A Study on the Effect of Seam, Color and Boundary Priors on Salient Region Detection

Abstract

Being the best creation of almighty Allah, human beings have the ability to analyse a visual scenario. They not only see an image, but can judge the importance of different parts of a visual area. They can easily differentiate a running car from its background, can tell the color of different objects in an image and can focus attention to some important parts of a visible scene. The value of this visual power of human is easily understood if one thinks about these. Achieving this wonderful human quality using machine is one of the most precious goals of today’s scientific researches. With the continuous advancement of imaging technologies, more and more visual data are being collected all over the world. But, a major portion of these data are left unprocessed. Image processing is used to process these types of visual data. For all image processing techniques, the initial goal is to find some target region for extracting information from the image. Saliency detection is the technique of computationally finding important regions of an image. It is usually done using the contrast information present in an image. Seam map is the combination of cumulative summation of energy values from different directions. In this thesis, a combined method is proposed which uses seam importance map along with boundary aware color importance map. Color importance map is the weighted average of different color channels of Lab color space. Some intermediate combinations which are closer to the proposed optimized version but differ in the optimization technique are also presented in this thesis. Several standard benchmark datasets including the famous MSRA 10k and ECSSD datasets are used to evaluate performance of the suggested method. The proposed saliency model has been compared with several state of the art methods for each dataset. The qualitative and quantitative results from those omparisons make things easier to understand. Besides that, comparison with those state of the art techniques and precision recall curves and F-beta values found from the experiments on several datasets prove the superiority of the proposed method. This robust saliency model can improve almost all types of vision related applications including object detection, robotics, medical image analysis, content aware image resizing etc. In this research, the proposed architecture of saliency detection technique has been applied into a simple implementation of pedestrian detection. Its application has significantly improved the result of pedestrian detection in terms of performance and accuracy.