Probabilistic Contextual Models For Object Class Recognition In Uncontrived Images

Loading...
Thumbnail Image
Date
2011-05
Authors
Hasanat, Mozaherul Hoque Abul
Journal Title
Journal ISSN
Volume Title
Publisher
Universiti Sains Malaysia
Abstract
Context is a vital element in deriving meaningful explanation of an image for both biological, as well as, artificial vision systems. This thesis proposes to model contextual relation among real-world objects in uncontrived images in order to improve object class recognition performance. Two probabilistic models are proposed – Semantic Context Model (SCM), and Spatial Context Model (SpCM) to model high-level semantic and spatial contextual relations respectively. SCM learns a directed graph structure from a given dataset of uncontrived images to model the dependency relation among real-world objects. The nodes of the graph represent the objects of the problem domain and the directed edges represent the dependency relation between a pair of objects. With respect to SpCM, it learns probability distributions of pair-wise spatial relations for all the objects in the problem domain. Both SCM and SpCM can learn contextual relation independently of each other and of any other learning process within an object class recognition system. This allows for modular integration of these models with an existing object class recognition system. In this regard, this thesis also proposes a framework dubbed as ConVes that integrates the proposed models with object class recognition systems. The performance improvements achieved due to the usage of the proposed models were compared against two local appearance-based recognition systems. Performance metrics used to evaluate the recognition performance are: confusion matrix, accuracy, precision, recall, F-measure, ROC curve, and area under the ROC curve.
Description
Keywords
Probabilistic Contextual Models , Object Class Recognition
Citation