TW200919233A - Content-based image retrieval method by integrating navigation patterns mining and relevance feedback - Google Patents

Abstract

A content-based image retrieval method by integrating navigation patterns mining and relevance feedback is retrieved. This method comprises an offline stage and an online stage. The offline stage includes a data transformation step, a navigation pattern mining step and a pattern indexing step. The online stage includes an initial query processing phase and a Q3 search phase. The initial query processing phase includes a feature extraction step, an initial feedback step and a user interaction step. The Q3 search phase includes a query-point generation step, a query expansion step, a query re-weighting step and a result filtering step.

Description

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method for retrieving an image content, and more particularly to an image content retrieval method for integrating a survey style and relevance feedback. [Prior Art] With the advancement of information technology in recent years, coupled with storage devices, network transmission, and data compression technology, the use of multimedia data has rapidly expanded, such as photos, video movies, and music. In order to deal with the ever-increasing variety of multimedia materials and their diversity, the search technology of multimedia content has always been a very important topic, and image search or query technology is a very important part. Image search or query technology is used to provide images related to the query image entered by the user. The technique of "Smantic Image Retrieval" is based on the knowledge of the file name, category and annotation. But when it comes to developing such a text-based search method, it takes a lot of manpower to give the appropriate semantics to the picture, so that the concept of the picture can be similar to the subjective consciousness of human beings. However, such prior art is quite labor intensive and requires a large amount of cost. Another conventional technique uses an automated R&D identification system to semantically mark the content of the picture. However, due to the considerable discrepancies between the computer system and human subjectivity, the efficiency of using such an automated identification system is rather low. . 5 200919233 In order to make image retrieval easier and more accurate, the techniques of Content-Based Image Retrieval (CBIR) and Relevance Feedback (RF) have gradually become the focus of recent research and development. However, the conventional CBIR and related feedback techniques have the following disadvantages: 1. When the user queries a certain picture, the user needs to go through the process of long browsing picture results and countless feedback interactions to reach the user. The society dedicated to the mind, FB t +

Tm jaundice, and the problem of Redundant Browsing. 2. The concept that the user is looking for may be concentrated in different areas of the feature space due to different feature values, and cannot be retrieved. However, these pictures are in line with the concept of human beings, which is to find the pictures that users want, thus creating the problem of visual diversity. For example: Suppose the concept of the query is flower 'but is divided into white flowers and safflower. At this time, if the low-level image feature value is used, the space between the white flower and the red flower sign vector is different in the effect. The area cannot be retrieved, but in the human concept, both groups belong to the category of flowers' and should be taken into account. 3. The same-picture may result in different results for different users. The Local Optimal problem of Convergence is generated. SUMMARY OF THE INVENTION Therefore, it is highly desirable to develop an image content retrieval method that integrates browsing style exploration and 200919233 relevance feedback to overcome the shortcomings of the prior art and thereby increase the rate of retrieval. One aspect of the present invention provides an image content retrieval method that integrates browsing style exploration and related feedback, thereby reducing the number of times a user queries a view' and quickly returns a result related to a picture that the user desires to query. Another aspect of the present invention is to provide an image content retrieval method that integrates browsing style exploration and relevance feedback to improve the quasi-break rate of each feedback, and to reduce the influence of visual differences and search convergence. An embodiment of the present invention provides an image content retrieval method that integrates browsing style exploration and relevance feedback. In this method, a plurality of query records (Log Data) corresponding to a plurality of historical query (Query) pictures and historical query pictures are first obtained, wherein the query records include at least one feedback, and each feedback has a plurality of historical related pictures. Next, the offline processing phase and the online processing phase are performed. In the offline processing stage, the data conversion step is first performed to calculate the center point of the feature vector of each of the historically related pictures of each feedback, and a plurality of historical query center points corresponding to each feedback of the historical query pictures are obtained ( Query Point), and using a grouping algorithm to divide the historical query pictures into a plurality of historical query picture clusters with a preset number, and divide the historical query center points of each feedback into a plurality of queries having the preset number Central point cluster. Then, the browsing style exploration step is performed to find the frequency of the historical query image cluster and the query center point cluster of each feedback by using the continuous correlation algorithm and according to the foregoing query record (Frequent Patterns). And get a plurality of Liu styles. Next, the online processing stage is performed. In the online processing stage, the user first enters a query picture. Then, a feature extraction step is performed to capture the feature values of the query picture. Then, an initial feedback step is performed to perform (10) the sum of all the pictures in the library for the query picture & eigenvalues, and obtain the closest approximation to the picture from the historical query picture. Next, a user interaction step is performed for the use of the present + secret and 0 related pictures, "the first: the first feedback picture selects a plurality of first pictures as the picture of the plurality of first feedback pictures except the first related picture Irrelevant picture. Then, in the first frame, the u point generating step is performed to calculate the first 杳-two of the first related picture: the algorithm and according to the history query step, the money (four) is calculated in the distance-inquiry , 'and the style of the tide' to obtain a plurality of second feedback pictures that are most similar to this point. In the index line processing stage, the more styles of the number of views are combined into a plurality of pre-numbers starting from the query picture. In another embodiment, in the offline processing stage, the step of deleting the redundant style is further performed, and the historical query center is deleted from the original query picture and the browsing style (or the structure of the tree). The number of points in the query center is the same, but the number of points in the middle is larger. Query: Online processing stage I is more calculated - check the number of the first picture: the accuracy is the number of the first related picture I. Judging whether the accuracy of the query is greater than 200919233 or equal to a quasi-definite φ + deterministic threshold, and when the yoghurt is YES, the first pass - the judgment result. When judging the knot 1 expanding the query step, the second relevant picture is calculated The statistical distance algorithm is used to divide the similarity of the first related picture query picture, and the first number of first seed pictures are obtained when the broken result is no; when the expansion is completed, the first calculation is performed. Related pictures f D steps are based on the statistical distance algorithm to query the similarity of the picture, σ on the first - irrelevant pictures and historical search

A number of first seed maps 4 are similar to the first replies of the first related picture, in the browsing tree structure of the seed picture of the tree, and corresponding to the first in the structure of the tree

The plural number of Nodes) - find the relevant leaf nodes (Leaf ^ related query center, algorithm, to calculate the first point. Then, by statistical distance - query the similarity of the central stagnation point, and get the most approximate —·'/, the first relevant query center returns the query center point. Then, the first pair of the number of the first to the first-to-feedback query in the middle of the sentence is selected. After that, the statistical distance algorithm is used to calculate the picture. However, the younger one queries the similarity between the center point and the corresponding picture of the flute, and obtains the closest approximation, and the younger ~~ takes the place--the query center point two feedback The first result is a non-correlation graph. In still another embodiment, in the online processing stage, the step of filtering is removed to delete one of the first seed pictures corresponding to the most recent one of the first pictures. In another embodiment, in the online processing stage, the feature weight value adjustment step is to utilize the attribute weight value of the first related picture. In still another embodiment, in the online processing stage, f enters > 'Use 9 200919233 to return to the second phase of the picture to check the center point of the center point. The algorithm is closest to the first number of times, the result; can be compared with the search to receive 4 plural pictures, the second check 2 query statistics to get the inquiry The related difference is browsed for the user to select a closed picture from the second feedback picture, wherein the second feedback picture is divided into a plurality of second irrelevant pictures. Then, the generating step is performed to calculate the The second related picture then 'performs a second extended query step to use and according to the historical query picture and the plurality of viewing styles, and the second plurality of feedback pictures of the central point of the query. Applying the embodiment of the present invention can reduce The user checks and quickly returns the accuracy of each feedback with the user's desired picture, and can reduce the influence of the problem such as visual convergence. [Embodiment] The present invention mainly uses a mobile query center at the same time. Query-Point-Movement (QPM), Query-Reweighting (QR), EXTery-Expan Si〇n ° QEX)' is called the Q3 model to improve the accuracy of each feedback and reduce the impact of visual differences and search convergence. For the visual difference between images and the convergence of search results, users may There are different browsing paths, which result in different image search results. Therefore, the present invention incorporates the concept of extended query and associated feedback technology to cover these different related picture results to reduce the impact of these problems. The problem of redundant viewing, the invention is to obtain relevant rules by exploring the style of the user to zoom in on the picture, so as to reduce the number of times the user browses the feedback. 200919233 Please refer to FIG. 1 , which is a block diagram showing an image content retrieval method for integrated browsing style exploration and relevance feedback according to an embodiment of the present invention. This embodiment can be divided into an offline processing stage 1 and an online processing stage 200. In this embodiment, first, the offline processing stage ι is performed by mining the user's quaint (four) style to establish a rule of the exploration style to provide the user with the best picture search. The stage 1 is processed offline and these rule exploration styles (browsing styles) are stored in the rule database. Then,

U can perform an online processing stage 200 to retrieve the content of the 5th picture 50 entered by the user (not shown). In the online processing stage 200, first, an initial query processing step 202 is performed, by comparing the feature values of all the pictures in the query picture 50 and the picture database 1 to provide related pictures (not shown) to the user. . Then, step 290 is performed to determine whether the user is satisfied with the content of the related picture. When the result of step 290 is "YES", the online processing stage 200 is stopped. When the result of the step 290 is "NO", the Q3 search is performed to obtain a feedback picture to the user according to the rule exploration pattern in the rule database 3 and in conjunction with the algorithm of the Q3 model. Next, a step 280 of user feedback is performed to allow the user to select a plurality of related pictures 60 from the feedback picture. Then, step 29 is performed to determine whether the user is satisfied with the content of the related picture 60. When the result of step 29〇 is 疋", the online processing stage 200 is stopped. When the result of step 290 is "NO", the Q3 search step 2〇4 of the next feedback is performed. At the same time, each time the Q3 search step 204 is performed, the user's information (inquiry record, query picture and related pictures) is stored in the query record database 2 and the picture 200919233 piece database u). By recording the user's (4) behavior record for a long time, the offline processing stage 100 can create a new and useful user profile: . Knowledge base to speed up the search accuracy of the picture. The offline processing stage 1 and the online stage 200 of the present embodiment will be separately described below. % Miscellaneous processing step 埒inn Offline processing stage 100 is to transfer the data to the user's query record, and then perform the exploration operation (4) to dig out the hidden lacquer! Users get fast and accurate results. The input of the offline processing stage 100 is the historical query picture of the user (the picture database is called the query record corresponding to the historical query picture (query record database) and the output of the offline processing stage i 00 is the searched rule (browsing style or The Liu Shushu structure) is stored in the rule database 3G. The offline processing stage 100 includes at least a data conversion step 110, a review style exploration step 120, and a style indexing step 130. (j Refer to FIG. 1 and FIG. 2' Figure 2 is a schematic diagram showing the data conversion step 11 of the embodiment of the present invention, wherein the arrow dotted line is used to display the lake view path of the user in the query record. First, a plurality of historical query pictures are obtained. The plurality of query records corresponding to the historical query pictures, wherein the query records include at least one return to the library, and each of the feedbacks has eight historical related pictures. Then, the data conversion step 110 is performed to calculate each Obtaining the complex number of each feedback corresponding to the historical query picture by the center of the feature vector of the history-related picture of one feedback Historical inquiry center point of the central point of the feature vector computing picture of (ie Information Center 12 200 919 233 points) method can also, for example. Space vector equation (Space-Vector

Formula), the eigenvalue weighted Euclidean Distance or the general Euclidean distance (Generalized)

Euclidean Distance) and so on. Then, a grouping algorithm is used to divide the historical query picture into a plurality of historical query picture clusters Qi to Qm having a preset number m; and the historical query center points of each feedback (0th feedback to the nth feedback) are divided into A plurality of query center point clusters Cq! to Cnm having the same preset number of melons, and sequentially encode the historical query picture cluster Qi to the query center point cluster C〇1 to Cnm with each feedback, and make the parent a history The query picture clusters Qi to Qm have a unique identification code with each of the query center point clusters Co! to Cnm. Referring to FIG. 1 and FIG. 3, FIG. 3 is a schematic diagram showing a browsing pattern searching step 12 of an embodiment of the present invention, wherein for convenience of explanation, the history query picture cluster of FIG. 3 and The default number of query center point clusters is 5. After the data conversion step 11〇, the browsing style exploration step 120 is continued to use the continuous association algorithm (for example: Apri...寅 algorithm, etc.) and find the historical query picture clusters Q1 to... according to the user's query record. Each time the feedback center point group, 15 frequent item sets, and a plurality of Pan J styles are obtained (the view of these (four) styles is clustered by historical query pictures to... and one" heart point cluster CG1 to cis identification The code is to be represented. The group H style exploration step 12G task must be tested in each feedback continuous relationship. Basically, the work of the Liu Ming module is established - i browse the transaction data table and generate the browsing transaction style two 13 200919233 steps As shown in Table 1, a (Transaction). A winning cycle can be regarded as a transaction query picture) with multiple feedbacks on the starting query point project (historical frequent items. As shown in Figure 3; The starting query item of the (historical query center point) is shown by the dotted line of the Q head. It is assumed that the record 〇〇1 belongs to C〇3, and the first query of the query center point item of the 0th feedback is the second feedback. Inquiries - 6 tenth center point project c12..., point item C This trader 1 changed the transaction to {qi C〇3 C 12, so the step of querying the record 001 data transfer table will use all ^/. /", 匸^卜Create a browse transaction information table, to effectively check the record of the transaction to establish a browse and exchange as shown in Table 1, to investigate the exploration of the cut-off style. /〇1 is the same as the start of the query record 002; the query record. 2: Check the same but the destination group is the same. The initial search group of apricot, 〇〇4 is not the starting date. 〇〇3 is the same as the query record 004 Ο ” ” group but the destination group is different; in addition, the query 〇 5 may be another important and different path. Therefore, from the above observation, the present embodiment It is true that we can fully consider the information of this hidden connotation., ' __录 - 1 - >, frequent items 001 Ql, C〇3, C] 2, C^l'. 32, C42 002 -~~ ~~- Ql,C〇2,C",C23,C32,C42 003 Q 2,C (Η,C 】2,C 2 1,C 3 2,C 4 1 14 200919233 〇〇4 Q2,C〇3 ,Ci2,C2 1,C31,c42 〇〇5 · ---- _ Q3,C〇 1,C13,C22,C32,C43 Table 1 View Transaction Data Sheet

After completing the transaction data sheet, Cheng Pan proceeds to the step of generating a browse transaction style. This step is mainly about 'focusing on finding out the relevant feedback relationships between different users. This step can use, for example, the Apriori algorithm to find frequent itemsets that are greater than the minimum support, which can be considered as Sequential Association Rules, since relatively ambiguous codes are continually attached at the time of data conversion. Sexual traits such as: {c"->c32_>C42} are extended by frequent consecutive itemsets. After the story style exploration step 1 120, a number of Liu styles are obtained, and these browsing styles are stored in the rule database. A tidal style can be thought of as a path, and these paths can be stored using the tree's structure to reduce storage space and search time. Referring to Figures 1 and 4, FIG. 4 is a schematic diagram showing a sample indexing step 130 for explaining an embodiment of the present invention. After performing the browsing style exploration step 12, the style indexing step 1 3 0 can be performed to combine all the browsing styles into a plurality of browsing tree structures starting from the historical query picture, wherein the number of browsing tree structures and the number of query pictures the same. For example, the history query picture is divided into k groups in the offline processing stage 100. Therefore, in the picture search process, the image search can be performed using the established k tree structure for all possible paths. That is to say, when 15 200919233 starts to check the 3rd picture (historical query picture) and breaks 77% of the k divided groups, the center point of each group starting from the historical query picture will be treated as a seed. Then build a browsing tree. As shown in Figure 4, as shown in Figure 4, a single J-tree tree package has multiple paths, and this ambiguity #门仏格 begins at the same Query Root, each Node (N〇de) 々J represents a feedback in a browse path. Ο

Since the nodes of the browsing tree structure are quite complicated and complicated, the step of deleting the redundant style can be performed to delete the starting query picture in the browsing tree structure and the same as the terminating historical query center point, but the number of historical query center points in the middle is the same. More browsing styles (paths), thus increasing the speed of the search. The definition of redundant styles is explained below. Suppose there is two browsing styles in a browse style collection:

Fitemsetl = {Cij,...............,Cmn} and

Fitemset2={CPg,...............,Cxy} Right Cij=Cpg’ Cmn=Cxy and I Fitemsetl| $ j Fitemset2|, then Fitemsetl is a redundant style. For example: Pan I view style 1: 1=>6 = >12 = >17 = >25 and browse style 2 '· 1==>8=>25 because the initial query image and last feedback The query center point is the same, so the browse style 1 is defined as a redundant style, and the Pan style 1 is deleted from the browse style set (browsing tree structure). After the step of removing the redundant pattern, the subsequent search can be simplified, so that it is not necessary to search the entire database, and the calculation cost can be reduced. Online Processing Segment 200 The online processing phase 200 includes at least an initial query processing step 2〇2 200919233 and a Q3 search step 204. After the user enters the query picture 50, the initial query processing step 202 is first performed. In the initial query processing step 202, the feature extraction step 210 is performed to capture the feature value of the query image 50, which may be a low-order feature value such as color, texture, shape, and the like of the image. Then, an initial feedback step 22 is performed to compare the feature values of the query picture 5〇 with the feature values of all the pictures in the picture database 10, and obtain the most approximate number of the search pictures 50 from the database picture. A feedback image for one feedback. In this embodiment, the number of feedback pictures for each feedback may be a fixed number, for example: 1〇·, except for the feedback picture of the initial feedback is the database picture, and the other feedback pictures are the month " Query the historical query image of the image. Then, the user interaction step 2 fly η 勒 7骒23〇, for the user to select the lane from the feedback image of the first feedback, ^ 卬筏 several first feedback related pictures , the basin φ back to the museum for the first feedback, the picture with the picture other than the first feedback in the picture is called the irrelevant picture of the first feedback. Ο Then proceed to step 29 to determine if the user is satisfied with the content of the relevant picture. The volume is also im, line: slice - if the result of step 290 is "YES", the processing stage 200 is stopped. The result of step 29 is the Q search step 204. j π % 订 发明 参见 参见 参见 参见 实施 实施 实施 实施 实施 实施 实施 实施 实施 实施 实施 实施 实施 实施 实施 实施 实施 实施 实施 实施 实施 实施 实施 实施 实施 实施 实施 实施 实施骒 In the Q3 search step 204, the query center point generation step 240 is first performed to calculate the correlation of the first return to the library and the calculation accuracy of the query is calculated. ^ 17 200919233 The number of related pictures is the number of related pictures. Divide by the number of images returned. Next, an augmented query step 250 is performed. As shown in block 254 of Figure 5, when the "5» query accuracy rate of Putian is greater than or equal to the accuracy threshold, the statistical distance algorithm is used to calculate the relevant pictures of the first feedback separately (2, 5, 6 pictures) similarity with the historical query pictures, and obtain a plurality of seed pictures which are respectively closest to the related pictures of the first feedback. As shown in block 252 of FIG. 5, when the query accuracy is less than the accuracy threshold, the statistical feedback algorithm is used to calculate the feedback image of the first feedback (related pictures (2, 3 pictures) And the irrelevant pictures (4th, 6th, and 7th pictures)) are similar to the historical query pictures, and obtain a plurality of seed pictures respectively corresponding to the related pictures of the first feedbacks. In other words, when the query accuracy is greater than or equal to the accuracy threshold, the embodiment only needs to consider the relevant picture selected by the user before; when the query accuracy is less than the accuracy threshold, the present embodiment will be the previous time. The feedback images provided are all taken into account to improve the accuracy of the next feedback. In this embodiment, the number of seed pictures for each feedback may be a fixed number, for example, 10 sheets, that is, the seed picture is a historical query picture of the top 1 most similar picture. As shown in block 252 of FIG. 5, when the query accuracy is less than the accuracy threshold, a result filtering step 270 may be performed to delete one of the seed pictures corresponding to the most irrelevant picture, such as the fifth. Figure 7 The seventh picture in block 252 to increase the search rate. Then, a plurality of phases of the relevant leaf nodes are searched for in the structure of the browsing tree corresponding to the seed picture in the browsing tree structure shown in Fig. 4 18 200919233

Close the query center point. Then, using the aforementioned statistical distance algorithm, the similarity between the query center point qpnew of the related picture of the _ under-return and the related query center point is calculated, and a plurality of feedback query center points closest to the query center point qpuu are obtained. Then, a plurality of corresponding pictures corresponding to the feedback query center points are selected from the historical related pictures. Then, using the statistical distance algorithm described above, the similarity between the query center point qPnew of the first feedback related picture and the corresponding pictures is calculated, and the second feedback closest to the first-query center point is obtained. Give feedback to the film such as: 丨〇 )). The statistical distance wide algorithm used in this embodiment may be, for example, a 35-degree formula of the eigenvalue weight or a Euclidean distance used. One " Next, the user feedback step 28〇 is performed to allow the user to retrieve the right dry picture 60 from the feedback picture of the first feedback. After the absence, step 290 is performed to determine whether the use of a is not discrete ..., and whether the user is satisfied with the inner picture of the related picture. When step 29 〇 . _ L L 疋 疋 ,, the line j· processing stage 200 is stopped. When step 29 拎 订 一 一 一 ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ In the next method of @_ 贞, the method used in the right embodiment can be used to query the information clip of the Euclidean distance calculus 60 of the special tongue-weight 260 ' The feature weight value is determined using the associated edge of the graph. The characteristics to increase or decrease the color, according to the related picture 60 ~ π 巴 知, texture or shape to increase the query accuracy. The weight value of the special feature, the definition of the weight value of the hand feature, such as the melon β and the different characteristics of the η {F丨···*.Fn} and the corresponding 19 200919233 standard deviation set (σ!, σ2,... , ση}. And assume that each feature F is expressed in m dimension vectors by k_p point X, x=={fl, f2, ... fm}, and there are k pictures with the query center ', ( qP) performs similarity comparison. Therefore, the weight value definition of the i-th feature Fi

(3)

Where σ, and Fi ={fl5f2,

The present embodiment will be described below based on Fig. 5. Assuming that k pictures are returned to the user, the user will decide which pictures are related and unrelated pictures. Then, according to the information, the present embodiment generates a new query center point qPNEW, and calculates the query accuracy of the feedback. . If the query accuracy is lower than the threshold, as shown in block 252, then this embodiment will collect the seeds (1st, 2nd, and 7th pictures) corresponding to the related pictures and correspond to irrelevant pictures. Seeds (4th, 6th, and 7th pictures), and find the seed that corresponds to the irrelevant picture (the 7th picture) and delete it from the seed set to get the seed to be searched (1st, 2nd, 2nd) 4, 6 pictures) ° pick up to the corresponding browsing tree of various sub-categories, find the relevant query center points of the relevant leaf nodes from the path of the browsing tree, and get the first s query center points that are the most similar to qpNEW . Next, the most similar pictures of the first k pictures are taken out from the pictures corresponding to the s query center points, and are transmitted back to the user. When the query accuracy is greater than or equal to the threshold value, as shown in block 254, it is only 20 less than 200919233. The part that processes the irrelevant picture is only seeded by the related picture information (the 2nd, 5th, and 6th pictures), and These seeds are searched for the corresponding k pictures in the most relevant s query center points and returned to the user. It can be seen from the above embodiments that the present invention can achieve high accuracy, and at the same time consider the problems of visual diversity, search convergence, and redundant browsing; it can be avoided by exploring the rules of the user's Liu Ying picture style to avoid The length of the lengthy feedback, thus improving efficiency; can overcome the problem of visual eigenvalue difference.

Although the present invention has been disclosed in the above preferred embodiments, it is not intended to limit the scope of the invention, and the skilled person will be able to make various changes and modifications without departing from the spirit and scope of the present invention. The scope of protection of the invention is subject to the definition of the scope of the patent application.

21 200919233 [Simplified description of the drawings] " For a more complete understanding of the present invention and its advantages, please refer to the above description, and with the following drawings, wherein: Figure 1 shows an integrated browsing style according to an embodiment of the present invention. A block diagram of the image content retrieval method of exploration and relevance feedback. Figure 2 is a schematic diagram showing the data conversion steps for explaining an embodiment of the present invention. ^ Figure 3 is a schematic diagram showing the browsing pattern exploration steps for explaining an embodiment of the present invention. Figure 4 is a diagram showing the steps of the pattern indexing for explaining an embodiment of the present invention. Figure 5 is a schematic diagram showing the Q3 search step for explaining an embodiment of the present invention. [Main component symbol description] 10 Picture database I' 20 Query record database 30 Rule database 50 Query picture 60 Related pictures 100 Offline processing stage ' 110 Data conversion 120 Browse style exploration 130 Style index 22 200919233 201 Online processing stage 202 Initial Query Processing 204 Q3 Search Step 210 Feature Capture 220 Initial Feedback 230 User Interaction 240 Query Center Point Generation 250 Extended Query 252, 254 Block 260 Query Feature Weight Values 270 Results Filter 280 User Feedback 290 User Satisfied qp ' qpNEw query center

Claims (1)

200919233 X. Patent application scope 1. An image content retrieval method for integrated browsing style exploration and relevance feedback includes at least: obtaining a plurality of historical query (Query) images and a plurality of query records corresponding to the historical query images (Log Data), wherein the query records include at least one feedback, and each of the at least one feedback has a plurality of historical related pictures; performing an offline processing stage, wherein the offline processing stage comprises at least: performing a data conversion step to calculate And each of the at least one feedback center points of the feature vectors of the historical related pictures, and obtaining a plurality of historical query center points (Query Points) corresponding to each of the at least one feedback of the historical query pictures, and using one The grouping algorithm divides the historical query pictures into a plurality of historical query picture clusters having a predetermined number, and divides the historical query center points of each of the at least one feedback into a plurality of queries having the preset number Center point clustering; and performing a browse style exploration step to use a continuous association algorithm and according to the query records, to find the historical query picture cluster and each of the at least one feedback of the Fresent Patterns of the query heart point clusters, and obtain a plurality of Browsing the style; and performing an online processing stage, wherein the online processing stage includes at least: 24 200919233 input by a user - performing a feature capture quotation value; querying a picture; step 'to capture the special picture of the query picture - An initial feedback step of comparing the feature values of the query picture with the feature values of all the pictures in a picture database, and obtaining a plurality of first feedback pictures that are most approximate to the query picture from the historical query pictures; Ο Performing a user interaction step for the user to select a plurality of first related pictures from the first feedback pictures, wherein the pictures of the first feedback pictures other than the first related pictures are plural a first unrelated picture; performing a first query center point generating step to calculate a first query center point of the first related pictures; and performing a first extended query step to use a statistical distance algorithm Obtaining a plurality of second feedback pictures that are closest to the first query center point according to the historical query pictures and the browsing styles. 2. The integrated browsing style exploration and the Μ _ _ ^. The feedback image content retrieval method 'where the offline processing stage includes at least: the process of deleting the redundant style to delete The query images at the beginning of the oscillating view style are the same as the history point of the terminated historical query, but the number of the historical query center points of the middle 3 is more. 25 200919233 The method of retrieving the integrated Liu Lian style exploration and the video content retrieval method described in the 1st article of the patent 11 coffers, wherein the offline processing stage is further related to: the first style indexing step to the browsing style The composition is composed of: a plurality of the predetermined number of plurals starting from the query picture (4) browsing tree knot C Ο Guan Cong Ni, as described in the patent scope of the application, the integrated browsing style exploration and feeding image content retrieval method, wherein The style indexing step is more advanced) includes: constructing a step of deleting the redundant style to delete the middle of the browsing tree. The 5th-day picture is the same as the end of the historical query center point, but the history has a large number of browsing points. In the video content retrieval method of the integrated Liuguan style exploration and joint township described in the third or fourth paragraph of the patent (2) patent, the online processing text at least includes: relative calculation - query accuracy The 1^ query is the number of the first closed pictures divided by the number of the first feedback pictures divided by the second; the value 2 is the accuracy of the query is greater than or equal to the accuracy threshold, and Get a judgment result. For example, the integrated browsing style exploration described in claim 5 of the patent scope and the image content retrieval method of the 2009200923 association feedback method, wherein when the determination result is YES, the first extended query step further comprises: at least the statistical distance The algorithm is configured to separately calculate the similarity between the first related pictures and the historical query pictures, and obtain a plurality of first seed pictures that are respectively closest to the first related pictures; corresponding to the browsing tree structure In the browsing tree structure of the first seed picture, 'find the plurality of first-related query center points of the relevant leaf nodes (FeastNodes); F' * ^ using the statistical distance algorithm to calculate the first query center point and the The first related query center points are similarly obtained, and the plurality of first feedback query center points that are closest to the first query center point are obtained; and the first feedback query center points are selected from the historical related pictures. a plurality of first corresponding pictures; and calculating the phase of the first query center point and the first corresponding pictures by using the statistical distance algorithm Of 'recently obtained similar to those of the first query the central point of the second feedback picture. 〇7_ </ RTI> <RTIgt; </ RTI> <RTIgt; </ RTI> <RTIgt; </ RTI> <RTIgt; </ RTI> <RTIgt; </ RTI> <RTIgt; The distance algorithm is used to calculate the similarity of the μ ^ called the wide ^ historical query picture, and obtain a plurality of first seed pictures which are respectively closest to the second brother-reward picture; in the browsing tree structure Corresponding to the first seed picture browsing tree 27 200919233 structure, finding a plurality of first related query center points of the relevant leaf nodes; using the statistical distance algorithm to calculate the first query center point and the first Correlating the similarity of the central point of the query, and obtaining a plurality of first feedback query center points that are closest to the first query center point; and selecting a plurality of the first related feedback query center points from the historical related pictures a first corresponding picture; and using the statistical distance algorithm to calculate a similarity between the first query center point and the first corresponding pictures, and obtain The first query approximate center point of the second feedback these pictures. 8 . The image content retrieval method of the integrated browsing style exploration and the related feedback according to claim 7 , wherein the online processing stage further comprises: performing a result filtering step to delete the first seed pictures. The middle corresponds to one of the first unrelated pictures. 9. The image content retrieval method of the integrated browsing style exploration and the joint student feedback as described in the patent application scope item, wherein the statistical distance algorithm is a statistical distance algorithm of the feature value weight. The patent content retrieval method described in item 1 (4) of the patent application (4) refers to the image content retrieval method of the joint feedback, wherein the statistical distance calculation is a Euclidean distance algorithm with a eigenvalue weight. 28 200919233 - u. The method for image content retrieval of the integrated browsing style exploration and related feedback according to claim 9 of the patent application scope, wherein the online processing stage further comprises: performing a query feature weight value adjustment step to The information of the first related pictures is used to determine the feature weight value. 12. The image content retrieval method of the integrated browsing style exploration and related feedback according to claim 1, wherein the online processing stage further comprises: performing a user feedback step for the user to use the a plurality of second related pictures are selected in the second feedback picture, wherein the pictures other than the second related pictures in the second feedback picture are a plurality of second unrelated pictures; and performing a second query center point generating step Calculating a second query center point of one of the second related pictures; and C) performing a second extended query step to use the statistical distance algorithm and according to the historical query pictures and the cut patterns, To obtain a plurality of third feedback pictures that are closest to the second query center point. 13. The image content retrieval method of the integrated browsing style exploration and relevance feedback described in claim 12, wherein the offline processing stage further comprises: performing a style indexing step to combine the browsing styles into From 29 200919233, the query pictures start with a plurality of browsing tree structures with the preset number. 14. The image content retrieval method of the integrated browsing style exploration and the associated feedback according to claim 13 of the patent application scope, wherein the pattern indexing step further comprises: performing the step of deleting the redundant pattern to delete the pans! The tree view in the structure is the same as the history point of the terminated history query, but the number of historical query center points in the middle is larger. 15. The image content retrieval method for integrated browsing style exploration and relevance feedback according to claim 13 or claim 14, wherein the online processing stage further comprises: • • 10 calculation accuracy, wherein The query accuracy is the number of the first related pictures divided by the number of the first feedback pictures; whether the rate of the 4-5 is the maximum &amp; or the &amp; an accuracy rate threshold, and obtains a critical result. 16. If the image content retrieval method of the integrated (4) style exploration and related feedback mentioned in Item 15 of the patent application is applied, if the judgment result is yes, the second extended query step further includes: The distance algorithm, the different method calculates the similarity of the similarity between the second related pictures and the historical query pictures to obtain a plurality of second seed pictures that respectively approximate the second related pictures; 200919233, in the browsing tree structure corresponding to the second seed picture in the browsing tree structure, finding a plurality of second correlation check 1 center points of the relevant leaf nodes; calculating the number by using the statistical distance algorithm In the second query, the similarities of the second related query center points are obtained, and the plurality of second feedback query center points that are closest to the second query center point are obtained; and the corresponding historical images are selected to correspond to the first The second query picture of the plurality of second corresponding pictures; and the statistical distance algorithm 'to calculate the second query center point and the second corresponding pictures Similarity, obtained recently disappeared consultation like the second center point of the third back some pictures. 17. If you apply for a patent scope! The image content retrieval method of the integrated browsing style exploration and the related feedback, wherein the second extended query step further comprises: using the statistical distance algorithm to calculate the respective a second feedback pattern ◎ a similarity between the slices and the historical query pictures, and obtaining a plurality of second seed pictures that are respectively closest to the second feedback pictures; and browsing to the second seed picture in the browsing tree structures In the tree structure, a plurality of second related query center points of the relevant leaf nodes are found; and the statistical distance algorithm is used to calculate the similarity between the second query center point and the second related query center points, thereby obtaining a plurality of second feedback query center points that are similar to the second query center point; 31 200919233 selecting a plurality of second corresponding pictures corresponding to the second feedback query center points from the historical related pictures, and Calculating a distance algorithm to calculate a similarity between the second query center point and the second corresponding pictures to obtain the most similar second check The central point of the third back some pictures. 18. The image content retrieval method of the integrated browsing style exploration and the related feedback according to claim 17, wherein the online processing stage P further comprises: performing a result filtering step to delete the second seeds. The picture corresponds to the most one of the second irrelevant pictures. 19. The image content retrieval method for integrated browsing style exploration and relevance feedback according to claim 12, wherein the statistical distance calculation is a statistical distance algorithm with a feature value weight. 2〇. As described in the scope of patent application, the integrated browsing style exploration ^ relevance (4) image content retrieval method, the statistical distance calculus, is a eigenvalue weight Euclidean distance algorithm. 21. The image content retrieval according to claim 19 and the related feedback further includes: the integrated browsing style exploration method described in the item, wherein the online processing stage performs a query feature weight value adjustment step to utilize the second Related 32 200919233 The picture of "Ifl to determine the feature weight value. The exploration and replacement steps are more in turn for the historical query image cluster and each 哕 &gt; lL mine six y inverse at least one feedback Upstreaming some query center point clusters for encoding, so that each of the historical sentence picture clusters has a unique identification with each of the query center point clusters. 23. As shown in the scope of claim 22 The image content retrieval method of the style exploration and the associated feedback 'the browsing style exploration step further obtains the browsing styles according to the query records and the identification of the query center point clusters. 33