This research presents biologically inspired visual models and algorithms for automatic part extraction and object recognition from gray-scale images with approximate rotational invariance. Motivated by the efficiency of the primate visual system, the models employ sparse representation and unsupervised learning to achieve fault tolerance, low power consumption, and adaptive feature discovery without prior knowledge. A hierarchical architecture inspired by the primate ventral visual pathway was developed, including models of V1 and V2 for low-level feature extraction, V4 for parts-based shape representation, and IT for high-level object recognition. The models demonstrate efficient, transformation-tolerant visual processing, with V4 units exhibiting biologically consistent curvature and object-centered tuning. Using flexible constellations of rigid parts, the IT model achieves robust object recognition across viewpoints. Overall, this work integrates sparse coding and unsupervised learning into a biologically plausible framework with strong performance in visual recognition tasks.
