A General Framework for Image Retrieval using Reinforcement Learning

In this paper, we present a general framework for image retrieval that is invariant to translation, rotation, scaling, local texture distortion and local geometrical deformation of an image. We first solve the problems in image database retrieval by transforming the image space into the Trace transform space. The Trace transform captures the key characteristics of an object by suppressing variations of an image, while maintaining discriminability. A robust feature from the Trace transform can be thought of as texture representations which can be used in an image retrieval. In addition, the redundant features are ignored by using only the flagged line in the Trace transform with the help of weighted Trace transform (WTT). The optimal parameter of the algorithm is discovered using reinforcement learning for which the within-class variance is minimized. We obtain a high indexing rate of 98.72% on 353 images.