A Multiscale Image Saliency Detection Method

Abstract:

In order to improve the accuracy and robustness of the saliency detection algorithm, this paper proposed a multiscale image saliency detection method. First, the smoothing algorithm was adopted to filter out the noise characteristics in the image. Then, the multiscale representation of an image was performed and saliency maps were computed at different scales. Finally, according to the conditional random field theory, the saliency detection results at different scales were weighted together to get the final results. Extensive experiments in which the proposed method was compared with 9 existing stateoftheart methods on five benchmark data sets, ECSSD and MSRA10K, show that the proposed method performs better in terms of various evaluation metrics.

Key words:saliency; multiscale fusion; conditional random field