7

Applications in Computer Vision, Image Retrieval and Robotics

Computer vision is a key area in computer science and electronic engineering. In computer vision, the image or video information is acquired by the aid of digital photoelectric sensors (image or video cameras) that act like a human or animal eye, then the visual information is processed by computer software or hardware in order to obtain object detection and recognition in scenes, object tracking and scene understanding and so on. Computer vision can be used in many real-world scenarios such as military target detection, diagnosis with medical images, video surveillance, identity recognition, industrial automation, remote-sensing imagery processing, human computer interface, image retrieval, and so forth. For robotics, computer vision techniques are applied in automatic localization and navigation with the help of camera sensors on the robots. Although there are a lot of methods and algorithms used in real-world environments where the system receives a mass of ongoing information from cameras, the limited computing ability and memory of the system are still the obstacle in practice. The human or animal vision system has an attention mechanism that can select the most important information as the focus from real-world environments and allot the focus for the limited resources of the brain to process it. Obviously, a computer vision system also requires the capability to focus on task-relevant events.

The applications of visual attention models in computer vision emerged during 1990s, almost at the same time as the development of computational visual attention models themselves. The applications were for object detection and recognition first in natural images (or videos) [1–7] and then in complex satellite imagery (i.e., remote sensing and radar imagery [8–11]). More recently, object detection methods have been combined with the visual attention concept, called salient object detection [12–20].

Along with the rapid development of computer, internet and multimedia technology, image retrieval has become a hot topic in computer science. Since visual attention can capture the salient region of an image, that is in general related to the content of image, another application of visual attention modelling is for content-based image retrieval in larger image databases [21–23].

The applications in robotics are more of an open issue. In the initial stage, the robots had no camera in their systems. The information come only from various sensors of distance measurement, such as laser or infrared sensors, to detect the robot's location in the environment, or acoustic sensors (receiving commands from humans) or computer instruction to complete simple operations (as in industrial automation). With the enhancement of computational ability and augmentation of the memory size of computers, people expect a robot to do more complex tasks, such as to stimulate the development of self-localizing humanoid robots with vision and movement capabilities. The challenge of humanoid robots is not only to solve general computer vision issues such as object detection and recognition in still scenes, but also to adapt to the environmental changes when a robot is moving. In addition, the input of abundant continuous visual images received by a moving robot may result in memory overflow and system breakdown. In that case, adding visual attention in robot systems is very necessary. Much literature discusses and explores the behaviour of robots with cameras, such as self-localization, landmark recognition and drawing the surrounding map, with the help of visual attention computation [24–35].