结合整体一致性和局部差异性的图像目标显著性检测

目的 为准确描述图像的显著信息,提出一种结合整体一致性和局部差异性的显著性检测方法,并将显著性特征融入到目标分割中。方法 首先,利用频率调谐法(IG)对目标整体特征的一致性进行显著性检测。然后,引入NIF算法检测显著目标的局部差异性。最后结合两种算法形成最终的显著性检测方法,并应用于图像目标分割。结果 在公认的Weizmann数据集上验证本文方法显示目标的绝对效率并与其他算法对比,实验结果表明本文方法在精确率,召回率,F1-measure(分别为0.445 6,0.751 2,0.576 4)等方面优于当前流行的算法。并且在融合显著性的图像目标分割中,取得满意的实验结果。结论 提出一种新的显著性检测算法,综合体现目标的整体和局部特征,并在公开数据集上取得较高的统计评价。实验结果表明,该算法能够对自然图像进行较准确的显著性检测,并成功地应用于自然图像的目标分割。

Image target saliency detection by combining the global consistency and local difference

Objective Visual saliency, which is a relative property that depends on the degree of difference between a pixel or a region and its background, is difficult to obtain accurately. IG approach, a global-based contrast method, achieves good results on a public dataset. However, this method ignores important local details and inhibits contrast definition between target and background. Thus, this method needs to be improved in terms of the local details. This study proposes a novel saliency detection method based on global consistency and local difference, in which global contrast and local difference are considered two factors for saliency detection. In addition, the application of saliency characteristics to target segmentation is proposed and has achieved better experimental results. Method Different angles show different saliency characteristics. Thus, we compute image saliency from two angles: global feature and local details. The method consists of three basic steps. First, we detect the global consistency by IG, and the global saliency value of the target is achieved by the detection. However, to avoid ignoring the local details and inhibiting the contrast definition between the target and background,we propose the use of the NIF algorithm to overcome the drawbacks of IG. Second, we introduce the NIF method to detect the local difference of the target. Concretely, each pixel is the center pixel, and the square neighborhood of each center pixel is taken as its local scope. Thus, for every pixel, we can obtain corresponding local difference values in relation to its local scope. Finally, we derive the final saliency value by combining the saliency values that originate from the global and local contrasts. Result The saliency map produced by the proposed algorithm achieved a consistent effect compared with artificial segmentation map. To verify the efficiency of the proposed method, experiments are performed using MSRA-1 000 dataset, which is one of the largest public available datasets. Results show that our method outperforms several existing salient object detection methods in terms of precision, recall and F-feasure. In comparison with other popular algorithms, our algorithm has a higher efficiency in image segmentation. Conclusion In this paper, we propose a novel saliency detection method, which treats an image as being composed of global consistency and local difference. We mainly focus on the problem of the local difference of the target and combine both global contrast and local difference for final saliency detection. Compared with some popular methods on a public dataset, our approach achieves the best results in terms of saliency measure evaluation. Furthermore, it obtains good performance in integrating saliency feature in image segmentation. Experimental results show that the algorithm can detect natural images more accurately and can be successfully applied to natural image segmentation.