Barlow (1961) proposed that the goal of sensory coding is to transform the input signals such that it reduces the redundancy between the inputs. Atick (1992) and Atick and Redlich (1993) have used correlation-based methods suggesting that the principle of redundancy reduction may be applied towards the understanding of coding principles in retinal cells in the visual cortex (Field, 1994). This strategy may be used for the purpose of efficient coding for natural images since they are not purely random but contain structure. Natural images have oriented lines, edges and other structures that have dependencies of higher-order statistics. These localized structures can be described mainly by the phase spectrum. For example, an edge occurs locally and has its phase spectrum aligned across different spatial frequencies. Correlation-based approaches are phase-blind, i.e. they can capture only the power spectrum and higher-order methods are needed to account for localized oriented structures. Olshausen and Field (1996) considered a network that maximizes the sparseness of the representation of natural images and showed that the extracted features are localized and oriented. Those features are similar to receptive fields in the primate striate cortex.
KeywordsFace Recognition Receptive Field Natural Image Natural Scene Phase Spectrum
Unable to display preview. Download preview PDF.
- 1.Gabor wavelets are of similar shape as the receptive fields of simple cells in the primary visual cortex (V1). They are localized in both space and frequency domains and have the shape of plane waves restricted by a Gaussian envelope function.Google Scholar
- 2.The images are available in (ftp.cnl.salk.edu/pub/tony/VRimages).Google Scholar