Salient Region Detection Using Multilevel Image Features

In this paper, we propose a novel salient region detection approach. First, segment the original image into a set of superpixels to extract patch level features using low-level features in the patch. Next, global level features like element uniqueness and

  • PDF / 459,461 Bytes
  • 10 Pages / 439.37 x 666.142 pts Page_size
  • 75 Downloads / 242 Views

DOWNLOAD

REPORT


Abstract In this paper, we propose a novel salient region detection approach. First, segment the original image into a set of superpixels to extract patch level features using low-level features in the patch. Next, global level features like element uniqueness and color contrast are created by previous patch level features. And then both patch level and global level features are gathered to a region to create region level features. Finally, all three level features are utilized to train support vector machines (SVM) classifier, and the trained SVM classifier is used to compute saliency map. The experiment results on the datasets show that the approach we propose performs outstanding in several state-of-the-art approaches. Keywords Element uniqueness SLIC SVM





Color contrast



Saliency detection



1 Introduction The primary goal of computer vision is to understand the surrounding environment utilizing images and videos. In this field, there are three critical works which are correct perception of the main object in the scene, outline recognition of the objects, and access to environmental context of the objects. In order to achieve this goal, saliency detection in the images is the most basic step. Thus, it has turned into an active field in computer vision on account of its various applications, such as image retargeting, object recognition, and object detection. Salient region detection approaches can be usually classified into two groups, which are top-down and bottom-up [1]. Bottom-up approach is stimulus and data-driven, which utilizes lots of low-level features [2], such as texture, intensity and color, since salient objects have strong contrast in comparison with the background. Nowadays, in order to value saliency, many approaches [1, 3, 4] use color Q. Duan ⋅ S. Li (✉) ⋅ M. Mao School of Automation, Chongqing University, Chongqing 400044, China e-mail: [email protected] © Springer Science+Business Media Singapore 2016 Y. Jia et al. (eds.), Proceedings of 2016 Chinese Intelligent Systems Conference, Lecture Notes in Electrical Engineering 404, DOI 10.1007/978-981-10-2338-5_23

233

234

Q. Duan et al.

contrast features along with other high-level features which are called color distribution and uniqueness, since salient objects are spatial compact rather than widely spread background in the image. Except for these features, center prior [1] is another useful feature, which can evaluate salient objects center with color contrast and position. Currently, more and more approaches are using multiple level detection methods, which however operate on either global level [4, 5], region level [6, 7] or patch level [1]. Patch level features usually called low-level features like color, edge, and texture would usually fail to suppress background noises and cannot highlight a salient object more uniformly [8]. While region level features may solve this difficulty a little when smaller salient objects are in the image. If a salient object is large enough, global level features can figure out the entire salient objec