Conclusions and Future Research

Abstract

Several artificial neural networks were presented and applied to computer vision problems such as static and motion stereo, computation of optical flow, and image restoration. To ensure quick convergence of the networks, the deterministic decision rule was used in all the algorithms. Experimental results using natural images confirm that neural networks provide simple but very efficient means to solve computer vision problems, especially at the low-level. Experimental results also provide a strong support to the hypothesis that the first order derivatives of the intensity function and the Gabor features are appropriate measurement primitives for stereo matching, and the principal curvatures are useful for computing optical flow. The utilization of multiple frames for computing depth and optical flow gives much better results and is useful for real time applications. Since no matrices are inverted during restoration, the serious problem of ringing due to the ill conditioned blur matrix is avoided and hence the neural network algorithm gives high quality images compared to some of the existing methods. Although the artificial neural networks have been applied to only a few low-level computer vision problems so far, it is potentially useful for many computer vision problems.