An elastic net-regularized HMAX model of visual processing

Abstract

The hierarchical MAX (HMAX) model of human visual system has been used in robotics and autonomous systems widely. However, there is still a stark gap between human and robotic vision in observing the environment and intelligently categorizing the objects. Therefore, improving models such as the HMAX is still topical. In this work, in order to enhance the performance of HMAX in an object recognition task, we augmented it using an elastic net-regularised dictionary learning approach. We used the notion of sparse coding in the S layers of the HMAX model to extract mid- and high-level, i.e. abstract, features from input images. In addition, we used spatial pyramid pooling (SPP) at the output of higher layers to create a fixed feature vectors before feeding them into a softmax classifier. In our model, the sparse coefficients calculated by the elastic net-regularised dictionary learning algorithm were used to train and test the model. With this setup, we achieved a classification accuracy of 82.6387%∓3.7183% averaged across 5-folds which is significantly better than that achieved with the original HMAX.