Due to the huge amount of potentially interesting documents available over the Internet, searching for relevant information has become very difficult.  Since image and video are  a major source  of these  data, grouping  images into (semantically)  meaningful  categories using  low-level  visual features is an important (and challenging) problem in content-based image retrieval.  Using  Bayesian classifiers, we attempt to capture high-level concepts from low-level  image features.  Specifically, we have developed Bayesian  classifiers  for semantic   image  classification  (indoor  vs.  outdoor,   city vs.  landscape,  and sunset vs. forest vs. mountain), image orientation detection, and object detection (detecting regions of sky   and   vegetation   in   outdoor   images).   

We   demonstrate   that   a   small   codebook   (the   optimal  codebook  size is selected using a modified MDL criterion) extracted from a learning vector quantizer  can be  used to estimate  the class-conditional  densities  of the  observed features needed for  image  classification.  We  have developed an incremental  learning  paradigm,  a feature selection scheme,  a rejection   scheme,   and   a   classifier   combination   strategy   using   bagging   to   improve   classifier performance.  Empirical  results on a large database (400  images) show that semantic  categorization and organization of the database using  the proposed classification schemes improves both retrieval  accuracy and efficiency.