Hierarchical Object Recognition Model of Increased Invariance

Abstract

The object recognition model described in this paper enhances the performance of recent pioneering attempts that simulate the primary visual cortex operations. Images are transformed into the log-polar space in order to achieve rotation invariance, resembling the receptive fields (RF) of retinal cells. Via the L*a*b colour-opponent space and log-Gabor filters, colour and shape features are processed in a manner similar to V1 cortical cells. Visual attention is employed to isolate an object's regions of interest (ROI) and through hierarchicallayers visual information is reduced to vector sequences learned by a classifier. Template matching is performed with the normalised cross-correlation coefficient and results are obtained from the frequently used Support Vector Machine (SVM) and a Spectral Regression Discriminant Analysis (SRDA) classifier. Experiments on five different datasets demonstrate that the proposed model has an improved recognition rate and robust rotation invariance with low standard deviation values across the rotation angles examined. © Springer-Verlag Berlin Heidelberg 2013.