Object´s Depth Ordering in Monocular Image by Using Multi-Neural Network Classification

The problem of finding the orders of objects in a given monocular image is studied. The study is conducted under the constraints that no prior knowledge on the imaging device specification and focus depth are provided. Most of the work on depth ordering of the sub-areas in an image is focused on binocular vision (stereopsis). The depth can be easily estimated from the overlapping distance of the binocular images. In case of monocular image, the depth ordering is more complex than the binocular case. A supervised neural network is applied to learn focused and blurred images by using Fourier coefficient of each image as the input to multi-neural network classifier. The degree of blurring as well as the location of each sub-area in the image is used to determine the order of depth by a set of conditional rules. Both real and synthetic images are experimented