Segmentation and region of interest based image retrieval in low depth of field observations

In this paper we address the problem of extracting the focused region and its use in retrieving similar images from a low depth of field image database. We compute the histogram of the local contrast at each pixel and model it as a mixture of two exponential distributions – one for the focused and the other for the defocused region. Unlike the mixture of Gaussian distributions, a mixture of exponential distributions overlaps with same monotonicity over the entire range in [0, ∞) and it is difficult to separate into components. We estimate the parameters of these distributions using the EM algorithm. This is followed by a hypothesis testing which segments the focused region in the low depth of field image. A content-based retrieval scheme is now confined to the detected region for a proper retrieval. Experimental results for both segmentation and image retrieval using a database consisting of 4986 images are presented to show the efficacy of the suggested scheme.