融合双层信息的显著性检测

目的 针对已有工作在颜色及结构显著性描述方面的缺陷,提出一种新的图像显著性检测方法。方法 本文方法在不同的图像区域表达上从颜色与空间结构角度计算图像的显著性,充分考虑图像的特征与像素聚类方式之间的适应性。首先,根据颜色复杂度、边缘与连通性等信息,将图像从像素空间映射到双层区域表示空间。然后,根据两个层次空间的特性,与每个图像区域的边界特性,计算图像的结构和颜色显著度。最后,由于不同图像表示中的显著性信息存在互补性,将所有这些信息进行融合得到最终的显著性图。结果 在公认的MSRA-1000数据集上验证本文方法并与目前国际上流行的方法进行对比。实验结果表明,本文方法在精确率、召回率以及绝对误差(分别为75.03%、89.39%、85.61%)等方面要优于当前前沿的方法。结论 提出了一种融合双层信息的显著性检测算法。根据图像本身信息控制区域数目构建图像双层表示,提高了方法的普适性;利用图像不同层次的特性从不同角度计算显著性,增强了方法鲁棒性。

Image saliency detection based on two-layer information fusion

Objective Image saliency detection is a method used to eliminate the redundant image information. Moreover, this method is used in many computer vision applications, such as adaptive compression of images, content-aware image editing, and image retrieval. In this study, a new image saliency detection method is proposed to compute for image saliency from different perspectives. In fact, many methods are used to compute for saliency, and most of these approaches use different types of features to detect saliency in single regional representation. However, only a few methods consider the adaptability between the feature and image representation. Method According to the different characters of different types of regional representations, we compute image saliency from different angles by using a wide variety of information, including color. The method consists of three basic steps. First, the image is mapped from the pixel space to a two-layer regional representation space on the basis of connectivity and edge information. The first layer is related to the spatial structure of the image, whereas the second one is superior in describing color information. Then, on the basis of the diverse properties of the constructed two-layer representations, we adopt a number of features to abstract image saliency. In the first layer, we use the spatial distribution of region in the image and the structure feature to obtain the spatial structure saliency. In the second layer, we use the color feature to compute for color saliency. Given the complementarity between the two kinds of saliency, the last step is to integrate the two kinds to obtain the final saliency map. In practice, color saliency has higher significance and discriminative power than spatial structure saliency. Thus, we use an exponential function to combine the two kinds of saliency while highlighting color saliency. In addition, the boundary prior is also a reasonable and popular method for enhancing saliency detection and has thus been widely used for image saliency detection. In contrast to existing methods that set a region containing boundary pixels directly to the background, we employ the percentage of boundary pixels in each region to adjust saliency values. Result Given that the extracted saliency clues correspond to the attributes of the local image regions quite well, our method has several advantages over existing methods. To verify the efficiency of the proposed method, experiments are performed using the MSRA-1000 dataset, which is one of the largest publicly available datasets. Results show that our method outperforms state-of-the-art methods in terms of precision, recall, F-measure, and mean absolute error. Conclusion Image saliency detection is a promising approach in the field of image processing and analysis. This study presents a new saliency detection method based on two-layer regional representation through both color saliency and spatial structure saliency. Experimental evaluation results suggest that our method outperforms other methods in image saliency detection.