Interactive image retrieval using self-organizing maps

Abstract

Digital image libraries are becoming more common and widely used as visual information is produced at a rapidly growing rate. Creating and storing digital images is nowadays easy and getting more affordable all the time as the needed technologies are maturing and becoming eligible for general use. As a result, the amount of data in visual form is increasing and there is a strong need for effective ways to manage and process it. In many settings, the existing and widely adopted methods for text-based indexing and information retrieval are inadequate for these new purposes.

Content-based image retrieval addresses the problem of finding images relevant to the users' information needs from image databases, based principally on low-level visual features for which automatic extraction methods are available. Due to the inherently weak connection between the high-level semantic concepts that humans naturally associate with images and the low-level visual features that the computer is relying upon, the task of developing this kind of systems is very challenging. A popular method to improve retrieval performance is to shift from single-round queries to navigational queries where a single retrieval instance consists of multiple rounds of user-system interaction and query reformulation. This kind of operation is commonly referred to as relevance feedback and can be considered as supervised learning to adjust the subsequent retrieval process by using information gathered from the user's feedback.

In this thesis, an image retrieval system named PicSOM is presented, including detailed descriptions of using multiple parallel Self-Organizing Maps (SOMs) for image indexing and a novel relevance feedback technique. The proposed relevance feedback technique is based on spreading the user responses to local SOM neighborhoods by a convolution with a kernel function. A broad set of evaluations with different image features, retrieval tasks, and parameter settings demonstrating the validity of the retrieval method is described. In particular, the results establish that relevance feedback with the proposed method is able to adapt to different retrieval tasks and scenarios.

Furthermore, a method for using the relevance assessments of previous retrieval sessions or potentially available keyword annotations as sources of semantic information is presented. With performed experiments, it is confirmed that the efficiency of semantic image retrieval can be substantially increased by using these features in parallel with the standard low-level visual features.