Role of Complex Cells in Invariant Object Recognition

A major puzzle about visual system is how to deal with image variations in terms of lighting, deformation and others. It is important to have a code that is robust with respect to such changes and yet can distinguish between different shapes. We present here arguments that show that the complex cells of the visual cortex are precisely that. There is increasing evidence for an important role of complex cells in object representation and recognition. On the other hand, complex cells are insensitive to spatial phase information, and thus can lead to ambiguities with respect to position, contrast polarity and mirror symmetry, aspects which appear to be important in characterizing natural images. We investigate the richness of information in such a neural coding scheme, by using Gabor magnitudes as a model of complex cells, through a mathematical treatment of the ambiguities of such filter responses, and by empirical examination of alternate reconstructions of images from their Gabor magnitudes. The results suggest that while individual complex cell responses are ambiguous with respect to exact position, contrast polarity and mirror symmetry, population responses contain sufficient information to capture the perceptual essence of images and for inducing recognition. A complex cell type representation seems to be not only sufficiently discriminating for shape perception, but also superior to a ‘phase sensitive code;’ its inherent ambiguities providing robustness to changes in background, lighting, and distortion.