Algorithms for image saliency via sparse representation and multi-scale inputs image retargeting

Abstract

Saliency detection is an important yet challenging task in computer vision. In this report we investigate the use of sparse coding over redundant dictionary for saliency detection. We attempt to present a small fraction of the growing knowledge regarding sparse representation over redundant dictionary and discuss some potential usage of this powerful tool for saliency detection task. We propose a new algorithm for saliency detection based on the likelihood that an image patch can be encoded sparsely using a dictionary learned from other patches. experimental results based on saliency ground of truth of 1000 real images shows a superior performance of the new algorithm in comparison with other existing saliency algorithms. We also propose an image retargeting algorithm which is capable of combining the strength of the Shift-map framework and warping-based algorithms. The Shift-map algorithm experiences problems with extreme resizing ratio: important objects might be removed due to limited space in the output. We tackle this problem by introducing a stack of multi-scale inputs. This kind of input allows the Shift-map framework to produce output with great flexibility: regions can be removed or scaled in order to achieve the optimal and desired retargeted image. Experiments are conducted based on a benchmark image database to demonstrate potential power of this approach.