Hierarchical model for object recognition based on natural-stimuli adapted filters

We examine a set of biologically inspired features and apply it to the multiclass object recognition problem. To obtain these features we modify HMAX, which is based on a hierarchical model of visual cortex. Instead of using a set of standard Gabor filters we use a set of natural-stimuli adapted filters. These filters emerge as a result of optimization based in part on smooth L1-norm based sparseness maximization. These filters in conjunction with the HMAX model give more biological plausibility to the model. The features thus obtained are largely scale, translation and rotation invariant, and are fed to a support vector machine classifier. We successfully demonstrate the applicability of this modified HMAX model to the object recognition task by testing it on Caltech-5 and Caltech-101 datasets.