Image Saliency Detection In Frequency Domain

Saliency detection is widely used in many computer vision applications, such as im-age segmentation, object recognition, image retargeting, video compression, etc. Being simple, fast, effective, and independent of categories, or other forms of prior knowledge, saliency detection in the frequency domain (spectral saliency) attracts more study since it was presented. This paper summarizes the detailed procedure of saliency detection algo-rithm in the frequency domain and reviews the state-of-the-art spectral saliency detection models. Through analysis, it is found that many computational methods can not uniformly highlight the salient objects and satisfactorily detect the saliency details.Assuming that a natural image consists of several salient regions and many repeated patterns (nosaliency), the spikes in the amplitude spectrum turn out to correspond to re-peated patterns in spatial domain. If the spikes in the amplitude spectrum are smoothed at an appropriate scale, the nonsalient patterns will be suppressed, thereby producing the pop-out of the salient objects. To address this issue, a novel saliency detection model-salient object detection via automatic adaptive amplitude spectrum analysis is proposed in this paper. The main contributions are shown as follows:1. Presenting the specific relation between optimal scale for smoothing the amplitude spectrum and the size of salient region. This relation plays a very important role in the research of salient region detection in frequency domain.2. Proposing a method of automatic optimal scales selection based on the Spectrum Scale-Space. This method can highlight the salient objects uniformly.3. Introducing a strategy of adaptive weighted combination. This new strategy can retain meaningful saliency information by incorporating different saliency maps adaptively.The performance evaluation of quantitative and qualitative comparisons on four chal-lenging popular benchmark data sets validates that the output of the proposed model is more uniform and precise when compared with other spectral saliency models.