Region-based salient object detection of natural images using contour and color spatial compactness prior

Abstract

This thesis deals with the salient object detection model for natural images by combining color contrast, location, contour and color spatial compactness prior to detect attention-grabbing objects. All priors are computed in region-based approach by performing quantized color-based segmentation. Color contrast of a region is computed from the distance between average color in a region and the average color of image, location prior is measured based on relative distance to center of image, contour prior is determined by finding the area inside strong contour and color spatial compactness prior is computed from variance of spatial position of certain color. In the implementation of proposed model, location, contour and color spatial compactness prior is firstly summed together as initial saliency prior before being multiplied with the contrast prior in multiscale computation.

In the experimental evaluation, we demonstrate the salient object detection of proposed model on 1000 images in ASD dataset. By putting the location prior as the baseline, we show the contribution of color spatial compactness and contour prior in improving the saliency detection. Then, based on the precision and recall rates, F-measure and mean absolute error, we show that the proposed model outperforms the state-of-the-art approaches. Moreover, the result obtained through initial saliency prior is closer to ground truth compare to that of the state-of-the-art models.