10-08-2012, 02:27 PM
A Co-saliency model of image pairs
cosaliency.doc (Size: 310.5 KB / Downloads: 32)
Abstract
Co-saliency detection is a challenging task in computer vision. This paper gives a method to detect co-saliency from an image pair that may have some objects in common. The co-saliency is a linear combination of the single-image saliency map (SISM) and the multi-image saliency map (MISM). The SISM is designed to describe the local attention. To compute the MISM, a co-multilayer graph is constructed by dividing the image pair into a spatial pyramid representation. Each node in the graph is represented by two types of visual descriptors, which are extracted from a representation of some aspects of local appearance, e.g., color and texture properties. In order to evaluate the similarity between two nodes, we uses a normalized single-pair SimRank algorithm.
Introduction
VISUAL attention is an effective simulation of perceptual behavior, which aims to find a salient object from its surroundings by computing a spatial saliency map. In the past several years, attention models have been successfully applied to many fields, such as
object recognition [1] [2] [3], image segmentation and understanding [4], adaptive coding [5], object tracking [6], image database querying, video retrieval [7] [8].
In this paper, we introduce a perceptual model to describe the similar entity (e.g., a region or object) within an image pair. We will refer to such entity in a pair of images as co-saliency, which is defined as follows: 1) Each region in the pair should have strong local saliency with respect to its surroundings.2) The region pair should exhibit high similarity of certain features (e.g., intensity, color, texture, or shape). Given an image pair, co-saliency is closely related to how we perceive visual stimuli and fixate on the most valuable information from the image pair. Compared with image co-segmentation that partitions an image pair into different segments (i.e., similar regions and backgrounds), saliency as a concept from human vision implies a selection and/or ranking by importance. More precisely, co-saliency is the subjective perceptual quality that makes similar objects in an image pair stand out from their neighbors and capture our attention by visually salient stimuli.
Single Image Saliency Map (SISM)
The salient object detection is very helpful in computer vision and image processing [12]. However, it is still a challenging task to solve this problem. In the current literature, there is no method that can detect the saliency accurately for all images. In order to achieve robust saliency detection, a weighted saliency detection method is proposed, which aims to improve detection performance by combining several saliency maps linearly. Assume I denotes an input image, while Sl represents the corresponding single-image saliency map. We have
Multi Image Saliency map (MISM)
The goal of MISM is to extract the multi-image saliency information from multiple images. Given a pair of images, the multi-image saliency is defined as the inter-image correspondence, which can be obtained by feature matches. The subject is searching for a particular or interesting object, and the attention is geared to react when it appears [13]. Therefore, if the two images contain a similar object, the object region in each image should be assigned high visual saliency values. It means that more visual attention will be attracted by this object. Otherwise, low multi-image saliency values should be considered for the dissimilar regions. According to this principle, the multi-image