Research On Object Detection Based On Image Structures

Object detection is a vital part of image understanding. How to handle the variabilities in images is one of the most difficult problems that object detection has to suffer. As the inherent organization manner of objects, structures represent the internal relationships of an object as well as the contextual constraints with the object. Structures are also one kind of the most stable information in images. Therefore, structures can offer great help in effectively dealing with the variabilities in images. This dissertation focuses on object detection based on image structures. The different object detection methods using different image structures are proposed according to different detection task.The main content and innovations are summarized as follows:(1) We propose an object detection approach based on sparse image structure learning. It firstly forms the structural presentation of an image, and then trains a strong classifier using Gentle Boosting algorithms. In the process of detection, we introduce a method, which searches the local maximum of belief, to solve the problem that multiple detection results are obtained from a single object.(2) We propose a silhouette based detection approach. The silhouettes of objects in the train dataset are averaged to define an object template. Then objects are captured through a detection process using TPS transformation. The process begins with a seed region selection. TPS is then applied to locate the difference between the seed region and template, where the boundaries are continually adjusted to approaching the object gradually.(3) We propose a topological hierarchy based detection approach. This approach models the topological relationships among image regions into a hierarchy, and models the contextual constraints of an object as a path in the hierarchy. The potential objects in images can be detected via path matching method.(4) We propose a multi-scale shape context based detection approach. This approach uses the different scale part of an object to form the object's hierarchical presentation. We model the detection as a Bayesian MAP optimization problem, and then transform MAP into the two problems of maximizing likelihood. By maximizing the two likelihood function, the detection task is achieved.(5) We propose a shared structure learning based detection approach. In the learning of shared structure, we use a simple likelihood model. This model does not define any structural form, does not use any code word table, does not specify the owner of features, which provides flexibility for the extraction of various complex structures in images. The shared structure learning is performed in an unsupervised way, with applications in a variety of complex tasks, including multi-object detection task.