Aggregating complementary boundary contrast with smoothing for salient region detection

Abstract

Listen

Translate article

Automatic to locate the salient regions in the images are useful for many computer vision and computer graphics tasks. However, the previous techniques prefer to give noisy and fuzzy saliency maps, which will be a crucial limitation for the performance of subsequent image processing. In this paper, we present a novel framework by aggregating various bottom-up cues and bias to enhance visual saliency detection. It can produce high-resolution, full-field saliency map which can be close to binary one and more effective in real-world applications. First, the proposed method concentrates on multiple saliency cues in a global context, such as regional contrast, spatial relationship and color histogram smoothing to produce a coarse saliency map. Second, combining complementary boundary prior with smoothing, we iteratively refine the coarse saliency map to improve the contrast between salient and non-salient regions until a close to binary saliency map is reached. Finally, we evaluate our salient region detection on two publicly available datasets with pixel accurate annotations. The experimental results show that the proposed method performs equally or better than the 12 alternative methods and retains comparable detection accuracy, even in extreme cases. Furthermore, we demonstrate that the saliency map produced by our approach can serve as a good initialization for automatic alpha matting and image retargeting.