Video Saliency Detection Incorporating Local Descriptor

Abstract

Human eye and brain is one of the most important parts of the Human Vision System. The eye captures visual data and transmit it to the brain for making it more informative to the respective person. Human Vision System has very effective and important characteristic in the field of computer vision. Human Vision System has to grape the most salient regions which can help the people to understand the contextual information. We proposed a framework to detect the visual saliency by combining both spatial and temporal saliency by using pre and post processing so that the most significant portions of the image are extracted from the image. It detects the spatial saliency map using feature base techniques and apply local descriptor to refine the output image for the final result. It then calculates the optical flow by using latest technique of contrast enhancement Non-parametric modified histogram equalization and edge detection Spatial stimuli gradient sketch model and finally calculates the motion contrast by using the optical flow result. It binarize the motion contrast result with help of OTSU technique to determine the finale temporal saliency. It then calculates the uncertainty of the output map and fuse the uncertainty with the spatial and static saliency map to find the final video saliency result. The improved results are shown in the experimental portion of this thesis. The results are evaluated by both qualitatively and quantitatively by different techniques like Kullback-Leibler divergence, Normalized Scanpath Saliency, Correlation Coefficient and Area Under the Receiver Operating Characteristics Curve. The results of each dataset are shown in the form of table.