A Hybrid Scheme for Detection of Salient Regions

Abstract

Salient region detection intends to find visually important and noticeable regions in visual scenes; as such regions hold significant information and straightforwardly capture human attention. Identified salient objects can be later used in more complex computer vision and image processing applications such as object detection, recognition and tracking, image retrieval, content based image editing & cropping, and image compression. Salient object detection faces challenge in uniformly highlighting desired objects and suppressing irrelevant background. Due to rapid development in technology, wide number of techniques have been proposed for saliency detection, yet there is a need for an effective and reliable saliency detection technique. Existing heuristic methods acquire false detection while dealing with complex scenarios i.e. cluttered backgrounds or foreground object camouflaged in background. In this thesis, an effective framework is proposed to improve the salient region detection in complex scenarios. The contrast of image is enhanced using weighted approximated histogram equalization as pre-processing step. Edge preserving guided filter is used to minimize the unwanted details (texture) while maintaining the edges and semantics. Iterative rolling guidance filter is applied to perform scale-aware local operations for image abstraction. Cellular automata is then used to obtain and optimize saliency cues by exploiting local similarity. A cost reduction framework is further employed to integrate low level cues in order to produce cleaner saliency maps. Proposed technique effectively deals with problems found in saliency detection and produce accurate saliency maps in challenging scenarios where existing techniques fail to claim sharp boundaries of salient object. Visual and quantitative comparisons with state of art existing techniques verify the significance of proposed technique. Salient regions detected by proposed technique can be further used in many image processing areas including object recognition and tracking, image retrieval, automatic cropping and image compression.