Background contrast based salient region detection

Abstract

This paper proposes a novel salient region detection method based on background contrast. Salient regions are widely considered as those distinct to the image background. However, owning to the lack of information about the image background, existing methods measure saliency of an image pixel/region using its contrast to local neighborhoods or the entire image rather than the image background. Inappropriate contrast region leads to difficulty in highlighting the whole large salient region. In this paper, we discover that the absence of eye fixations provides an important clue to the image background. Regions without eye fixations are very likely the image background. To further acquire the spatial extent of image background, the complementary area of the convex-hull of eye fixations is used to represent the possible image background. Then we measure the saliency of each region by computing its contrast to the estimated image background. The experimental results demonstrate that our approach outperforms previous state-of-the-art methods.

Introduction

Humans can rapidly and accurately identify salient regions in their visual fields. Simulating such an ability in machine is critical to make machine handle visual content like humans. In the past decades, extensive research efforts have been conducted to construct the computational model for detecting salient region. Most of these methods [1], [2], [3], [4] intend to predict human eye fixations, where the eye fixates for a while to gather and process interesting information. Itti et al. [1] directly summed multi-scale center-surround responses of visual features to generate saliency map. Murray et al. [2] integrated the information at multiple scales by performing inverse wavelet transform on the scale-weighted center-surround responses. Hou et al. [3] measured the saliency of each pixel in the frequency domain. Ban et al. [4] introduced psychological distance into the prediction of eye fixations. However, eye fixations are just some spatial discrete points, they cannot accurately represent the whole salient region [5], [6].

Recent years have witnessed more interest in salient region detection, which is useful in applications such as image segmentation [7], object detection [8], and content aware image resizing [9]. We focus on data-driven salient region detection.

It is widely agreed that salient regions are distinct to the image background [5], [6], [10]. However, due to the lack of any knowledge about the image background, current methods measure saliency of an image pixel or region using its contrast with respect to local neighborhoods [11], [5], [6], [12] or the entire image [13], [14], [15], [16], [17]. Though these methods have achieved promising results, they still have difficulty in highlighting the entire salient region uniformly (Fig. 1(1)). Besides, in the presence of large salient objects or complex backgrounds, they may fail to correctly highlight the salient regions (Fig. 1(2)).

Though the eye fixations cannot mark the whole salient region [5], [6], the absence of eye fixations provides a clue to image background. Observing the outside world, we move our fovea in the retina to specific salient regions to fixate on them [18]. During the fixation, we get a detailed impression of them. We then move, or saccade, to the next location of interest. Thus human eye fixations always fix on the interesting parts of image [19], seldom on the image background. The regions without eye fixations are very likely the image background.

In this paper, guided by the eye fixations, we estimate the spatial extent of image background. Then we measure the saliency of each small image region using its contrast to the estimated image background. In Section 2 of the paper, the limitations of existing methods are described. Section 3 introduces the proposed salient region detection approach. Experimental results and conclusion are respectively presented in 4 Experimental results, 5 Conclusion.