Fast visual saliency detection with bisection search to scale selection

Inan image, objects are meaningful entity only at certain scales; however appropriate scales of object cannot be predicted a priori. The description of the image at multiple scales is thus necessary for salient objects detection. Multiple computations involved in this process make the model slow. To reduce the computational complexity bisection search to scale selection is proposed in this paper. The saliency detection with Hypercomplex Fourier Transform is made faster with the proposed algorithm. Minimum and maximum scales act as distinct seeds. To begin with, saliency maps and their entropy are obtained for these seeds. A set of lower half or upper half scales is selected recursively based on minimum entropy for further search. A saliency map of approximate minimum entropy is obtained efficiently. The proposed method reduces computational complexity from O (N) to O (log₂N+1) compared to the HFT method in the literature.