Color Spectrum Normalization: Saliency Detection Based on Energy Re-allocation

Abstract

Spectrum normalization is a process shared by two saliency detection methods, Spectral Residual (SR) and Phase Fourier Transform (PFT). In this paper, we point out that the essence of spectrum normalization is the re-allocation of energy. By re-allocating normalized energy in particular frequency region to the whole background, the salient objects are effectively highlighted and the energy of the background is weakened. Considering energy distribution in both spectral domain and color channels, we propose a simple and effective visual saliency model based on Energy Re-allocation mechanism (ER). We combine color energy normalization, spectrum normalization and channel energy normalization to attain an energy re-allocation map. Then, we convert the map to the corresponding saliency map using a low-pass filter. Compared with other state-of-the-art models, experiments on both natural images and psychological images indicate that ER can better detect the salient objects with a competitive computational speed.