ROI image retrieval based on multiple features of mean shift and expectation–maximisation

• An image retrieval by using MS and EM to adaptively change the location and scale of the region of interest (ROI).
• EM-like merged into the colour and spatial features to detect the scale and position of the ROI.
• A novel similarity measure based on the colour and spatial histograms over the lattice structure.
• Huge speed up the EM-like iteration due to the proposed similarity measure.

In this paper, a novel region of interest (ROI) query method is proposed for image retrieval by combining a mean shift tracking (MST) algorithm and an improved expectation–maximisation (EM)-like (IEML) method. In the proposed combination, the MST is used to seek the initial location of the target candidate model and then IEML is used to adaptively change the location and scale of the target candidate model to include the relevant region and exclude the irrelevant region as far as possible. In order to improve the performance and effectiveness using IEML to track the target candidate model, a new similarity measure is built based on spatial and colour features and a new image retrieval framework for this new environment is proposed. Extensive experiments confirm that compared with the latest developed approaches, such as the generalized Hough transform (GHT) and EM-like tracking methods, our method can provide a much better performance in effectiveness. On the other hand, for the IEML, the new similarity measure model also substantially decreases computational complexity and improves the precision tracking of the target candidate model. Compared with the conventional ROI-based image retrieval methods, the most significant highlight is that the proposed method can directly find the target candidate model in the candidate image without pre-segmentation in advance.