Automatic region-of-interest extraction in low depth-of-field images

Abstract

Automatic extraction of focused regions from images with low depth-of-field (DOF) is a problem without an efficient solution yet. The capability of extracting focused regions can help to bridge the semantic gap by integrating image regions which are meaningfully relevant and generally do not exhibit uniform visual characteristics. There exist two main difficulties for extracting focused regions from low DOF images using high-frequency based techniques: computational complexity and performance. A novel unsupervised segmentation approach based on ensemble clustering is proposed to extract the focused regions from low DOF images in two stages. The first stage is to cluster image blocks in a joint contrast-energy feature space into three constituent groups. To achieve this, we make use of a normal mixture-based model along with standard expectation-maximization (EM) algorithm at two consecutive levels of block size. To avoid the common problem of local optima experienced in many models, an ensemble EM clustering algorithm is proposed. As a result, relevant blocks, i.e., block-based region-of-interest (ROI), closely conforming to image objects are extracted. In stage two, two different approaches have been developed to extract pixel-based ROI. In the first approach, a binary saliency map is constructed from the relevant blocks at the pixel level, which is based on difference of Gaussian (DOG) and binarization methods. Then, a set of morphological operations is employed to create the pixel-based ROI from the map. Experimental results demonstrate that the proposed approach achieves an average segmentation performance of 91.3% and is computationally 3 times faster than the best existing approach. In the second approach, a minimal graph cut is constructed by using the max-flow method and also by using object/background seeds provided by the ensemble clustering algorithm. Experimental results demonstrate an average segmentation performance of 91.7% and approximately 50% reduction of the average computational time by the proposed colour based approach compared with existing unsupervised approaches.