal of the C matrix, the greater the probability that point w will become a cluster center. Its value is the reference degree q. The transmission mechanism mainly includes attraction factors and attribution factors. The attraction factor and attribution factor matrices are updated. To avoid result oscillation, the attenuation coefficient k is introduced. When updating information, each information is set to k times the last iteration update value plus 1-k times the information update value, where the attenuation coefficient k is a real number between 0 and 1. That is, the iterative value of t(j, w) and d(j, w) for the r+1th time:

[0043] d r+1 (j, w)←(1-k)d r+1 (j,w)+kd r (j, w)

[0044] Specific steps for map content image clustering:

[0045] Step 1: Initialize parameters and read map content data;

[0046] Step 2: Calculate the proximity matrix, find the median value of the proximity matrix and assign it to the reference degree q;

[0047] Step 3: Update the attraction factor matrix and the attribution factor matrix;

[0048] Step 4: Determine whether the maximum number of iterations has been reached or the termination condition has been met. If so, go to step 2, otherwise go to step 5;

[0049] Step 5: Get the final cluster center Z and divide each data object into the corresponding cluster.

[0050] Content-adaptive-aware web map intelligent recommendation system, further, based on the clustering method of image groups: using the content-based map service preview adaptive sampling method, three preview images are generated for each layer. These three preview images are expressions of the changes in information of different zoom ratios of this layer. Whether it is clustering or layer recommendation, it is necessary to integrate the three preview images of each layer;

[0051] The three preview images of each layer are taken as an image group. The image groups of each layer are clustered. First, the features of the three preview images in the image group are analyzed. The three images come from different zoom levels of the same layer. Layers at different zoom levels express the same content of the layer at different scales, and the color expression of the same object at different scales will not change.

[0052] When clustering image groups according to color features, the color features of the image groups are comprehensively considered. Suppose the color quantization unified histograms of the three images A, B, and C are R A ,R B ,R C Due to the similarity of the color features of the images in the image group, the color feature of the image group is R = R A +R B +R C , the calculation takes into account the characteristics of the images in the image group and comprehensively considers the color characteristics of the image group.

[0053] Content-adaptive-aware web map intelligent recommendation system, further, map service recommendation based on image groups: similar layer recommendation based on image content comprehensively considers the three preview images of the layer, and on this basis, a similar service recommendation method based on map service sampling images is proposed;

[0054] Suppose there are layers D and E, each layer has 3 preview images, namely D1, D2, D3 and E1, E2, E3, calculate the similarity between the sampled image groups of layer D and layer E; considering that the 3 images of each layer reduce the low similarity of the two images caused by similar services being located at different levels or different areas, the following image group similarity calculation method is proposed:

[0055] Step 1: Calculate the image similarity between image groups D1, D2, D3 and image groups E1, E2, E3 respectively, using the color plus texture feature calculation method. The similarities are C ji , where j = {1, 2, 3} is D j The subscript of i={1,2,3} is E i Subscript, assuming C ji The maximum value in is C 11 , then among these 9 groups of comparisons, D1 and E1 are the most similar, and C1=C 11 , excluding those containing Di and E i All groups of C 22 , C 23 , C 32 , C 33 ;

[0056] Step 2: Similarly, C 22 , C 23 , C 32 , C 33 Find the maximum value among the four groups and set it as C 32 , let C2=C 32 , excluding images D3 and E2, now only the group of images D2 and E3 is left, let C3=C 23 ;

[0057] Step 3: After the first and second steps, the three most similar image combinations are found between the two image groups. To reduce the impact of different zoom levels and different positions on the similarity, a different weight value k is set for each group of similarities. The final similarity between the image groups is:

[0058] C=k1×C1+k2×C2+k3×C3

[0059] Among them, 1≥k1>k2>k3≥0.

[0060] A content-adaptive web map intelligent recommendation system, further, builds a web map intelligent retrieval recommendation platform: including geographic information service data collection, data processing, keyword retrieval library and image feature retrieval design and construction, builds a web-based web map service keyword retrieval and similar service recommendation platform, crawls a large number of map services through web crawlers, parses their metadata, stores them in the database and establishes a keyword retrieval library, publishes keyword retrieval services based on Apache Solr, samples representative images for each layer, and then calculates the color and texture features of each image to build a feature library, and stores the clustering results in the category library through clustering. When similar images need to be recommended, the system first finds the category to which they belong, and then finds the layer to which the most similar image belongs under this category and recommends it to the user.

[0061] (I) Map intelligent retrieval and recommendation platform architecture

[0062] By building a web crawler to collect web map services scattered on various nodes of the Internet, the collected structured web map service data is parsed and stored in the database as metadata for display to users. In addition, through the layer acquisition interface provided by the service, combined with the automatic generation method of web map service thumbnails, thumbnails are generated for each layer. These thumbnails represent the service content and are the data basis for subsequent image feature extraction and image clustering. The map intelligent retrieval recommendation platform framework is divided into three layers:

[0063] Layer 1: Web map service collection and preprocessing layer: By compiling web map service theme crawlers, we can crawl scattered map services on the Internet, and then download and parse their metadata according to their service versions and specifications. Based on the structured organization form, we can establish a database table structure according to its various parameter information, and store the processed data in the MySQL database.

[0064] Layer 2: Image preprocessing and keyword search library construction: This layer is divided into two parts, one for the platform's keyword search and the other for similar service recommendations. For keyword search, a directional index library from keywords to services is established. For similar service recommendations, thumbnails are generated for each service. The color and texture features of the services that have been reserved in this application are calculated in advance in the background and stored in the feature library to facilitate subsequent clustering.

[0065] Layer 3: Platform query and recommendation logic implementation: Implement the background logic of the query recommendation platform, mainly including: map service query and display based on keywords, and similar service recommendations based on thumbnail content.

[0066] (II) Network map data collection and processing

[0067] The geographic information service theme crawler is developed to collect OGC WMS and ArcGIS MapService on the Internet. The ideas and logic of the geographic information service theme crawler include:

[0068] (1) Obtain the seed point URL using keywords through a search engine;

[0069] (2) Read the seed URL webpage by compiling a crawler program and extract other URL links obtained from the webpage;

[0070] (3) Parse the web page content. If the page contains the keywords WMS, web map service, or ARCGIS REST Services, it will be retained as a URL to be processed.

[0071] (4) If the URL link contains the keywords service=wms&request=getcapbilities and the page contains wms-capbilities, it is determined to be an OGC WMS service; if the link contains / arcgis / rest / services and the page content contains ArcGIS REST Services Directory, it is determined to be an ArcGISService service and the service link is stored in the database.

[0072] The captured services are stored in the database in linked form, and the metadata of the services are obtained. According to the service publishing specification type and its service version, web pages are requested and metadata parsed, and only the more important information is selected and stored in the database.

[0073] Content-adaptive-aware web map intelligent recommendation system,Further,(iii) map keyword retrieval library: establish a reverse index library of keywords to services, establish a mapping of keywords to service layers, and select the description information and title of each service layer as the fields for keyword matching;

[0074] (1) Treat each layer of WMS and ArcGIS Map Service as a document, select the Title and Description in each layer as the fields to be segmented, and pass the information to be indexed to the Tokenizer component to obtain the word-unit Token information of each document by removing punctuation and pauses;

[0075] (2) The language analysis tool is then used to process the previously obtained lemmas, including converting uppercase letters to lowercase letters and converting single words to root forms, to obtain the word Term.

[0076] (3) Pass the term to the indexer, create a word dictionary with the indexed terms and sort them in alphabetical order, merge the same terms into a posting list;

[0077] The objects of this application are OGC WMS and ArcGIS MapService map services. The purpose is to index the service layer information with keywords and find the appropriate service description information in its metadata.

[0078] The search service is established by creating a word-based inverted index table. English words are separated by spaces and punctuation marks. The search results display information including: layer title, layer description information, layer name, and sampling map connection. The corresponding fields are mapRiRle, mapAesc, mapMame, and images.

[0079] Keyword search service publishing: Publish keyword search as a separate service to decouple the keyword search service from the entire system. Select Tomcat as the publishing server for the Solr service. After the service is successfully published, view the information in Solr through a visual web page. Query the search results by configuring query conditions on the page. Integrate the service into the platform using the SolrJ tool to call the Solr search service.

[0080] Compared with the prior art, the innovations and advantages of this application are:

[0081] First, this application integrates image content-based clustering and image recommendation into the retrieval and recommendation of network map services, which greatly improves the experience of map service search in the existing technology and improves the efficiency of service discovery. It proposes a complete set of technical processes from data collection and processing, thumbnail generation, image clustering, and platform construction, and then designs and implements a web-based map service retrieval and recommendation platform on this basis. The proposed image content-based map service recommendation method accurately grasps and identifies map content, and can perform intelligent network map recommendations based on content adaptive perception. It also develops a map service retrieval platform based on keyword search and image content recommendation. The recommended layer has a high similarity with the comparison layer, and the retrieved map content has a high degree of match with actual needs, which verifies the rationality and effectiveness of the method of this application and has good practical value.

[0082] Second, this application introduces image content-based search into web map services, develops an image-based map service recommendation function, assists keyword-based search, improves the existing web map service search experience, and proposes an automatic sampling method for web map services based on information volume and proximity. The information volume is combined with a quadtree structure to locate content-rich areas in the map, and the information volume is approximately expressed by image complexity; the proximity is calculated by color quantization unified histogram to screen the scale ratio that has undergone mutations. The two respectively realize the orderly traversal of map services from the plane position dimension and the scale ratio dimension, providing effective data for clustering recommendations based on image content. Experiments have shown that image complexity is very effective in locating key locations of maps, and can accurately select the map tiles with the richest content at the same scale ratio. Using proximity to screen key scales has also achieved good results, improving the efficiency of map service discovery;

[0083] Third, the present application proposes to use color features and texture features to comprehensively represent image features. The color features are represented by a color quantized unified histogram, and the texture features are represented by a grayscale co-occurrence matrix. The inclusion relationship of the image group and the color similarity are considered. The image group recommendation method reduces the similarity difference caused by layer positioning by assigning different weight values ​​to different similarity groups. The color and texture driven image clustering map recommendation method considers the inclusion relationship of the image group and the color similarity. The image group recommendation method reduces the similarity difference caused by layer positioning by assigning different weight values ​​to different similarity groups. Experiments show that the present application comprehensively considers the color and texture features of the image group, takes into account the characteristics of the image group, and has good recommendation results for different types of layers.

[0084] Fourth, this application has built an intelligent retrieval and recommendation platform for web maps. It crawls a large number of map services by compiling web crawlers, parses their metadata and stores them in a database, establishes a keyword retrieval library based on the Apache Solr service and publishes a keyword retrieval service, generates thumbnails for each service and stores them, calculates the color and texture features of each thumbnail and stores them in a feature library, stores the clustering results in a category library, and when recommending similar images, finds the category to which they belong and then finds the most similar images under this category and recommends them to users. This is an important and effective supplement to the text-based service search in the prior art, and makes up for the single keyword-search-based geographic information service discovery method in the prior art. It plays an important role and has great application value in enhancing user experience and improving the efficiency of map service discovery. BRIEF DESCRIPTION OF THE DRAWINGS

[0085] Figure 1 This is a schematic diagram of the data volume of PNG images outputted by map symbols of different complexity.

[0086] Figure 2 This is an example of locating complex areas of a map based on PNG data.

[0087] Figure 3 It is a multi-level conceptual division diagram of the multi-source vector road network real-time fusion model.

[0088] Figure 4 It is a display diagram of similar color features in the image group of this application.

[0089] Figure 5 This is a schematic diagram of the first step of screening for map service recommendation based on image groups.

[0090] Figure 6 This is a schematic diagram of the second step of screening for map service recommendation based on image groups.

[0091] Figure 7 This is a schematic diagram of the input image group of Experiment 1 of this application.

[0092] Figure 8 This is a schematic diagram of the recommended image group for Experiment 1 of this application.

[0093] Fig. 9 This is a schematic diagram of the input image group of Experiment 2 of this application.

[0094] Fig.10 This is a schematic diagram of the recommended image group for Experiment 2 of this application.

[0095] Fig.11 This is a schematic diagram of the input image group of Experiment 3 of this application.

[0096] Fig.12 This is a schematic diagram of the recommended image group for Experiment 3 of this application.

[0097] Fig.13 This is a framework diagram of a content-adaptive web map intelligent recommendation platform.

[0098] Fig.14 It is a schematic diagram of the specific steps of keyword indexing of this application.

[0099] Fig.15 It is a schematic diagram of the steps of map image preprocessing and image retrieval process. Specific implementation methods

[0100] The following, in conjunction with the accompanying drawings, further describes the technical solution of the content-adaptive-aware web map intelligent recommendation system set up in the present application, so that those skilled in the art can better understand the present application and implement it.

[0101] The user experience of map service search and the efficiency of service discovery are crucial. Compared with keyword-based text search, image-based search has a better visual experience and better image search efficiency. Baidu, Google, etc. have launched image-based search functions. If this idea is introduced into network map services and image-based map service recommendation functions are developed, it can effectively assist keyword-based searches, improve the existing network map service search experience, and improve the efficiency of map service discovery.

[0102] This application integrates image content-based clustering and image recommendation into the retrieval and recommendation of network map services to improve the experience of existing map service search and improve the efficiency of service discovery. It proposes a complete set of technical processes from data collection and processing, thumbnail generation, image clustering, and platform construction. On this basis, a web-based map service retrieval and recommendation platform is designed and implemented, which mainly includes:

[0103] First, an automatic sampling method for web map services based on information content and proximity, including: calculating and evaluating the information content of map tiles, and calculating the proximity of cross-scale map tiles;

[0104] The method of generating map service thumbnails based on information volume and proximity adopts information volume combined with quadtree structure to locate the content-rich areas in the map. The information volume is approximately expressed by image complexity. The proximity is calculated by color quantization unified histogram to screen the sudden change of scaling ratio. The two realize the ordered traversal of map services from the plane position dimension and scaling ratio dimension respectively, providing effective data for clustering recommendation based on image content.

[0105] Second, color and texture driven image clustering map recommendation, including: map feature collaborative representation of integrated color and texture features, map content image clustering, clustering method based on image group, map service recommendation based on image group;

[0106] The color features and texture features are used to comprehensively represent the image features. The color features are represented by the color quantization unified histogram, and the texture features are represented by the gray-level co-occurrence matrix. The inclusion relationship and color similarity of the image group are considered. The image group recommendation method reduces the proximity difference caused by layer positioning by assigning different weight values ​​to different proximity groups. A multi-feature representation method that integrates color and texture is proposed to realize map service recommendation based on image groups.

[0107] Third, build a network map intelligent retrieval recommendation platform, including: map intelligent retrieval recommendation platform architecture, network map data collection and processing, map keyword retrieval library, map image preprocessing and image retrieval, network map intelligent retrieval recommendation platform construction and access;

[0108] Map service retrieval platform based on keyword search and image content recommendation: crawl a large number of map services by compiling web crawlers, parse their metadata and store them in the database, build a keyword retrieval library based on Apache Solr service and publish keyword retrieval services, generate thumbnails for each service and store them, calculate the color and texture features of each thumbnail and store them in the feature library, and store the clustering results in the category library. When recommending similar images, first find the category they belong to and then find the most similar images under this category and recommend them to users. This platform uses a B / C structure and can run stably through web pages after release.

[0109] This application proposes a map service recommendation method based on image content, and develops a map service retrieval platform based on keyword search and image content recommendation. The recommended layer has a high similarity with the comparison layer, which verifies the rationality and effectiveness of the method.

[0110] 1. Content-based adaptive sampling method for map service preview

[0111] Most of the existing map services exist in the form of map tiles, which are small-sized map images pre-rendered by the server and segmented according to the quadtree rule. For a network map service with n levels of zoom ratios, the smallest zoom ratio L0 only contains one tile covering the entire map range, and the largest zoom ratio L n-1 The level contains 4n-1 map tiles. The size of map tiles is close to the requirement of thumbnails, so there is no need to compress them. They can also be flexibly spliced. Therefore, some network map tiles are used as preview images. In tile map mode, the map service is simply defined as selecting a number of tiles that can represent the map service content from a large number of map tiles, which is further decomposed into two sub-modules: tile selection in the plane position dimension and tile selection in the scaling dimension.

[0112] (1) Tile selection in the plane position dimension: Based on a specific zoom ratio, the representativeness of the map tile is quantified. For a static map, people tend to pay more attention to the more complex parts, such as areas with densely distributed symbols or compactly placed annotations, because such areas can provide richer information. Based on this idea, this application uses the complexity factor of the map tile to approximate its information content and as a quantitative indicator of tile representativeness;

[0113] (2) Tile selection in the scaling dimension: The key is to find tiles with mutations. Map tiles of different scales have both similarities and differences. More similarities indicate subtle changes between scales; more differences indicate significant changes between scales. Subtle and significant correspond to slow changes and mutations in the map scale space, while mutations are more likely to attract user attention and should be included in the map service thumbnail. This application uses image similarity to quantitatively evaluate the similarity between tiles of two scales, and uses this as a basis to determine the scaling ratio that has undergone significant changes.

[0114] Based on the two quantitative indicators of complexity and approximation, two rounds of screening are used to achieve full coverage of the plane position dimension and automatic selection of multi-scale tiles in the scaling dimension:

[0115] In the first round of screening, starting from the smallest scale, only the most complex tiles are selected at each level, and the geographical range of the tile is used as a constraint to process the next scale. The first round of screening filters out most tiles;

[0116] Second round of screening: Compare the similarity of tiles between scaling ratios, further remove redundancy, and define the sampling set as R. The specific steps are as follows:

[0117] Step 1: Starting from level L0, add the only map tile in level L0 to R and use it as the seed tile;

[0118] Step 2: Request four higher-level map tiles with the same geographical range as the seed tile, and use the tile Ringo with the richest information as the new seed tile;

[0119] Step 3: Combine the four tiles to get Rpvda. If the similarity between Rpvda and the seed tile is greater than the critical value, add Rpvda to R, otherwise do not add it.

[0120] Repeat steps 2 and 3 until all scaling ratios are traversed to obtain the final set of sampling images;

[0121] Flexible adjustment of multiple variables: The number of tiles with adjacent scaling ratios does not necessarily satisfy the 4-fold relationship, which will affect the number of tiles to be stitched together. The first r (r≥1) tiles with the most information are used as seed tiles to increase the probability of obtaining representative tiles.

[0122] (I) Calculation and evaluation of map tile information

[0123] From coarse to detailed, the number of map tiles at each zoom level increases exponentially. If all tiles are compared, the amount of calculation is very large. This application performs the first round of screening based on the amount of map information. At each zoom level, the map tile with the richest information is selected as the seed tile, and the geographic space covered by the seed tile is used as a constraint to heuristically traverse the next zoom level. Except for L0 and L1 levels, it cannot be guaranteed that the seed tiles at other levels have the maximum amount of information, but a higher hit rate can be obtained at a lower computational cost. After coarse selection, the number of map tiles involved in the comparison at each level does not exceed 4, and the amount of calculation is greatly reduced;

[0124] In terms of map information calculation and evaluation, most existing technologies are based on statistical analysis of vector targets and are not suitable for information calculation of map tiles (raster images). Compared with natural scene images, maps contain regular text and map symbols, with a small number of colors, sharp regional boundaries, and high homogeneity of color blocks. A high compression ratio can be obtained by using lossless compression methods. This application uses the amount of data of map tiles after compression as a characterization indicator of their information volume. Map tiles are mostly in PNG format, which is very suitable for processing homogeneous areas in images and has a high compression ratio, which can better meet the compression requirements of maps. Figure 1 The data volume of PNG images outputted by map symbols of different complexity is listed. The data volume of square symbols, dot symbols, ring symbols and color symbols increases in sequence, indicating that PNG is very sensitive to image changes caused by visual variables such as shape, texture and color, and can reflect the complexity of map content. Figure 2An example of locating a complex area of ​​a map based on PNG data volume is given. Each image is composed of four tiles. The dotted box indicates the tile with the largest data volume and requests the next zoom ratio based on this range. From the three zoom processes, it can be concluded that the data volume of the PNG image can better reflect the complexity factor of the map content. Therefore, this application uses the data volume of the map tile in PNG format as the information volume indicator.

[0125] (II) Calculating cross-scale map tile approximation

[0126] The screening is performed within each zoom ratio based on the amount of information, and the relationship between different zoom ratios is not considered. After screening by the amount of information, each zoom ratio obtains a map tile, forming a map tile set. This set may contain several very similar tiles, and there is expression redundancy when used as a preview image. This application compares the similarity of map tiles between two adjacent zoom ratios, and judges sudden changes or slow changes based on this, and only retains map tiles that have changed significantly. Although maps with different zoom ratios differ in the degree of content detail and symbol style, the overall layout and distribution pattern of features are similar, which reduces the difficulty of calculating tile similarity to a certain extent. As an important visual variable of map symbols, color features play a decisive role in human map perception and cognition. Therefore, starting from color features, the similarity of map tiles is analyzed. This application uses a color quantization unified histogram to quantify the color features of the map, and on this basis, the similarity of map tiles is measured based on the cosine distance weight.

[0127] This application combines map expression features to make two improvements to the histogram generation method:

[0128] Improvement 1: Color quantization: To ensure readability and overview, the types of symbols contained in the map are limited, and the number of corresponding colors is much lower than that of natural scene images. However, due to the application of anti-aliasing and other technologies, the edge pixels of map symbols introduce color gradients, which indirectly increases the total amount of colors. Figure 3 As shown in Figure 1, the map tiles in A contain a total of 8238 colors. Therefore, the map tiles are first color quantized. Figure 3 B, C and D are the quantization results of A with different color quantities. It can be seen that the tile only needs 16 colors to accurately express. If the two tiles are quantized separately, the obtained color sets may be inconsistent. In the subsequent histogram distance calculation, the similarity of the colors themselves needs to be considered, and the calculation complexity is high. Here, the two tiles to be compared are first combined together, all the colors of the two are uniformly managed, and quantized to obtain a significant and unique color set;

[0129] Improvement 2, calculate the frequency of occurrence of colors: The conventional method is to count the number of pixels belonging to a certain color. This method cannot reasonably reflect the composition of map content, because map symbols, especially linear symbols and point symbols, usually express important information with a small number of pixels. The histogram based on the number of pixels will greatly weaken the status of these two types of symbols. To address this problem, this application first divides the map into several continuous color blocks, and uses the perimeter of these color blocks to calculate the frequency of occurrence of the corresponding colors. The division standard of continuous color blocks is that pixels of the same color are adjacent in four neighborhoods, and the perimeter calculation adopts the Moore neighborhood boundary tracking method.

[0130] In the RGB color space, color quantization uses the minimum variance method to specify the number of colors to be retained. Based on the quantized color set, the frequency of occurrence of each color in the two tiles is counted to obtain a unified histogram of color quantization, and finally the cosine distance is used to calculate the similarity between the histograms. According to the histograms of the two tiles when the color volume is 16, the approximation is 0.172 at this time. The color volume will affect the value of the approximation. After many experimental attempts, it was found that when the number of colors is more than the actual colors appearing in the map, the approximation gradually tends to stabilize. Therefore, subsequent experiments set a relatively high color volume. In addition, thematic maps often contain large areas of blank areas or background color fillings. These areas with excessively large weight values ​​will weaken the status of other map symbols, thereby interfering with the approximation index. This application makes adjustments during the approximation calculation, setting the color weight value with the largest proportion in the histogram to 0 to suppress its influence on the approximation.

[0131] 3. Analysis of the adaptive sampling experiment of map service preview

[0132] The experiment selected three map services to verify the method, including the public map service and two thematic map services (medical facility distribution map and river level map). First, the map tiles of the three services at the minimum zoom level (i.e., level L0) were obtained as the initial seed tiles. From the representative map tiles of the 10-level zoom level automatically obtained, it can be concluded that the public map services finally located the densely populated building areas with rich content. The entire zoom process better reflects the content changes of the map services. The map tiles obtained by the two thematic map services at the maximum zoom level have less content, but from the previous zoom levels, the positioning of complex areas is accurate.

[0133] Color quantization is performed with 32 colors, and then the approximation of tiles between zoom ratios is calculated. Overall, the two types of map services show different rules during the zooming process: the mutation of public map services occurs in the later stage, due to the introduction of elements such as blocks, buildings, and green spaces, the content of the map changes significantly; the mutation of thematic map services occurs in the early stage, due to the alleviation of symbol conflicts, the selection of element targets, the conversion of expression models, and the replacement of symbol styles. After repeated experiments and judgments, this application sets 0.1 as the critical value for thumbnail screening. If the approximation is greater than 0.1, it is considered that a mutation has occurred and it will be included in the thumbnail collection; otherwise, it will not be used as a thumbnail.

[0134] 2. Color and texture driven image cluster map recommendation

[0135] Map images are classified according to image features. Before layer recommendation, the layers are clustered to narrow the search range of similar layers and improve search efficiency. The color quantization unified histogram and grayscale co-occurrence matrix are used to represent the color and texture features of the image respectively. Based on the generation of 3 sampling images for each layer, this application proposes an image group clustering recommendation method for network map service characteristics.

[0136] 1. Collaborative representation of map features based on integrated color and texture features

[0137] For a color map, its color feature and texture feature are denoted as G s and G r , for two given images D and E, the color feature difference between them is recorded as A s , the texture feature difference is A r , k1 and k2 are the weight values ​​of the distance between the two features, then the comprehensive distance between D and E is:

[0138] A(D,E)=k1*A s (G SD , G SE )+k2*A r (G rD , G rE )

[0139] This is used as a similarity comparison method for map images and retrieval.

[0140] 2. Map content image clustering

[0141] At the beginning of clustering, all objects in the map data set are used as cluster centers. In each iteration, each cluster center is updated through the message passing mode. After each iteration, the object that failed in the competition selects a cluster center as its center. Through continuous iteration and selection, multiple high-quality clustering results are finally obtained, avoiding the influence of the initial random cluster center in the division method on the final result, and the result is very stable.

[0142] Clustering is performed based on the approximation between M map data points. This approximation constitutes an M×M similarity matrix. All map data points are regarded as potential cluster centers Z. The larger the value c(w, w) on the diagonal of the C matrix, the greater the probability that point w will become a cluster center. Its value is the reference degree q. The transmission mechanism mainly includes attraction factors and attribution factors. The attraction factor and attribution factor matrices are updated. To avoid result oscillation, the attenuation coefficient k is introduced. When updating information, each information is set to k times the last iteration update value plus 1-k times the information update value, where the attenuation coefficient k is a real number between 0 and 1. That is, the iterative value of t(j, w) and d(j, w) for the r+1th time:

[0143] d r+1 (j, w)←(1-k)d r+1 (j,w)+kd r (j, w)

[0144] Specific steps for map content image clustering:

[0145] Step 1: Initialize parameters and read map content data;

[0146] Step 2: Calculate the proximity matrix, find the median value of the proximity matrix and assign it to the reference degree q;

[0147] Step 3: Update the attraction factor matrix and the attribution factor matrix;

[0148] Step 4: Determine whether the maximum number of iterations has been reached or the termination condition has been met. If so, go to step 2, otherwise go to step 5;

[0149] Step 5: Get the final cluster center Z and divide each data object into the corresponding cluster.

[0150] (III) Clustering method based on image groups

[0151] This application selects 2355 WMS layers and adopts a content-based adaptive sampling method for map service preview to generate three preview images for each layer. These three preview images express the changes in information at different zoom ratios of this layer. Whether it is clustering or layer recommendation, it is necessary to integrate the three preview images of each layer.

[0152] The three preview images of each layer are taken as an image group. The image groups of each layer are clustered. First, the features of the three preview images in the image group are analyzed. The three images come from different zoom levels of the same layer. Layers of different zoom levels express the same content of the layer at different scales, and the color expression of the same object at different scales will not change. Most of the three images in the image group have similar color features, such as Figure 4 shown.

[0153] When clustering image groups according to color features, the color features of the image groups are comprehensively considered. Suppose the color quantization unified histograms of the three images A, B, and C are R A ,R B ,R C Due to the similarity of the color features of the images in the image group, the color feature of the image group is R = R A +R B +R C , the calculation takes into account the characteristics of the images in the image group and comprehensively considers the color characteristics of the image group.

[0154] (IV) Map service recommendation based on image group

[0155] Similar layer recommendation based on image content comprehensively considers the three preview images of the layer. On this basis, this application proposes a similar service recommendation method based on map service sampling images.

[0156] Suppose there are layers D and E, each layer has 3 preview images, namely D1, D2, D3 and E1, E2, E3. Calculate the similarity between the sampling image groups of layer D and layer E.

[0157] Considering that the three images of each layer are from different zoom levels, even if layer D is similar to layer E (the layers at each zoom level are very similar), during the image sampling and screening process, the local information of a single level is different, and the three images of each layer are likely to be sampled from different areas or at different zoom levels, which is likely to lead to lower similarity, which will lead to lower similarity between this set of images.

[0158] Considering that the three images of each layer reduce the low similarity between two images caused by similar services being located at different levels or different areas, the following image group similarity calculation method is proposed:

[0159] Step 1: Calculate the image similarity between image groups D1, D2, D3 and image groups E1, E2, E3 respectively, using the color plus texture feature calculation method. The similarities are C ji , where j = {1, 2, 3} is D j The subscript of i={1,2,3} is E i Subscript, assuming Cji The maximum value in is C 11 , then among these 9 groups of comparisons, D1 and E1 are the most similar, and C1=C 11 ,like Figure 5 As shown, at this time, exclude the i and E i All groups of C 22 , C 23 , C 32 , C 33 ;

[0160] Step 2: Similarly, C 22 , C 23 , C 32 , C 33 Find the maximum value among the four groups and set it as C 32 , let C2=C 32 ,like Figure 6 As shown, exclude image D3 and image E2, and the remaining group of images is D2 and E3, and C3 = C 23 ;

[0161] Step 3: After the first and second steps, the three most similar image combinations are found between the two image groups. To reduce the impact of different zoom levels and different positions on the similarity, a different weight value k is set for each group of similarities. The final similarity between the image groups is:

[0162] C=k1×C1+k2×C2+k3×C3

[0163] Among them, 1≥k1>k2>k3≥0.

[0164] (V) Experimental analysis of image clustering map recommendation

[0165] Based on the results of map content image clustering, the image group similarity measurement method proposed in this application was used to conduct similarity recommendation experiments. The experimental results are as follows:

[0166] Experiment 1: Image Group Figure 7 The recommended results are as follows Figure 8 As shown, from Experiment 1, it can be concluded that the input data is map base map data, and the recommended results are all map layers. Therefore, it can be seen that the image group similarity calculation method proposed in this application has a good recommendation result for map surface layers. The recommendation result takes into account the color characteristics and texture characteristics of the image group.

[0167] Experiment 2: Image Group Fig. 9 The recommended results are as follows Fig.10As shown in the figure, the input image of Experiment 2 is a complex image with complex textures and complex colors. The recommended results are all images with similar textures, and those with similar colors are ranked at the top. This shows that the image group similarity measurement method proposed in this application has a good effect on complex images.

[0168] Experiment 3: Image Group Fig.11 The recommended results are as follows Fig.12 As shown, in Experiment 3, the input is a layer of simple linear symbols, and the recommended results are all linear layers, and linear image groups with similar colors have high similarity, and linear image groups with large color differences are ranked later. This shows that the image group similarity measurement method proposed in this application has good recommendations for image groups with linear textures.

[0169] The above three groups of experiments show that the image group recommendation method for network map service features proposed in this application comprehensively considers the color and texture characteristics of the image group, and has good recommendation results for different types of land objects.

[0170] 3. Build a network map intelligent retrieval and recommendation platform

[0171] It includes geographic information service data collection, data processing, design and construction of keyword retrieval library and image feature retrieval, construction of a web-based network map service keyword retrieval and similar service recommendation platform, crawling a large number of map services through web crawlers, parsing their metadata, storing them in the database and establishing a keyword retrieval library, publishing keyword retrieval services based on Apache Solr, sampling representative images for each layer, and then calculating the color and texture features of each image to build a feature library, and storing the clustering results in the category library through clustering. When similar images need to be recommended, the system first finds the category to which they belong, and then finds the layer to which the most similar image belongs under this category and recommends it to the user.

[0172] (I) Map intelligent retrieval and recommendation platform architecture

[0173] By building a web crawler to collect web map services scattered on various nodes of the Internet, the collected structured web map service data is parsed and stored in the database as metadata for display to users. In addition, through the layer acquisition interface provided by the service, combined with the automatic generation method of web map service thumbnails, thumbnails are generated for each layer. These thumbnails, as representatives of service content, are the data basis for subsequent image feature extraction and image clustering. By building a retrieval recommendation platform, it is conducive to the rapid retrieval of web map services and the acquisition of related map services, thereby improving the ability to obtain information. Based on this, this application proposes the following Fig.13 The map intelligent retrieval and recommendation platform framework is shown.

[0174] The map intelligent retrieval recommendation platform framework is divided into three layers:

[0175] Layer 1: Web map service collection and preprocessing layer: By compiling web map service theme crawlers, we can crawl scattered map services on the Internet, and then download and parse their metadata according to their service versions and specifications. Based on the structured organization form, we can establish a database table structure according to its various parameter information, and store the processed data in the MySQL database.

[0176] Layer 2: Image preprocessing and keyword search library construction: This layer is divided into two parts, one for the platform's keyword search and the other for similar service recommendations; for keyword search, a directional index library from keywords to services is established; for similar service recommendations, thumbnails are generated for each service; this application has reserved the color and texture features of the service. Since the feature extraction calculation is relatively complex, its features are calculated in advance in the background and stored in the feature library to facilitate subsequent clustering;

[0177] Layer 3: Platform query and recommendation logic implementation: Implement the background logic of the query recommendation platform, mainly including: map service query and display based on keywords, and similar service recommendations based on thumbnail content.

[0178] (II) Network map data collection and processing

[0179] The geographic information service theme crawler is developed to collect OGC WMS and ArcGIS MapService on the Internet. The idea and logic of the geographic information service theme crawler are as follows:

[0180] (1) Obtain the seed point URL using keywords through a search engine;

[0181] (2) Read the seed URL webpage by compiling a crawler program and extract other URL links obtained from the webpage;

[0182] (3) Parse the web page content. If the page contains the keywords WMS, web map service, or ARCGIS REST Services, it will be retained as a URL to be processed.

[0183] (4) If the URL link contains the keywords service=wms&request=getcapbilities and the page contains wms-capbilities, it is determined to be an OGC WMS service; if the link contains / arcgis / rest / services and the page content contains ArcGIS REST Services Directory, it is determined to be an ArcGISService service and the service link is stored in the database.

[0184] The captured services are stored in the database in linked form. To obtain the metadata of the services, web pages are requested and metadata parsed according to the service publishing specification type and its service version. Since the service information to be displayed is limited, this application only selects the more important information for storage.

[0185] (III) Map keyword search library

[0186] If the query of keyword-based service is simply matched from the database using SQL statements, the speed is very slow and it is easy to get stuck in high concurrency, which is inefficient. Therefore, in order to solve these problems and improve efficiency, a reverse index library of keyword-based service is established, a mapping of keyword-based service layer is established, and the description information and title fields of each service layer are selected as the keyword matching fields. The specific steps of establishing a keyword index are as follows: Fig.14 shown.

[0187] (1) Treat each layer of WMS and ArcGIS Map Service as a document, select the Title and Description in each layer as the fields to be segmented, and pass the information to be indexed to the Tokenizer component to obtain the word-unit Token information of each document by removing punctuation and pauses;

[0188] (2) The language analysis tool is then used to process the previously obtained lemmas, including converting uppercase letters to lowercase letters and converting single words to root forms, to obtain the word Term.

[0189] (3) Pass the term to the indexer, create a word dictionary with the indexed terms and sort them in alphabetical order, merge the same terms into a posting list;

[0190] This application targets two map services, OGC WMS and ArcGIS MapService, in order to perform keyword indexing on the service layer information and find appropriate service description information in its metadata.

[0191] The search service is established by creating a word-based inverted index table. English words are separated by spaces and punctuation marks. The search results display information including: layer title, layer description information, layer name, and sampling map connection. The corresponding fields are mapRiRle, mapAesc, mapMame, and images.

[0192] Keyword search service release: To relieve server pressure under high concurrency and improve system stability, keyword search is released as a separate service, decoupling the keyword search service from the entire system. This not only improves system stability but also facilitates later system maintenance. Tomcat is selected as the publishing server for the Solr service. After the service is successfully released, the information in Solr can be viewed through a visual web page. The search results can be queried by configuring query conditions on the page. The service is integrated into the platform using the SolrJ tool to call the Solr search service.

[0193] (IV) Map image preprocessing and image retrieval

[0194] Generate sampling maps for OGC WMS, refer to their respective image acquisition interfaces, and then preview the adaptive sampling method based on the content of the map service to generate thumbnails for the layers of each service. For a recommendation system, an important indicator for evaluating its performance is the retrieval speed. The image resource data volume is large, and there are many images in the image library. The sequential retrieval efficiency is low. If similar images are clustered into one category, when recommending similar layers, first find the class where the image group of this layer is located, and then only search within this class, which can greatly narrow the layer search range.

[0195] After extracting the image feature database, clustering operations are performed on the color and texture features of the image to establish an image classification table. The images in each class are a collection of similar images. In the process of keyword query, the thumbnail of the layer selected by the user is automatically used as the input image. The category to which this input image belongs is directly searched in the clustering result library, and then a search is performed within this category to find the most similar image or the image that meets the specified conditions and recommend it to the user.

[0196] The map image preprocessing and image retrieval process includes Fig.15 process.

[0197] (V) Construction and access of the network map intelligent retrieval and recommendation platform

[0198] The platform is based on B / S architecture and uses Java technology stack as the implementation technology of B / S architecture. It integrates network map service management, keyword-based service retrieval, and image content-based service recommendation, providing a technical solution that integrates all services. The platform front-end uses Bootstrap technology to build display pages, and the back-end service is developed using JavaEE's SpringMVC+Spring+MyBatis integration framework. Tomcat is the back-end server, and Solr and Tomcat servers are used to publish retrieval services. A keyword query interface is provided for back-end calls, and MySQL database is used to store map service metadata information. Redis is used as the cache database to cache service metadata information.

[0199] (1) Database and Redis cache

[0200] Each map service contains multiple layers of information, ranging from a few to thousands. These layers of information are fixed. If the layers under the same service are retrieved from MySQL for display each time the service information is displayed, the efficiency is bound to be very low. Generating a layer information cache for each service is the solution of this application.

[0201] The specific timing and method of cache generation are as follows: when accessing this service for the first time, find the service description information and the layer information it contains from MySQL and store this information in Redis in json format; when accessing it for the second time or later, directly get the service information corresponding to this service from Redis to improve the page response speed.

[0202] (2) Map service publishing and access

[0203] The website backend server uses Tomcat, and the main functions are demonstrated as follows: website entry page, search by entering keywords.

Claims

1. A content-adaptive web map intelligent recommendation system, characterized in that: Integrate image content-based clustering and image recommendation into the retrieval and recommendation of network map services. From data collection and processing, thumbnail generation, image clustering, to platform construction, a complete set of technical processes are built. On this basis, a web-based map service retrieval and recommendation platform is designed and implemented, including: First, an automatic sampling method for web map services based on information volume and similarity includes: first, calculating and evaluating the information volume of map tiles, and second, calculating the similarity of cross-scale map tiles; using information volume combined with a quadtree structure to locate content-rich areas in the map, and the information volume is approximately expressed by image complexity; the similarity is calculated by a color quantized unified histogram to screen the scale ratio that has undergone mutations. The two methods realize the orderly traversal of map services from the plane position dimension and the scale ratio dimension respectively, providing effective data for clustering recommendations based on image content; Second, color and texture driven image clustering map recommendation, including: first, map feature collaborative representation of integrated color and texture features, second, map content image clustering, third, clustering method based on image group, fourth, map service recommendation based on image group; color features and texture features are used to comprehensively represent image features. Color features are represented by color quantization unified histogram, and texture features are represented by grayscale co-occurrence matrix. The inclusion relationship and color similarity of image groups are considered. The image group recommendation method reduces the proximity difference caused by layer positioning by assigning different weight values ​​to different proximity groups. A multi-feature representation method of integrated color and texture is proposed to realize map service recommendation based on image groups. Third, build an intelligent retrieval and recommendation platform for online maps, including:

1. The map intelligent retrieval and recommendation platform architecture, 2. The network map data collection and processing, 3. The map keyword retrieval library, 4. The map image preprocessing and image retrieval, 5. The network map intelligent retrieval and recommendation platform construction and access; crawl a large number of map services by compiling web crawlers, parse their metadata and store them in the database, establish a keyword retrieval library based on the Apache Solr service and publish the keyword retrieval service, generate thumbnails for each service and store them, calculate the color and texture features of each thumbnail and store them in the feature library, store the clustering results in the category library, and when recommending similar images, first find the category to which it belongs and then find the most similar image under this category and recommend it to the user.

2. According to the content-adaptive-aware web map intelligent recommendation system of claim 1, it is characterized by: Content-based adaptive sampling method for map service preview: some network map tiles are used as preview images. In tile map mode, the map service is simply defined as selecting a number of tiles that can represent the map service content from a large number of map tiles, which is further decomposed into two sub-modules: tile selection in the plane position dimension and tile selection in the scaling dimension; (1) Tile selection in the plane position dimension: Based on a specific scaling ratio, the representativeness of the map tile is quantified. The complexity factor of the map tile is used to approximate its information content and is used as a quantitative indicator of tile representativeness. (2) Tile selection in the scaling dimension: The key is to find the tile where mutation occurs. Image approximation is used to quantitatively evaluate the similarity between two scale tiles, and based on this, the scaling ratio where significant changes occur is determined. Based on the two quantitative indicators of complexity and approximation, two rounds of screening are used to achieve full coverage of the plane position dimension and automatic selection of multi-scale tiles in the scaling dimension: In the first round of screening, starting from the smallest scale, only the most complex tiles are selected at each level, and the geographical range of the tile is used as a constraint to process the next scale. The first round of screening filters out most tiles; Second round of screening: Compare the similarity of tiles between scaling ratios, further remove redundancy, and define the sampling set as R. The specific steps are as follows: Step 1: Starting from level L0, add the only map tile in level L0 to R and use it as the seed tile; Step 2: Request four higher-level map tiles with the same geographical range as the seed tile, and use the tile Ringo with the richest information as the new seed tile; Step 3: Combine the four tiles to get Rpvda. If the similarity between Rpvda and the seed tile is greater than the critical value, add Rpvda to R, otherwise do not add it. Repeat steps 2 and 3 until all scaling ratios are traversed to obtain the final set of sampling images; Flexible adjustment of multiple variables: The number of tiles with adjacent scaling ratios does not necessarily satisfy the 4-fold relationship, which affects the number of tiles to be stitched together. The first r (r≥1) tiles with the most information are used as seed tiles to increase the probability of obtaining representative tiles.

3. According to claim 2, the content-adaptive-aware web map intelligent recommendation system is characterized in that: Evaluation of map tile information volume calculation: The first round of screening is based on the map information volume. The map tile with the richest information volume is selected as the seed tile at each zoom ratio, and the geographic space range covered by the seed tile is used as a constraint to heuristically traverse the next zoom ratio. Except for L0 and L1 levels, it is not guaranteed that the seed tiles of other levels have the maximum information volume, but a higher hit rate is obtained at a lower computational cost. After rough selection, no more than 4 map tiles are involved in the comparison at each level; In terms of the calculation and evaluation of map information volume, the data volume of the compressed map tiles is used as the representation indicator of its information volume, and the data volume of the map tiles in PNG format is used as the information volume indicator.

4. The content-adaptive-aware web map intelligent recommendation system according to claim 1 is characterized in that: Calculate the similarity of cross-scale map tiles: compare the similarity of map tiles between two adjacent zoom ratios, and judge sudden or slow changes based on this, and only retain map tiles that have changed significantly; analyze the similarity of map tiles based on color features, use color quantization unified histogram to quantify the color features of the map, and on this basis measure the similarity of map tiles based on cosine distance weight; Combining the map expression features, we make two improvements to the histogram generation method: Improvement 1, for color quantization: First, quantize the colors of the map tiles. First, combine the two tiles to be compared together, unify all the colors of the two tiles, and quantize to obtain a significant and unique color set; Improvement 2, calculate the frequency of color occurrence: first divide the map into several continuous color blocks, and use the perimeter of these color blocks to calculate the frequency of occurrence of the corresponding color. The division standard of continuous color blocks is that pixels of the same color are adjacent in four neighborhoods. The perimeter calculation adopts Moore neighborhood boundary tracking method; In RGB color space, color quantization uses the minimum variance method to specify the number of colors to be retained. Based on the quantized color set, the frequency of occurrence of each color in the two tiles is counted to obtain a color quantization unified histogram. Finally, the cosine distance is used to calculate the similarity between the histograms. Set the color weight value with the largest proportion in the histogram to 0 to suppress its influence on the approximation.

5. The content-adaptive-aware web map intelligent recommendation system according to claim 1 is characterized in that: Color and texture driven image clustering map recommendation: map images are classified according to image features. Before recommending layers, the layers are clustered to narrow the search range of similar layers. The color quantization unified histogram and gray level co-occurrence matrix are used to represent the color and texture features of the image. On the basis of generating 3 sampling images for each layer, image group clustering recommendation is performed based on the network map service features. Collaborative representation of map features based on color and texture features: For a color map, its color features and texture features are denoted as G s and G r , for two given images D and E, the color feature difference between them is recorded as A s , the texture feature difference is A r , k1 and k2 are the weight values ​​of the distance between the two features, then the comprehensive distance between D and E is: A(D,E)=k1*A s (G SD ,G SE )+k2*A r (G rD ,G rE ) This is used as a similarity comparison method for map images and retrieval.

6. The content-adaptive-aware web map intelligent recommendation system according to claim 1 is characterized in that: Map content image clustering: At the beginning of clustering, all objects in the map data set are used as cluster centers. In each iteration, each cluster center is updated through the message passing mode. After each iteration, the object that failed in the competition selects a cluster center as its center. Through continuous iteration and selection, multiple high-quality clustering results are finally obtained. Clustering is performed based on the approximation between M map data points. This approximation constitutes an M×M similarity matrix. All map data points are regarded as potential cluster centers Z. The larger the value c(w, w) on the diagonal of the C matrix, the greater the probability that point w will become a cluster center. Its value is the reference degree q. The transmission mechanism includes attraction factor and attribution factor. The attraction factor and attribution factor matrices are updated. To avoid result oscillation, an attenuation coefficient k is introduced. When updating information, each information is set to k times the last iterative update value plus 1-k times the information update value. The attenuation coefficient k is a real number between 0 and 1, that is, the iterative value of t(j, w) and d(j, w) for the r+1th time: d r+1 (j,w)←(1-k)d r+1 (j,w)+kd r (j,w) Specific steps for map content image clustering: Step 1: Initialize parameters and read map content data; Step 2: Calculate the proximity matrix, find the median value of the proximity matrix and assign it to the reference degree q; Step 3: Update the attraction factor matrix and the attribution factor matrix; Step 4: Determine whether the maximum number of iterations has been reached or the termination condition has been met. If so, go to step 2, otherwise go to step 5; Step 5: Get the final cluster center Z and divide each data object into the corresponding cluster.

7. The content-adaptive-aware web map intelligent recommendation system according to claim 1 is characterized in that: Clustering method based on image groups: Adopting the content-based adaptive sampling method for map service preview, three preview images are generated for each layer. These three preview images express the changes in information of different zoom ratios of this layer. Whether it is clustering or layer recommendation, it is necessary to integrate the three preview images of each layer; The three preview images of each layer are taken as an image group. The image groups of each layer are clustered. First, the features of the three preview images in the image group are analyzed. The three images come from different zoom levels of the same layer. Layers at different zoom levels express the same content of the layer at different scales, and the color expression of the same object at different scales will not change. When clustering image groups according to color features, the color features of the image groups are comprehensively considered. Suppose the color quantization unified histograms of the three images A, B, and C are R A ,R B ,R C Due to the similarity of the color features of the images in the image group, the color feature of the image group is R = R A +R B +R C , the calculation takes into account the characteristics of the images in the image group and comprehensively considers the color characteristics of the image group.

8. The content-adaptive-aware web map intelligent recommendation system according to claim 1 is characterized in that: Map service recommendation based on image groups: Similar layer recommendation based on image content comprehensively considers the three preview images of the layer, and on this basis, a similar service recommendation method based on map service sampling images is proposed; Suppose there are layers D and E, each layer has 3 preview images, namely D1, D2, D3 and E1, E2, E3, calculate the similarity between the sampled image groups of layer D and layer E; considering that the 3 images of each layer reduce the low similarity of the two images caused by similar services being located at different levels or different areas, the following image group similarity calculation method is proposed: Step 1: Calculate the image similarity between image groups D1, D2, D3 and image groups E1, E2, E3 respectively, using the color plus texture feature calculation method. The similarities are C ji , where j = {1, 2, 3} is D j The subscript of i={1,2,3} is E i Subscript, assuming C ji The maximum value in is C 11 , then among these 9 groups of comparisons, D1 and E1 are the most similar, and C1=C 11 , excluding those containing D i and E i All groups of C 22 , C 23 , C 32 , C 33 ; Step 2: Similarly, C 22 , C 23 , C 32 , C 33 Find the maximum value among the four groups and set it as C 32 , let C2=C 32 , excluding images D3 and E2, now only the group of images D2 and E3 is left, let C3=C 23 ; Step 3: After the first and second steps, the three most similar image combinations are found between the two image groups. To reduce the impact of different zoom levels and different positions on the similarity, a different weight value k is set for each group of similarities. The final similarity between the image groups is: C=k1×C1+k2×C2+k3×C3 Among them, 1≥k1>k2>k3≥0.

9. The content-adaptive-aware web map intelligent recommendation system according to claim 1, characterized in that: Build an intelligent retrieval and recommendation platform for web maps: including geographic information service data collection, data processing, design and construction of keyword retrieval library and image feature retrieval, build a web-based web map service keyword retrieval and similar service recommendation platform, crawl a large number of map services through web crawlers, parse their metadata, store them in the database and establish a keyword retrieval library, publish keyword retrieval services based on Apache Solr, sample representative images for each layer, and then calculate the color and texture features of each image to build a feature library, and store the clustering results in the category library through clustering. When similar images need to be recommended, the system first finds the category to which they belong, and then finds the layer to which the most similar image belongs under this category and recommends it to the user; (I) Map intelligent retrieval and recommendation platform architecture By building a web crawler to collect web map services scattered on various nodes of the Internet, the collected structured web map service data is parsed and stored in the database as metadata for display to users. In addition, through the layer acquisition interface provided by the service, combined with the automatic generation method of web map service thumbnails, thumbnails are generated for each layer. These thumbnails represent the service content and are the data basis for subsequent image feature extraction and image clustering. The map intelligent retrieval recommendation platform framework is divided into three layers: Layer 1: Web map service collection and preprocessing layer: By compiling web map service theme crawlers, we can crawl scattered map services on the Internet, and then download and parse their metadata according to their service versions and specifications. Based on the structured organization form, we can establish a database table structure according to its various parameter information, and store the processed data in the MySQL database. Layer 2: Image preprocessing and keyword retrieval library construction: This layer is divided into two parts, one for the platform's keyword retrieval and the other for similar service recommendations. For keyword retrieval, a directional index library from keywords to services is established. For similar service recommendations, thumbnails are generated for each service. The color and texture features of the booked services are calculated in advance in the background and stored in the feature library to facilitate subsequent clustering. Layer 3: Platform query and recommendation logic implementation: Implement the backend logic of the query recommendation platform, including: map service query and display based on keywords, and similar service recommendations based on thumbnail content; (II) Network map data collection and processing The geographic information service theme crawler is developed to collect OGC WMS and ArcGIS MapService on the Internet. The ideas and logic of the geographic information service theme crawler include: (1) Obtain the seed point URL using keywords through a search engine; (2) Read the seed URL webpage by compiling a crawler program and extract other URL links obtained from the webpage; (3) Parse the web page content. If the page contains the keywords WMS, web map service, or ARCGIS REST Services, it will be retained as a URL to be processed. (4) If the URL link contains the keyword service=wms&request=getcapbilities and the page contains wms-capbilities, it is determined to be an OGC WMS service; if the link contains / arcgis / rest / services and the page content contains ArcGIS REST Services Directory, it is determined to be an ArcGIS Service and the service link is stored in the database; The captured services are stored in the database in linked form, and the metadata of the services are obtained. According to the service publishing specification type and its service version, web pages are requested and metadata parsed, and only the more important information is selected and stored in the database.

10. The content-adaptive-aware web map intelligent recommendation system according to claim 9, characterized in that: (III) Map keyword search library: Establish a reverse index library of keywords to services, establish a mapping of keywords to service layers, and select the description information and title fields of each service layer as the fields for keyword matching; (1) Treat each layer of WMS and ArcGIS Map Service as a document, select the Title and Description in each layer as the fields to be segmented, and pass the information to be indexed to the Tokenizer component to obtain the word-unit Token information of each document by removing punctuation and pauses; (2) The language analysis tool is then used to process the previously obtained lemmas, including converting uppercase letters to lowercase letters and converting single words to root forms, to obtain the word Term. (3) Pass the term to the indexer, create a word dictionary with the indexed terms and sort them in alphabetical order, merge the same terms into a posting list; The objects are OGC WMS and ArcGIS MapService map services. The purpose is to perform keyword indexing on the service layer information and find the appropriate service description information in its metadata. The search service is established by establishing a word-based inverted index table. In English, words are separated by spaces and punctuation marks. The search results display information including: layer title, layer description information, layer name, and sampling map connection. The corresponding fields are mapRiRle, mapAesc, mapMame, and images. Keyword search service publishing: Publish keyword search as a separate service to decouple the keyword search service from the entire system. Select Tomcat as the publishing server for the Solr service. After the service is successfully published, view the information in Solr through a visual web page. Query the search results by configuring query conditions on the page. Integrate the service into the platform using the SolrJ tool to call the Solr search service.