基于背景原型对比度的显著性物体检测

Saliency Detection via Background Prototypes Contrast

摘要:

针对传统显著性模型在自然图像的显著性物体检测中存在的缺陷,提出了一种利用背景原型(background prototypes)进行对比的视觉关注模型,以实现显著性物体的检测与提取；传统显著性模型主要通过计算区域中心与四周区域差异性实现显著性检测,而自然场景中显著性区域和背景区域往往都存在较大差异,导致在复杂图像中难以获得理想检测效果；基于背景原型对比度的显著性物体检测方法在图像分割生成的超像素图基础上,选择距离图像中心较远的图像区域作为背景原型区域,通过计算图像中任意区域与这些背景原型区域的颜色对比度准确检测和提取图像中的显著性物体；实验结果表明,基于背景原型对比度的显著性模型可以更好地滤除杂乱背景,产生更稳定、准确的显著图,在准确率、召回率和F-measure等关键性能和直观视觉效果上均优于目前最先进的显著性模型,计算复杂度低,利于应用推广。

Abstract:

To overcome the disadvantages of existing saliency models in saliency detection, a novel object-based attention model is presented to predict visual saliency using the contrast against the background prototypes. Traditional saliency models mainly detect salient regions by comparing the differences between center and surround regions, which makes hard to get desired results in complex scenes for significant differences often appear both in salient and background regions in real images. Saliency detection via background prototypes contrast firstly over-segment the input image into perceptually homogeneous superpixels, and automatically identifies a series of regions far away from image center as background prototypes. The visual saliency is then accurately calculated using the color contrast with respect to the selected background prototypes. Promising experimental results demonstrate that the proposed model, which outperforms the compared state-of-the-art saliency models in average precision, recall, F-measure and visual effect, can better exclude the cluttered backgrounds, and thus produces more robust and accurate saliency maps. Moreover, due to its computational efficiency, our model is easy to be widely applied.