Data processing: database and file management or data structures – Database design – Data structure types
Reexamination Certificate
2000-06-29
2004-03-09
Metjahic, Safet (Department: 2171)
Data processing: database and file management or data structures
Database design
Data structure types
C382S305000
Reexamination Certificate
active
06704725
ABSTRACT:
BACKGROUND OF THE INVENTION
1. Field of the Invention
The present invention relates to multimedia data, and more particularly, to a method of searching multimedia data using normalizing information of feature elements included in a specified image.
2. Background of the Related Art
Recently, the digital image signal processing technology has been developing rapidly and has been applied in many fields. Some of these fields includes a search system for automatically editing only a face of a specific character in a moving picture file of a movie or drama; a security system for permitting access only to those registered in the system; and a search system for searching a particular data from an image or video detected by a detecting system. In any application field, the performance of a system essentially depends on the accuracy and speed of detecting or searching a desired object. Accordingly, various image searching methods have been proposed in the related art.
Generally, an image search system searches for similarity of images using features included in images such as a color histogram or color of partial regions. Thereafter, the system provides a user with the search result, i.e. similar images. If the user is satisfied with the search results, the search operation terminates. However, it may be difficult to find a desired image by only one search. Therefore, the user may search for the desired image using weights of features predetermined in the search system for sorting each image.
There are two typical methods for setting the weights of features. In the first method, a user sets and updates the weights of features using a user interface when the user searches the database. In the second method, a search system automatically sets and updates the weights of features using Relevance Feedback from users.
In the first method, a consistent standard is difficult to defined when general users set the weights of features. Also, an error may occur when learning the weights of the features because respective features may have different similarity distributions. For example, suppose a feature A in an image is set to have more weight than a feature B by a ratio of 6:4 when a user is searching for a reference image and that the average of the similarity distribution of feature A is 50 while the average of the similarity distribution of feature B is 70.
If image
1
and image
2
are in the database, and if feature A ranks image
1
higher than image
2
while feature B ranks image
2
higher than image
1
, image
1
should be ranked higher overall since feature A is more important. However, since the average similarity of feature A is lower than that of feature B, image
2
may be ranked higher than image
1
when judging only by the similarity, irrespective of the weights. Namely, when a search system determines the similarity of feature A and feature B in image
1
as 60 and 70 respectively, and the similarity of feature A and feature B in image
2
as 50 and 90 respectively, the search system would determine image
2
as having a higher rank than image
1
, even if weights are considered.
Therefore, the weights of features may not be reflected and an error may occur, even if users determines and corrects the weights of the features included in images as because similarity distributions of each feature are different.
In the second method in which a search system automatically sets and updates the weights of features using Relevance Feedback from users, an error may occur when setting and updating the weights of each feature since similarity (or difference) distribution of each feature in images differ. Namely, after a user searches for an image, the user gives positive relevance back to the search system for similar images to a reference image, and gives negative relevance for different images to the reference image. Thus, the search system automatically sets and updates the weights of features included in each image according to the relevance feedback by the user.
However, each feature in an image is generally evaluated in a different manner when determining a similarity of an image. For example, suppose that a search system searches for ten similar images and sorts the images by evaluating the similarity based on each feature. At this time, an image may be ranked fifth if similarity of feature A is 80 and the same image may also be ranked fifth if similarity of feature B is only 60. This is due to the different distribution of each feature in an image. Accordingly, an error may occur when the weights of features in the images are set and updated using similarity (or difference) of features having different distribution.
For example, suppose that an average of the similarity distribution for feature A is 50 and the average of the similarity distribution for feature B is 70, when a search system searches for a reference image using feature A and feature B. If feature A and feature B had the same distribution, the same weights would be assigned to feature A and B when the system ranks an image ranked the same by features A and B higher. However, if the system sets the weights of features A and B using an image ranked the same by feature A and B, an error of setting the weight of feature B greater than feature A would occur because the similarity for feature B is relatively greater than the similarity of feature A due to the different similarity distribution.
On the other hand, when a user gives feedback on one image as irrelevant, feature A with respectively low similarity will be assigned more weight, even if the irrelevant image was ranked the same by both features A and B. Therefore, the method for setting weights in the related art by user feedback may have problems in learning wrong weights since each feature included in images have different similarity (or difference) distribution.
SUMMARY OF THE INVENTION
Accordingly, an object of the present invention is to solve at least the problems and disadvantages of the related art.
An object of the present invention is to provide a method of searching multimedia data using normalizing information of features.
Another object of the present invention is to provide a method of normalizing information used in learning weights of features for searching an image.
A further object of the present invention is to provide a data structure used in learning the weights of features for searching an image.
Additional advantages, objects, and features of the invention will be set forth in part in the description which follows and in part will become apparent to those having ordinary skill in the art upon examination of the following or may be learned from practice of the invention. The objects and advantages of the invention may be realized and attained ash particularly pointed out in the appended claims.
To achieve the objects and in accordance with the purposes of the invention, as embodied and broadly described herein, a method of searching multimedia data comprises normalizing information of each feature in at least one multimedia data in a search system; normalizing distribution of similarities or differences of each feature corresponding to the multimedia data using the normalizing information in the search system; and updating weights of each feature using the normalized similarity or difference.
In another embodiment of the present invention, a method of searching multimedia data comprises obtaining a Probability Density Function (PDF) corresponding to similarity or difference of features in a prepared multimedia data set in a multimedia data search system; obtaining Cumulative Distribution Function (CDF) corresponding to each feature using said PDF in said multimedia data search system; obtaining a first slope and a first mean to approximate said CDF in the multimedia data search system; setting a second slope and a second mean of normalizing object respectively in the multimedia data search system; obtaining similarity or difference corresponding to each feature of target multimedia data and reference multimedia data according to the referenc
Filipczyk Marcin
Fleshner & Kim LLP
Metjahic Safet
LandOfFree
Method of searching multimedia data does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Method of searching multimedia data, we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Method of searching multimedia data will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-3243380