frequency-tuned Salient Region Detection
The authors found that most of the methods produced by saliency map are not suitable for image segmentation. Because they produce saliency map resolution is low, the object boundary is not clear, can not uniformly highlight significant objects. This is because the downward sampling narrows the spectral range of the image. In order to solve this problem, the author puts forward five requirements for the algorithm, and proposes a saliency detector suitable for image segmentation.
Five requirements are: emphasis on the largest outstanding objects, uniformly highlight the entire object, the creation of well-defined object boundaries, ignoring the high frequencies of texture,noise and blocking artifacts; output full Resolution's saliency map.
This is really not going to look in this article?? It's the result of the mean minus the Gaussian filter.
Paper reading: frequency-tuned Salient region Detection