“Artificial Neural Network Algorithms for Some Computer Vision Problems”

by Yi-Tong Zhou

June 1989

The problem considered here involves the use of an artificial neural network for solving some computer vision problems such as static and motion stereo, computation of optical flow, and image restoration. The network used in this research contains massive mutually interconnected and self-connected binary neurons. Two decision rules, deterministic and stochastic, are used. The stochastic decision rule guarantees convergence to a global minimum but computationally is very expensive, while the deterministic decision rule greatly reduces computing time, but only gives a local minimum.

Two basic methods, static and motion stereo, for extracting 3-D information from more than one image are considered. The static stereo method is based on images taken by two cameras separated by a known baseline. The motion stereo method infers depth information from a sequence of monocular images. The derivatives of the intensity function are used for matching. A window operator which functions very similar to the human eye in detecting the intensity changes is proposed for estimating the derivatives. Under the epipolar, photometric and smoothness constraints the neural network is employed for the matching procedure. For motion stereo, two algorithms, batch and recursive, which allow the use of arbitrarily many image frames are presented. No surface interpolation step is involved in the algorithms because of the dense derivatives used.

An algorithm using rotation invariant primitives extracted from successive monocular images is presented for computing optical flow. Under local rigidity assumption and a smoothness constraint, the neural network is used to compute optical flow. To locate motion discontinuities, the information about occluding elements is utilized by embedding it into the bias inputs of the network. A batch solution is also developed for the case of pure translation.

An approach for the restoration of grey level images degraded by a known shift invariant blur function and additive noise is developed. The neural network is employed to represent a possibly nonstationary image whose grey level function is the simple sum of the neuron state variables. The nonlinear restoration method is carried out iteratively by using a dynamic algorithm to minimize the energy function of the network. Owing to the model's fault-tolerant nature and computational capability, a high quality image is obtained using this approach . A practical algorithm with reduced computational complexity is also presented. The choice of the boundary values to reduce the ringing effect is discussed and comparisons to other restoration methods such as the SVD pseudoinverse filter, minimum mean square error (MMSE) filter and modified MMSE filter using Gaussian Markov random field model are given. A schematic diagram of optical implementation of the restoration algorithm is described

To demonstrate the efficacy of all these algorithms, experimental results using both synthetic and natural images are presented.