Multiscale Phase Spectrum Based Salient Object Detection

Automatic image segmentation is emerging field in image processing research domain. In a visual scene, the objects which are different from their surroundings get more visual importance and get high gaze attention of the viewer. Saliency detection predicts significantly important regions in a scene, which should be considered for further processing according to specific applications. There are several applications where saliency detection is used as core modules such as object based surveillance, content adaptive data delivery for low data rate systems, automatic foveation system. There are two different approaches to predict the viewer’s gaze in a visual field: topdown approach and bottom-up approach. Top-down approach is target driven while the bottom-up method is independent of target and stimuli driven. Hardware based eye tracker devices are also commercially available but the cost is comparatively very high. In this paper, an efficient multiscale phase spectrum based salient object detection method is proposed. It is observed that a fixed scale of the original image may not predict properly the salient objects. Saliency predicted in one resolution may not predict the same fixation region on another resolution. It is proposed to apply saliency detection to multiple scales of the original image. Saliency is detected using phase spectrum of Fourier transform as positional information is contained in the phase spectrum while amplitude spectrum contains the presence of frequency components. The proposed method performs much better than other previous methods and predicts more precisely salient objects. Simulation results of six state-of-art techniques for salient object detection are analyzed along with the ground truth and compared against the proposed method. The performance of the proposed method is measured on the basis of objective and subjective analysis. The simulation verifies that the proposed method is suitable candidate for prediction of salient objects.