Depth from defocusing using a neural network

This paper presents a technique for obtaining the distance of a step edge from the lens of a camera using a single defocused image of that edge. The proposed technique does not suffer from the correspondence and occlusion problems associated with methods based on multiple images. The technique employs a Multi-Layer Perceptron network trained by backpropagation to compute distances from derivative images of blurred edges. The paper gives experimental results which clearly show the accuracy of the proposed technique.