A robust scene descriptor based on largest singular values for cortex-like mechanisms

Analysis and recognition of images observed in different situations are the most important tasks of the visual cortex. Although many studies have been done in the field of computer vision and neuroscience, the underlying processes in the visual cortex is not completely understood. An inspired model of the visual cortex that lately has gained attention for object recognition is HMAX, which describes a feed-forward hierarchical structure. This model shows a degree of scale and translation invariance. Other capabilities of the visual cortex are not that much sensitive against rotation and noise in color space. In this paper, we introduce a novel method to increase the degree of robustness against noise and rotation. We describe a hierarchical system that closely follows the organization of visual cortex and builds an increasingly complex and invariant feature representation with R, G, B and gray channels as inputs. Similar to the recently published methods, the phase of learning is done only in the S2 layer of HMAX structure. While in the proposed model instead of using directly the distance between the patches and the C1 units, the distance between largest singular values of the patches and the C1 units are used. These values behave indeed insensitive significantly with respect to rotation. To this end, we present the experimental results of the COREL datasets and shows that the proposed model have better performance than the previous HMAX models in complex visual scenes.