Pattern Extraction by Modeling Image Spatial Relationship

Abstract

In this thesis, a universal framework that is able to extract image spatial relationship among multiple appearance components is proposed, which can be employed to extract additional pattern in wide computer vision tasks. In order to demonstrate its usefulness, three novel algorithms solving different computer vision problems are presented as three main contributions of this thesis, which exercise this framework to improve their performances.

Starting with the object tracking task, the framework is utilized to extract object's inner structure. The algorithm makes use of this inner structure to support a discriminative learning process for mitigating the classic error accumulation effect raising in numerous trackers. In this way, the tracking task is formulated as a prior regularized semi-supervised learning problem. To solve this particular problem, a multi-objective optimization approach is developed. The experiment conducted by the author demonstrates that this tracking algorithm advances state-of-the-art performance of object tracking.

Next, the background subtraction task is studied. In this algorithm, the background is represented by a probabilistic topic model to deal with the dynamic background challenge. This topic model takes advantage of the framework to control topic proportions, which is shown a good descriptor for recurring pixel movement in dynamic background. In order to make the topic model suitable for this on-line task, an incremental learning approach is designed. In the experiment, this background subtraction algorithm outperforms the alternatives in challenging benchmarks.

Finally, the proposed framework is expanded by applying it on a single image processing task, airborne ship detection. The algorithm handles this detection problem by modeling the ocean background and treating the ship pixels as outliers. For simultaneously encoding the dynamic nature and the local similarity of ocean background texture, the framework is used to explore the majority of pixel intensity across the image plane. An extensive experiment shows robustness and accuracy of the ship detection algorithm on a large number of tested images.