TWI647580B - Search filtering method that enhances the matching of text search results - Google Patents

Abstract

The present invention relates to a search filtering method that improves the matching of search results of text images. It inputs a keyword in a search engine and analyzes whether the source article contains the keyword. If not, the source article is discarded. , Then keep the article as the preliminary search result, and then analyze whether the label of the picture carried in the article of the preliminary search result contains the keyword, if not, discard the article of the preliminary search result, if there is, keep it as the article of the final search result ; In this way, you can achieve the best results of the search results and graphics.

Description

Improve the search and filtering method to match the optimized search results

The present invention relates to a search filtering method, in particular to a method which can improve the degree of matching between the articles and pictures of search results and make the search data more accurate and not distorted.

With the vigorous development of information technology, the phenomenon of information explosion is increasing day by day. At the same time, it has changed the relationship between people looking for information, learning, interacting ... and so on. However, in tens of millions of articles or even more materials or webpages, all the relevant materials or webpages that meet the issues people care about are obtained through retrieval; and the current common retrieval methods all use full text Retrieval, that is, as long as the entered keywords or words are included in the article after entering the keywords or words, the search engine will sequentially display the articles in an itemized manner to obtain a search result.

In addition, although the search results returned by the search engine have data that matches the keywords or words, it does not necessarily mean that the returned search results are all the data that the user needs, so the user must further search the returned search The screening of data in the results not only caused confusion on the topic of interest in searching, but also was quite time-consuming.

Now, in view of the current search method for retrieving data, the present inventor cannot provide the data that best fits the queryer's desire to obtain troubles during the search and the time spent filtering data, and further developed the present invention.

The main object of the present invention is to provide an optimized matching degree between articles and pictures that can improve the search results, and make the search data more accurate and distortion-free. Filtering method.

A search filtering method for improving the matching of search results of text images includes the following steps: input at least one keyword in a search engine; analyze whether the source article contains the keyword, if not, discard the source article , If so, keep the article as the preliminary search result; analyze whether the label of the picture carried in the preliminary search result article contains the keyword, if not, discard the preliminary search result article, and if so, keep it It becomes the final search result article and displays the result on the display device.

As mentioned above, the search filtering method that improves the matching of search results of text images is optimized. In the process of parsing whether the keywords contained in the tags of the pictures carried in the articles of the preliminary search results include the keywords, the The Google Cloud Vision API accesses the tag information in the picture.

A search filtering method for improving the matching of text search results with optimization includes the steps of: creating a semantic database; entering at least one keyword in a search engine; selecting all the synonyms in the semantic database that are included in the keywords Word vocabulary; compare the article in a source database with all the synonyms of the keyword included in the keyword, and analyze whether all the synonyms of the keyword included in the keyword appear in the article, if not, discard The article, if any, keep the article as the preliminary search result article; analyze whether all the synonyms of the keywords contained in the keyword appear in the label of the picture carried in the preliminary search result article, if not, discard it The preliminary search result article, if any, retains the preliminary search result article to become the final search result article, and displays the title of the final search result article on the display device.

As mentioned above, the search filtering method for improving the search results of the text image is optimized, wherein the step of creating a semantic database includes the steps of importing a training library, and a solution. The steps of analyzing the tag content that may appear in each picture in the gallery, the step of collecting the tag content that may appear and storing in the semantic database.

As mentioned above, the search filtering method that improves the search results of the text image is optimized, wherein, in the step of parsing the label content that may appear in each picture in the gallery, multiple pictures in the gallery are first calculated The picture produces a total of k labels with a suitability of S _m , and m = 1, ..., k, and then the level of S _j determines whether to retain the label (S _j represents the suitability of figure j); suppose that each picture gets q The content of the label and its corresponding score (q appears in the range of j), the corresponding score is represented by Y _ij , where i is the picture number (i = 1, ..., n and j = 0, ..., q), when j is 0 The image does not produce any tags. When calculating, add up the total score of all k tags. If there are many occurrences and the quality score is high, the calculated S _m will be relatively high, so the tag represents that it can be retained To the semantic database, the calculation formula is as follows:

As mentioned above, the search filtering method that improves the search results of the text image is optimized, wherein the step of creating a semantic database includes the steps of inputting Chinese keywords, and using a translation application to key the input Chinese The steps of translating characters into English and storing them in the semantic database.

As mentioned above, the search filtering method that improves the search results of the text image matches the optimization, wherein the step of creating a semantic database further includes a step of expanding the semantic database, including It includes the steps of finding synonyms through the existing semantic database, and adding the synonyms to the semantic database for storage.

As mentioned above, the search filtering method that improves the matching of search results of text images is optimized. In the process of parsing whether the keywords contained in the tags of the pictures carried in the articles of the preliminary search results include the keywords, the The Google Cloud Vision API accesses the tag information in the picture.

(S11) ‧‧‧Step 1

(S12) ‧‧‧Step 2

(S13) ‧‧‧Step 3

(S21) ‧‧‧Step 1

(S22) ‧‧‧Step 2

(S23) ‧‧‧Step 3

(S31) ‧‧‧Step 1

(S32) ‧‧‧Step 2

(S33) ‧‧‧Step 3

(S41) ‧‧‧Step 1

(S42) ‧‧‧Step 2

(S51) ‧‧‧Step 1

(S52) ‧‧‧Step 2

(S53) ‧‧‧Step 3

(S54) ‧‧‧Step 4

The first picture: one of the results returned by the search for the keyword "Southern Taiwan Earthquake" in the Google image search engine

Second figure: The present invention relates to a flow chart of a search and filtering method for improving the matching of text search results

Figure 3: The flow chart of the first step to build a semantic database

Figure 4: Flow chart of the second step to build a semantic database

Figure 5: Steps to expand the semantic database of the existing semantic database

Figure 6: After the establishment of the semantic database, the implementation of the present invention is a flow chart of the steps of the search and filtering method for improving the matching of the search results of the text and images to perform the text and text retrieval

Figure 7: Upload a picture of a polar bear to the Google Cloud Vision API to identify the results of the tagged operation

Figure 8: Display the label analysis of one of the pictures after entering the keyword "Nantai Earthquake" and searching by the method of the present invention

The ninth figure: shows the label analysis of the two pictures obtained by entering the keyword "Nantai earthquake" and searching by the method of the present invention

Figure 10: Display the label analysis of the three pictures obtained by entering the keyword "Nantai Earthquake" and searching by the method of the present invention

Figure 11: A schematic diagram showing that the tagged data of the pictures contained in the article does not match the input of the keyword "Southern Taiwan Earthquake"

Figure 12: The second schematic diagram showing the tagging data of the pictures contained in the article that does not match the keyword of "Southern Taiwan Earthquake"

Figure 13: Three schematic diagrams showing that the tag data of the pictures contained in the article does not match the keyword of "Southern Taiwan Earthquake"

In order to provide a more complete and clear disclosure of the technical content, the purpose of the invention and the effect achieved by the present invention, the following is a detailed description, and please refer to the disclosed drawings and figures:

When searching for data with currently common search engines such as google, yahoo, Openfind, Bing ... etc., keywords are typed in the search engine, and these search engines list relevant information based on the keyword bar; however, Although the returned information contains the entered keywords and matches the search results, it does not replace the many data returned The table will be the information that the searcher needs. For example, when a searcher wants to collect data on earthquake relief in southern Taiwan by entering the keyword "Southern Taiwan Earthquake" in the Google search engine, the search results will also appear non-related in addition to articles on disaster relief and earthquakes. For the data, please refer to the first picture, which is one of the results returned by the search for the keyword "Nantai Earthquake" in the Google image search engine. From this example, you can see that the title of this news contains the key entered Word, but it is actually not directly related to the content of the disaster relief that the searcher wants to care about. The problem of this search gap is also common in image retrieval.

In order to improve the problem that traditional search engines are not accurate enough when searching for data, the present invention proposes a more effective and precise search method for graphic retrieval, which is to decide whether to discard the keywords entered in the search engine by filtering the source article The pen data may be retained, and then check whether the label of the picture appearing in the retained data contains the entered keyword, if the picture label does not contain the entered keyword, the pen data is discarded, if the picture label contains the The entered keyword retains the data and becomes one of the materials of the final search result.

Please refer to the second picture. That is, the present invention relates to a search filtering method that improves the matching of search results of text images. The method includes at least the following steps: Step 1 (S11): Enter at least one keyword in a search engine; Step 2 (S12): Analyze whether the source article contains the keyword, if not, discard the source article, and if so, keep it as the preliminary search result article; Step three (S13): analyze whether the label of the picture carried in the article of the preliminary search result contains the keyword, if not, discard the article of the preliminary search result, and if so, keep the article of the preliminary search result to become One of the final search result articles, and displays the title of the final search result article on the display device.

Since the annotations of the images generated by the Google Cloud Vision API will have many possibilities for change, and each person's text describing the image is also different, resulting in semantic gaps. To solve this problem, the present invention further provides a semantic database To solve this problem by creating a semantic database for the search before actually performing a graphic search.

The step of creating a semantic database is performed before step 1 (S11) enters at least one keyword in a search engine or step 2 (S12) analyzes whether the source article contains the keyword, and the step of creating a semantic database The methods can be divided into two types, namely: (1) importing a small number of pictures to create the required semantic database and (2) directly entering keywords to create the required semantic database.

Among them, please refer to the third figure for the first establishment method. The steps include: step one (S21): import into training library; step two (S22): analyze the content of labels that may appear in each picture in the library; in this step , the computing system of the first plurality of pictures in the gallery labels generated in total fitness k S _m, and m = 1, ..., k, then S _j is determined by the level of whether to retain the label (S _j represents the j FIG. The suitability of each image); suppose each image gets q label content and its corresponding score (q appears in the range of j), the corresponding score is represented by Y _ij , where i is the image number (i = 1, ..., n and j = 0, ..., q), when j is 0, it means that the image did not produce any tags, and the total score of all k tags is added in the calculation. If there are many occurrences and the quality score is high, the calculated S _m is It will be relatively high, so this label means that it can be reserved to the semantic database, and its calculation formula is as follows:

Step three (S23): Collect the possible label content and store it in the semantic database.

The second way to create a semantic database is shown in the fourth figure. The steps include: Step 1 (S31): Enter Chinese keywords; Step 2 (S32): Translate the entered Chinese keywords into English with a translation application ; Step three (S33): Collect the translated English and store it in the semantic database.

In addition, the data of the existing semantic database can also be used to expand the semantic database of the present invention. Please refer to the fifth figure. The steps include: Step 1 (S41): Through the existing semantic database (eg: WordNet (WordNet) )) Search for synonyms; step two (S42): add the synonyms to the semantic database for storage.

After the establishment of the semantic database, you can perform a graphic search (see Figure 6); first, after entering the keyword, the system will select the semantic database according to the entered keyword Include the synonyms of vocabulary, then compare the article with the synonyms of vocabulary (keywords) in the semantic database in a source database, if the article contains the required keywords (including synonyms vocabulary), go to Parse the label of the article picture. If the parsed label contains the required keywords (including synonyms words), the system will display the article, otherwise the article will be discarded.

Specifically, after the establishment of the semantic database, the step of searching and filtering the search results of the improved text image search results in accordance with the present invention includes: step one (S51): input at least one search engine One keyword; Step two (S52): Select all the synonyms of the keyword included in the keyword in the semantic database; Step three (S53): Perform an article in the source database with all the synonyms of the keyword included in the keyword Compare, analyze whether all the synonyms words contained in these keywords appear in the source article, if not, discard the source article, if there is, keep it as the preliminary search result article; Step 4 (S54): analyze the preliminary Whether all the synonyms words contained in the keyword appear in the label of the picture carried in the search result article, if not, the preliminary search result article is discarded, and if there is, the preliminary search result article is retained to become the final search One of the result articles, and the title of the final search result article is displayed on the display device.

In addition, the purpose of obtaining image annotations (Image Annotation) or labeling (Labelling) is to compare with the entered keywords to identify the text of the article. In addition to matching keyword content, the pictures are also It can match the theme of the corresponding keyword, and then improve the optimization results that match the search results of the text.

Tagging is an important mode for building machine learning. It needs to import a large number of training pictures and provide machine learning algorithms (Machine Learning algorithms) for learning. Especially if the pictures corresponding to the new retrieval conditions are not outdated, they cannot be recognized and then provided. Tags, so before tagging, a lot of training data is required. However, because the training machine takes a lot of time and cost, the present invention uses Google Cloud Vision API1 to effectively solve this problem, the system can provide pictures Labeling function.

The seventh picture is from ImageNet [J.Deng, W.Dong, R.Socher, L.-J.Li, K.Li, and L.Fei-Fei, "Imagenet: A large-scale hierarchical image database," in Computer Vision and Pattern Recognition, 2009.CVPR 2009.IEEE Conference on.IEEE, 2009, pp.248-255.] The picture of the polar bear provided in the test case, to be uploaded to the Google Cloud Vision API to identify the picture, and return 3 pens The result of the tagging operation, the API will return the specified number of JSON data.

According to the test results of the seventh picture, the Google Cloud Vision API can correctly parse that this picture is a polar bear, and brings 98% compliance. In addition, the picture may also be a mammal or an animal, 96% respectively With 95% confidence Degree, these trust degrees can be used to distinguish the trust weight of the label source.

The eighth to tenth pictures show the analysis results of the pictures related to the Nantai earthquake. The left of the eighth to tenth pictures shows the links between the pictures related to the keywords of the Nantai earthquake and the news. The eighth to tenth pictures The right side shows the results after tagging.

Observing the results of the labeling of the eighth to tenth pictures, if the pictures are labeled with earthquake (Earthquake), rubble (Rubbles), disaster (Disaster) and other related tags, the data is likely to be the searcher. Required content; on the contrary, if the labeling result does not match the searched keyword content, the information may be unrelated items, so these articles that do not match the keyword content can be ignored, Figure 11 ~ Figure 13 Shows the tagging results of the pictures contained in the article. The tagging data of the pictures contained in these articles does not match the content of the searched keywords, so the pictures shown in Figures 11 to 13 will be Is discarded and will not appear in search results.

That is, the present invention can access tag information in its pictures by using Google Cloud Vision API based on artificial intelligence, and filter by keywords and tag information in pictures to achieve the purpose of accurate search.

The above mentioned are only some of the embodiments of the present invention, and are not intended to limit the present invention. Therefore, those who have been modified slightly according to the creative spirit and features of the present invention should also be included in the scope of this patent.

In summary, the embodiments of the present invention can indeed achieve the expected use effect, and the specific technical means disclosed by it have not only not been seen in similar products, nor have they been disclosed before application, and have fully complied with the patent law. Regulations and requirements, according to law If you apply for a patent, please ask for examination and grant the patent.

Claims (5)

A search filtering method that improves the matching of search results of text images includes the following steps: Step 1: Establish a semantic database; Step 2: Enter at least one keyword in a search engine; Step 3: Select a semantic database and All the synonyms words included in the keyword; Step 4: compare the articles in a source database with all the synonyms words included in the keyword, and analyze whether the keywords appear in the article All the synonyms with the same vocabulary, if not, discard the article, and if so, keep the article as the preliminary search result article; Step 5: Analyze whether the tags appear in the tags of the pictures carried in the preliminary search result article All the synonyms words included in the keywords, if not, discard the preliminary search result article, if so, keep the preliminary search result article to become the final search result article, and put the title of the final search result article in The display device displays; wherein, the step of creating a semantic database includes the steps of importing a training library, and parsing the images in the library The steps of possible tag content, the step of collecting the possible tag content and storing it in the semantic database; in the step of parsing the tag content that may appear in each picture in the gallery, the gallery is first calculated The multiple images of generate a total of k labels with a suitability of S _m , and m = 1, ..., k, and then the level of S _j determines whether to retain the label (S _j represents the suitability of figure j); suppose each image Obtain q label contents and their corresponding scores (q appears in the range of j). This corresponding score is represented by Y _ij , where i is the picture number (i = 1, ..., n and j = 0, ..., q), when When j is 0, the image does not produce any tags. The total score of all k tags is added during the calculation. If there are many occurrences and the quality score is high, the calculated S _m will be relatively high, so the tag represents This is reserved for the semantic database, and its calculation formula is as follows:

As mentioned in the first paragraph of the patent application scope, the search filtering method for improving the search results of the text matches the optimization, wherein the step of creating a semantic database includes the steps of inputting Chinese keywords and a translation application The step of translating the input Chinese keywords into English and storing in the semantic database. As described in item 1 or 2 of the patent application scope, the search filter method for improving the search results of the text image matches the optimization, wherein the step of creating a semantic database also includes a step of expanding the semantic database, including through the existing semantic The step of the library searching for synonyms is to add the synonyms to the semantic database for storage. As mentioned in item 3 of the patent application scope, the search filter method for improving the search results of the text image is optimized, wherein whether the keyword is included in the tag of the picture carried in the article of the preliminary search result is analyzed In the process, the tag information in the picture is accessed by linking to the Google Cloud Vision API. The search filtering method for improving the search results of the text image according to item 1 or 2 of the patent application scope is optimized, wherein whether the key is included in the label of the picture carried in the article of the preliminary search result is analyzed In the process of wording, the tag information in the picture is accessed by linking to the Google Cloud Vision API.