Saliency Detection Based on Frequency and Spatial Domain Analysis

As a component of low-level vision processing, saliency detection facilitates subsequent processing such as object detection or recognition. In this paper, we argue that a reasonable saliency detector should have the ability to: (1) Detect both small and large saliency regions. The size of salient regions vary greatly. When the salient region is large, however, because center-surround algorithms mainly use local information, they will respond heavily in boundary regions, where the texture, intensity or other features are locally different. (2) Detect saliency in cluttered scenes. Another drawback of local information-based saliency models is that heavily textured regions are always highlighted. Cluttered scenes are still a challenge for models depending on local information and some based on global information. (3) Inhibit repeating patterns. Objects in scenes viewed by the human visual system are thought to compete with each other to selectively focus attention on a subset [5]. These repeating patterns will suppress each other and then be inhibited. In this paper, inspired by [3, 4], we propose a new saliency detection model by combining global information from frequency domain analysis and local information from spatial domain analysis. In the frequency domain analysis, instead of modeling salient regions, we model the nonsalient regions using global information. Thus those so-called repeating patterns that are not distinctive in the scene are suppressed by using spectrum smoothing. In the spatial domain analysis, we enhance those regions that are more informative by using a center-surround mechanism similar to that found in the visual cortex. Finally, the outputs from these two channels are combined to produce the saliency map. Frequency Domain Analysis Frequency analysis presents an opportunity to deal with the global information in an image. In this paper, we investigate the relationship between the amplitude spectrum and nonsalient regions in the image. However, instead of searching for these socalled distinctive patterns, we model the regular patterns (repeating patterns) that would not attract much attention by our visual system. We refer to these as being non-salient. It is argued in [3] that the spectrum residual corresponds to the saliency in an image, while contradictorily in [2], the amplitude information was totally abandoned. However, in this paper, we illustrate that the amplitude spectrum also contains important information corresponding to image saliency. To be more exact, spikes in the amplitude spectrum correspond to repeating patterns, which should be suppressed for saliency detection. A Gaussian kernel h can be employed to suppress spikes in the amplitude spectrum (implemented by a log amplitude spectrum instead of using the amplitude spectrum) as follows: