Thesis of Lilei Zheng


Subject:
Feature Modeling for Object Class Recognition, Identification and Categorization in Image

Defense date: 10/05/2016

Advisor: Atilla Baskurt
Coadvisor: Khalid Idrissi

Summary:

Content based image retrieval remains an open issue and the semantic gap still
represents one of the most complicated problems until now. In general, images are
described with low level features as color, texture, shape, etc., but often, those
attributes are insufficient to describe the semantic content of the image.
Many approaches have been proposed in the literature. Some authors place the user in
the retrieval loop and try to model its behaviour in order to understand his waits.
Semantic classification is another way to bridge the semantic gap by classifying
images (e.g. inside/outside, City/landscape, etc.) with the use of relevant low level
features. More recently, some authors [WEI 05] [SUD 05] proposed methods based
on local description (such as SIFT or SURF) combined with generative classifier and
supervised learning (HDP) in order to create models allowing object identification.
Method:
We plan to follow this approach during this thesis. An image can be considered as
containing a set of semantic concepts, similarly that text contains many topics. Images
and concepts are described with different descriptors, and each image (respectively
concept) descriptor will be constituted by a set of sub descriptors of concepts
(respectively images). Images descriptors will be relevantly chosen and normally
computed, but the concepts ones have to be determined automatically.
The objectives of the training step are to detect all the concepts that exist in the
training data set, and to determine the relationships which exist between images
descriptors and concepts descriptors. The test step should allow the retrieval of
database images that are similar to a query.
The Aim of this thesis is to propose, to build and to evaluate a complete system:
Description/Learning/Modeling/Recognition. Face modeling and recognition will be
used as application domain for the developed system.
[SUD 2005] Sudderth E., Torralba A., Freeman A., Willsky. A Describing Visual
Scenes using Transformed Dirichlet Processes. Neural Information Processing
Systems, Dec. 2005.
[WEI 2005] Wei Zhang, Bing Yu, Greg Zelinsky, and Dimitris Samaras. Object class
recognition using multiple layer boosting with multiple features. IEEE Proc. of CVPR
2005, pp II:323-330.