Three-dimensional scene recovery from image defocus

Abstract

Three dimensional scene information consists of geometric and photometric information. The distance and shape of objects in a scene constitute the geometric information, while color and image irradiance of objects constitute the photometric information. Active methods of scene recovery use energy beams such as laser, infrared light and sonar. Passive methods, on the other hand, use only the ambient illumination of the scene. This dissertation deals with a new passive method of scene recovery using defocus information. The method is based on a new spatial domain deconvolution formula. It requires only two blurred images taken with different camera parameters. The camera parameters include position, focal length and aperture size of the optical system in the camera. The computations in the method are local and hence it can be used for obtaining the depth-map of a scene in real-time. The method has been implemented on a prototype camera system and a large number of experiments have been performed. The experiments indicate that the method is useful in practical applications such as autofocusing and robot vision. The method has also been extended to continuously focus on moving objects. A simple camera configuration has been suggested for real-time focusing on moving objects. This dissertation also deals with the problem of recovering the focused image of a scene from its blurred images. A spatial domain technique and a Fourier domain technique have been investigated for focused image recovery. Experimental results and comparisons between the different techniques are provided. The experimental results have also been compared with results obtained by computer simulation.