Saliency Detection for Content Aware Computer Vision Applications

In recent years, there has been an increased scope for intelligent computer vision systems, which analyse the content of multimedia data. These systems are expected to process a huge quantum of image/data with high speed and without compromising on effectiveness. Such systems are benefited by reducing the amount of visual information by selectively processing only a relevant portion of the input data. The core issue in building these systems is to reduce irrelevant information and retain only a relevant subset of the input visual information. To address this issue, we propose a region-based computational visual attention model for saliency detection in images. The proposed model determines the salient object or part of the salient object without prior knowledge of its shape and color. The proposed framework has three components. First, the input image is segmented into homogeneous regions and then smaller regions are merged with neighbouring regions based on color and spatial distance between them. Second, three attributes such as spatial position, color contrast and size of each region are evaluated to distinguish salient object/parts of salient object. Finally, irrelevant background regions are suppressed and the region level saliency map is generated based on the three attributes. The generated saliency map preserves the shape and precise location of salient regions and hence it can be used to create high quality segmentation masks for high-level machine vision applications. Experimental results show that our proposed approach qualitatively better than the state-of-the-art approaches and quantitatively comparable to human perception.