Models for visual saliency in images and videos
Gopalakrishnan Viswanath · 2010
Multimedia data consisting of videos and images are in general very rich in content and carry lot of information.With the digital cameras and camcorders becoming easily available to consumers, the amount of information in the form of multimedia data stored and shared across web has increased exponentially.Human beings are normally interested only in a 'subset' of this information available in images and videos, which they consider as salient.Given the huge amount of data to be processed, it is important to obtain information regarding the regions in an image or video which attract human attention so that the overhead of redundant data can be greatly reduced in multimedia applications.The detection of such salient regions or objects in images and videos is the objective of the proposed research work.We investigate the role of features like color and texture in deciding the salient regions in an image by analyzing its distributions in the spatial domain and feature domain.A color saliency framework and an orientation saliency framework is proposed to identify salient regions resulting from color and texture.The dominant colors of any image are modeled as mixtures of gaussians and a saliency value is attributed to each gaussian (color) depending on its 'isolation' in the feature domain and its 'compactness' in the spatial domain.The complexity of orientation histogram in different scales for a local patch is proposed as the texture feature.The orientation saliency map is evaluated by computing the spatial distribution of the different orientations and the contrast of the texture complexity to immediate neighborhood.The final saliency map that best describes the salient regions in an image is selected from the color and orientation maps in a systematic way.Graph representations of images possess great potential in describing the geometric dependencies of various regions in the image.This factor is utilized in our