Salient feature-based object recognition in cortex-like machine vision

摘要

This preliminary study concentrates on the refinement of early models in cortex-like machine vision, in particular Feature Hierarchy Library (FHLib), in order to remedy some of its drawbacks and introduces a salient feature approach to generic object recognition. Enhancements in three areas, a) extraction of features from the most salient region of interest (ROI) and their rearrangement in a ranked manner, rather than random extraction over the whole image, b) exploitation of larger patches to improve spatial resolutions and performance, c) a more versatile template matching technique, have been the main contributions of the present work. Under different classification methods, the improved model is validated using 4 different types of datasets and shows better recognition accuracy over the original FHLib model. This improvement has remained consistent even after the reduction of the library of features by approximately 40%.