Piecewise Frequency Domain Visual Saliency Detection

Previous spatial domain methods of visual saliency detection suffer from computational complexity, and recent frequency domain methods lack biological justification. We propose a saliency detection method that combines the speed of frequency domain methods with the topology of biologically based methods. We show that saliency detection can be achieved in frequency domain using frequency domain divisive normalization (FDN). However, this method is constrained by a global surround. Extending this model by conducting piecewise FDN (PFDN) using overlapping local patches overcomes this constraint to provide better biological plausibility. Experiments show that PFDN out-performs FDN and other state-of-the-art methods in eye fixation predication.