A robust image recognition algorithm has been developed aiming at direct implementation in a bio-inspired hardware accelerator chip. The characteristic features in an original gray-scale image of 64×64-pels (i.e., a 4096-dimension vector) are extracted by a newly-developed Principal Axes Projection (PAP) method and compressed to form a 64-dimension characteristic vector. Despite the large dimensionality reduction, the essential features in the original image are well retained in the vector representation. As a result, very robust image recognition has become possible by using a simple template matching technique. Although the matching with a large number of templates is computationally very expensive, we rely upon the already-developed vector matching LSI's featuring about 1000 GOPS performance for template matching [1-3]. The present algorithm has been successfully applied to medical radiograph analysis and handwriting pattern recognition, and its robust nature in recognition tasks has been proven by intensive computer simulation.