Electronics Project:Depth from a defocused image

Implementation of Depth from a single defocused image:

Estimation of depth using defocused images from multiple cameras is a much researched topic. But single defocused images have so far not been successfully used for the same. Some attempts have been made on different approaches, specifically using low level cues in order to obtain information about relative depth. Such depth maps are referred to as 2.5 D maps because they are not completely successful in providing information about the 3rd dimension. Previous work on estimation of relative depth from single images involve the use of projective geometry based techniques, learning appearance based models, statistics of wavelet co-efficient etc. The approach used here is different.

The project estimates the relative depth in the image directly by estimating the blur in various parts of the pictures. A poor assumption made in this method is that the focused part of the image is closer to the imaging system and the defocused (or blurred) part is farther away. Under this assumption, objects closer by are identified by their fine resolution and details and those farther away are characterized by their blurry nature. The reverse heat equation is used to measure the amount of blur at different regions of the image. A good amount of literature is available on the use of stabilized reverse heat equation for restoration purposes. Using the reverse heat equation, the image is restored in a spatially variant manner. The regions which are restored sooner will be finer resolution regions, assume to be closer to the imaging system, whereas those regions which take longer time to be restored will be the blurred regions.

The assumption made regarding the correlation between the blur of a region in the image and the depth of that particular region is certainly flawed. In many cases it may be possible for nearby objects to be blurred and objects farther away to be focused. Yet, in cases where accuracy may not be the prime goal, this method may yield invaluable information to computer vision scientists as supplementary information to be used to identify, classify and understand the interaction of various objects in a scene.