| Image segmentation plays a key role in a range of subfields of computer vision, including content-based image retrieval, object recognition, mobile object tracking, medical imaging, etc. However, strong segmentation, which corresponds to the process of partitioning an image into meaningful regions, remains a difficult and as yet unsolved problem. Bottom-up approaches utilize color, texture and other low-level visual features to partition images into homogeneous regions. However, regions that are homogeneous in low-level visual content do not necessarily (and usually do not) correspond to meaningful objects. This is mainly due to the lack of correspondence between low-level visual features and high-level semantics, which is known as the semantic gap. Hence, in recent years, much research interest on this topic has been focused on top-down designs that introduce high-level cues into the segmentation process.;In the context of content-based image retrieval, relevance feedback learning has been successfully used in the past as a means of reducing the semantic gap. Inspired by this, we developed in this thesis an adaptive image segmentation framework that achieves a task-dependent top-down adaption of the scale parameters of the mean shift-based segmentation algorithm. Different from previous learning-based segmentation schemes, the proposed method requires neither manual segmentations as training samples nor prior object-specific knowledge for parameter learning. This is made possible based on the assumption that the visual appearance of a particular object model has a Unique distribution in the feature space. Thus, once a quantized representation of this distribution is obtained through a learning-based means, an improved segmentation for images of the same object model can be conducted. More specifically, under the context of a content-based image retrieval system, each image in the retrieval set is labeled either positive or negative depending on the presence of an object of interest. Extracted features of labeled images constitute the training samples for multiple instance learning, which in turn induces a mapping from the object of interest to the scale parameters of the mean shift-based segmentation algorithm. Once the mapping is established, it is incorporated into the segmentation procedure so as to improve the performance for images of the same object model. Experimental results indicate both the capability and flexibility of our proposed method for practical usage. |
