Image Retargeting with Robust Multi-Scale Epitomic Resizing and Seamless Merging

Image retargeting by summarizing image using bidirectional similarity does not take matching error into account in M-step, causing noticeable blurriness when resizing ratio is small. To handle this problem, we take the following modifications to improve the visual quality of the resized image: Firstly, the matching errors in E-step are taken into account for M-step, which means the patch with larger matching error contributes less to the resized image, and vice versa; Secondly, multi-scale strategy is applied to accelerate the EM convergence; To highlight regions of interest (ROI), we integrate the resized background with the original foreground (ROI) or less resized one by seamless merging through Poisson editing. Experimental results show that new retargeting algorithm can get resized images with perceptually good quality. Considering the patch-based features, it's also suitable for the resizing in spatial scalable video coding.