Fast identification of images using neural networks

摘要

A hierarchical structure of neural networks has been constructed and utilized for fast identification of images. Applications include recognition and extraction of partially obscured objects. The networks are self-organizing. The selection and registration of necessary templates are automated at all levels. Training of the networks takes a minimal amount of time. The development is mostly carried out on a Sun-3 workstation, VAX and IRIS 4D/240 computers with a single processor. However, this algorithm is more effective if implemented in parallel computers. A preliminary parallel implementation study with an AMT DAP-610 systolic array computer was conducted and favorable results obtained. The experimentation of the algorithm with various image data with or without partial occlusion shows robustness of the networks. Further applications of the networks are in progress.