Measuring Tag Confidence

2008. 3 ~ current

 

The vast amount of online available images requires a certain amount of descriptive tags to images, because generic online image sharing services are using tags to deal with searching, managing and organizing images. To do that, users have possibility to register appropriate tags to images when uploading these images. Image searching engines can then retrieve images by relying on tag-based search techniques. However, those tags annotated by users are not always representing the actual image. They are likely to be highly subjective. Sometimes, some tags are not related to the visual perception of the image. To overcome the limitation, this research proposes a method to measure tag confidence for images in database so that one can differentiate confidence tags from noisy tags.

 

Proposed Method

 

Overall Procedure

We define tag confidence as how much the tag is related to associate the image. Tag confidence value is defined varying from 0 to 1. If the tag is not related to the image, tag confidence value is close to 0. Likewise, as tag confidence value is close to 1, the tag is highly related to the image.

Fig. 1 illustrates the overall procedure of the proposed tag confidence measurement. The tag confidence is calculated based on the visual information of image. Let us assume the image database in Fig. 1 consist of images from UGC and associated tags.  We suppose to measure tag confidence for all tags with associated images in the database. To do that, a target image is selected from the image database and the tag confidence for the target image is measured. All images in the database are selected as target images so that confidence values of all tags in the database are measured.

Fig. 1. Overall procedure of proposed tag confidence measurement

Semantic Vector Generation

A semantic vector aggregates the confidence values that are the result of applying a series of independent classifiers to an image. Each confidence value represents the degree of certainty in which a particular concept has been detected. Each element of a semantic vector, i.e. the confidence value for a particular concept, is generated from a collection of training images that all contain the same concept.

Likelihood Image Selection

Likelihood images in the database, which are visually similar to the target image, are obtained with the semantic vectors mentioned above. The distance between two vectors of the target image and an image in the database is calculated with L2 distance. Since each element of semantic vector represents the probability of the semantic concepts, the maximum distance is 1. The selected likelihood images are I1, , Itho which are represented in order of similarity distance. Image in the database having small similarity distance means it is close to the target image.

Tag Confidence Measurement

Each tag confidence in the target image is supposed to be measured. Let nth tag of target image be Ttn. It is calculated by analyzing the word similarity of tags in likelihood images. We adopt WordNet to calculate the similarity among tags. Fig. 2 shows the schematic structure of the tag confidence measurement.

Fig. 2. Proposed tag confidence measure

Sihyoung Lee

ijiat@kaist.ac.kr

  Ph.D Candidate in Dept. of Information and Communications Engineering in KAIST

  Senior Researcher in IVY Lab in KAIST

Experience of Research projects

    Home photo categorization

    Multimedia application format for e-learning

    Automatic image annotation

 

Research Interest

    Image Tag Recommendation, Social Media Processing

 Hyun Seok Min

hsmin@kaist.ac.kr

 

  Ph.D Candidate in Dept. of Information and Communications Engineering in KAIST 

  Senior Researcher in IVY Lab in KAIST

  

Experience of Research projects

    Use of Semantic Features for Filtering of Malicious Content in an IPTV Environment

    Enhancement of Face Detection Using Spatial Context Information

    Semantic Annotation of Personal Video Content Using an Image Folksonomy

 

Research Interest

    Video annotation, Video copy detection, Face detection, Video event detection

 
̽, , ̿, , UCC ± ŷڼ ѿ, ѱȣóýȸ ϰмȸ 91ȣ, pp.68~71, 泲б, June 28, 2008 (selected as An Excellent Paper Award)
Sihyoung Lee, Hyun-Seok Min, Young Bok Lee, Yong Man Ro, Measurement of Tag Confidence in User Generated Contents Retrieval, SPIE 2009, Jan.18~22, 2009, San Jose, USA (accepted)