The research paper propose a unique methodology for automatic annotation, categorization and annotation-based retrieval of pictures. The new methodology, that we have a tendency to decision Markovian Semantic indexing (MSI), is bestowed within the context of an internet image retrieval system. forward such a system, the users’ queries square measure wont to construct AN mixture Markoff chain (AMC) through that the connection between the keywords seen by the system is outlined. The users’ queries are wont to mechanically annotate the pictures. A random distance between pictures, supported their annotation and therefore the keyword connection captured within the AMC is then introduced. Geometric interpretations of the planned distance square measure provided and its relevancy a bunch within the keyword area is investigated. By means that of a brand new live of Markovian state similarity, the mean 1st cross passage time (CPT), optimality properties of the planned distance square measure tried. pictures square measure sculptured as points in a very vector area and their similarity is measured with MSI. The new methodology is shown to possess bound theoretical blessings and conjointly to realize higher exactitude versus Recall results when put next to Latent linguistics categorization (LSI) and probabilistic Latent linguistics categorization (pLSI) strategies in Annotation-Based Image Retrieval (ABIR)tasks.