Image Processing: Object Recognition

Abstract

HMAX is a powerful computational model of object recognition introduced by Riesenhuber and Poggio (Nat Neurosci (2):1019–1025, 1999) which attempts to follow the rapid object recognition as performed by the human brain. Hierarchical approaches to generic object recognition have become increasingly popular over the years. As advocated by Serre et al. (Proceedings of CVPR, IEEE Computer Society, San Diego, pp. 994–1000, 2005) and Mutch and Lowe (Int J Comput Vision 80(1):45–57, 2008), hierarchical approaches have been shown to consistently outperform flat single-template (holistic) object recognition systems on a variety of object recognition task. Recognition typically involves the computation of a set of target features at one step, and their combination in the next step. A combination of target features at one step is called a layer, and can be modeled by a 3D array of units which collectively represent the activity of set of features at a given location in a 2D input grid. In general, (some of the) layers are computationally intensive as they involve the application of complex 2D filters on large 2D images. Nonetheless, they also provide important opportunities for parallelization, e.g., both at feature and at layer level. This chapter provides a brief introduction to the HMAX algorithm and surveys the concrete results obtained using different flows developed in the SMECY project and targeting the STHORM platform.