Improving invariance in visual classification with biologically inspired mechanism

摘要

A computational model of visual cortex has raised great interest in developing algorithms mimicking human visual systems. The max-operation is employed in the model to emulate the scale and position invariant responses of the visual cells. We further extend this idea to enhance the tolerance of visual classification against the general intra-class variability. A general architecture of the basic block constituting the model is first presented. The architecture adaptively chooses the best matching template from a set of competing templates to predict the label of the incoming sample. To optimize the non-convex and non-smooth objective function resulted, we develop an algorithm to train each template alternately. Experiments show that the proposed method significantly outperforms linear classifiers as a template matching method in several image classification tasks, and is much more computationally efficient than other commonly used non-linear classifiers. In the image classification task on the Caltech 101 database, the performance of the biologically inspired model is obviously boosted by incorporating the proposed method.