Saliency Detection by Adaptive Channel Fusion
Kan Huang,Chunbiao Zhu,Ge Li
DOI: https://doi.org/10.1109/lsp.2018.2830755
2018-01-01
IEEE Signal Processing Letters
Abstract:Saliency detection has received tremendous attention from the research community and has become an increasingly important preprocessing tool in many computer vision tasks, such as object recognition. Saliency detection in the frequency domain possesses great advantages compared with that in the spatial domain due to its computational efficiency and succinct model design. Many excellent frequency domain based models, such as SR, phase spectrum of quaternion Fourier transform, hypercomplex Fourier transform (HFT), have made great progress in predicting human eye fixations. However, there still exists a vital drawback among them-ignoring the different contributions of image channels to saliency detection. In our observation, chrominance component of an image may play an unequally important role as luminance component in different situations. In this letter, we propose a frequency domain based framework, which exploits HFT-based saliency algorithm and finds a way to weigh the importance of different channels in saliency detection across space. In the first step, saliency maps corresponding to different channels are generated independently. In the second step, a combination of both saliency maps is performed using an adaptive entropy-based uncertainty weighting approach. We provide an extensive evaluation and show that our proposed method outperforms nine outstanding eye fixation prediction models on four most commonly-used datasets.