Saliency Detection Based on a Novel Color Channel Volume and Background Likelihood Weight

The objective of saliency detection is to extract a salient map in a natural image. However, the detection of salient regions located by the object is challenging due to some factors such as multi-object overlapping, salient object touching the image boundaries, and salient region on color, texture, and shape. To alleviate those issues, we proposed a two-phase strategy saliency model for saliency detection by using a robust color channel volume and background likelihood weight. Specifically, we first linearly combine several color space channels of Lab color space to construct a low-level image features extractor called color channel volume which is brilliant to stand out a salient region but suppress a background region. Then, based on super pixel segmentation, we design a robust background region measure called background likelihood weight (BLW) to find the background region from the color channel volume and obtain an initial salient map by removing the found background region. Finally, we proposed improved multimodal histogram thresholding to refine the initial salient map and further extract a final salient map. Experimental results on two public image datasets show that our approach achieves better performance results of saliency detection than several other saliency detection algorithms in metrics.