Investigation On Location-Aware Image Saliency Algorithm

Image saliency detection is a basic problem in the field of computer vision and image analysis. Related algorithms have been widely used in objects recognition and tracking, image retrieval, image segmentation, video compression and so on. In this paper, based on the analysis and study of the literature algorithms, some new ideas for saliency detection are proposed which can provide more reasonable detection results.In the first part of the paper, we investigated the current achievements and a key issue in this field. The existing saliency detection methods usually assume that the salient object is on the center of the image and therefore incorporate a center-bias assumption in the design of their algorithms. Obviously, this is not always true, especially for those imageries acquired by unmanned monitoring system or device(e.g., surveillance camera), in which the salient object could appear in any location within the image. Consequently, the resulted saliency detection performance could be greatly degraded. In particular, the Hypercomplex Fourier Transform(HFT) based saliency detection algorithm is investigated, which incorporates a ‘Center-bias Setting’ in computing the optimal scale for saliency detection. To remove the unreasonable assumption, a location-aware strategy is exploited to identify the optimal scale. Correspondingly, a location-self-aware algorithm and a location-prediction-aware algorithm are proposed based on this strategy. Extensive simulation results show that the proposed algorithms clearly outperform the existing state-of-the-art algorithms on saliency detection.In the second part of the paper, we investigated the influence of the used color space for the location-aware saliency detection. Most existing saliency detection algorithms are designed in the RGB color space. However, other color space, such as Lab color space, might be more suitable for saliency detection. In the Lab color space, there are one luminance channel and two color channels with 2D independent distribution, which makes it possible to extract luminance and color features separately. In order to take use of the RGB and Lab color space at the same time, a location-aware saliency detection algorithm in combined color space is proposed. The location-aware strategy is employed to identify the optimal scale in the HFT algorithm both in the RGB space and the Lab space, and then the saliency map with minimum entropy is selected as the final output, which outperforms the existing HFT algorithm clearly.