Abstract
The advent of digital technology and the Internet requires an efficient approach to the retrieval and indexing of large volumes of diverse images. Traditional content-based image retrieval (CBIR) approaches cannot satisfy potential high-level application tasks where semantics play an important role. Humans are accustomed to replying on immediate semantic impression generated the moment they observe images, thus the absence of semantics will limit the potential effective applications of image databases. Automatic semantic content recognition has become an open research issue; great difficulties in deriving semantics from primitive image features have constrained semantic based image retrieval (SBIR) to work within comparatively small sets of images. This thesis is an attempt to offer a systematic approach to automatic semantic content recognition in a large-scale image archive. The general research issues of realizing automatic semantic content recognition in a large-scale image database are discussed and explored. Due to the complicated and unpredictable variability in broad image databases, the capture of high-level features is attempted from different visual perspectives and classification theories, exploring techniques and theories of multiple classifiers. The solution proposed here is an improved classifier combination strategy which has potential generic scope. Domain knowledge plays an important role in the disambiguation and recognition of difficult patterns. Through exploring knowledge models existing in image interpretation, global information and spatial contexts are acquired and generalised through a knowledge elicitation subsystem. This knowledge is formalized and modelled in a Markov Random Field (MRF) based framework, with parameter estimations constructed from combining multiple classifiers. An optimal solution is implemented guided by global information and agreed structure analysis. A large-scale histological image database is chosen as the test bed in this research, which has shown encouraging empirical results with improved generalisation performance. Further benefit has additionally been demonstrated in a semantic based image retrieval system.