SALIENT POINT DETECTION
Abstract
In psychology, visual saliency is the study of how the human visual system collects, processes, and interprets the world around it. Psychologists hope to uncover data pathways in the brain to understand what brain functions take place and where (1). It is believed that the human visual system has low level feature detectors tuned to specific colors, orientations, intensities, etc., which guide attention before higher level cognition occurs. The modeling of this bottom-up process of visual attention is important for both psychologists and engineers. Visual saliency now has applications in robotics, data compression, and anomaly detection, among others. For example, we can guide a robot through an environment so it collects as much useful information as possible by focusing it on salient objects (2); or we can compress an image and budget more bits to describe salient image patches; or we can find salient anomalies in audio or video data.
Here we focus on two methods of detecting salient points in still images. As an example, consider the task of finding a vertical bar among horizontal bars. If such an image were flashed in front of you the single vertical bar would immediately "pop out" and attract your attention. The vertical bar is considered salient. We hope to solve generalizations of this problem by finding salient structures in both synthetic and real images.
By testing the saliency algorithms on synthetic data, such as the pop-out example given or a general line orientation test, the computer algorithms performance can be compared to human performance. If the two perform similarly perhaps the algorithm can reveal how a human might execute the task. A real world application that has arisen during this research is image ranking based on saliency. Unmanned aerial vehicles (UAVs) now commonly collect massive amounts of image and video surveillance data that needs to be sifted through and understood. By ranking the collected data in order from most salient to least salient, the saliency algorithms can reduce the time for a human to review the most important data.
The rest of this paper will be organized as follows: Chapter 1 provides a cursory review of the human visual system. Chapter 2 will describe the Itti et al model (3) along with improvements we have made. Chapter 3 will discuss a new clustering approach to saliency and Chapter 4 will detail the results of computer simulations
2
and compare against human performance. This paper concludes with a discussion of the results in Chapter 5.