Algorithmic optimizations in the HMAX model targeted for efficient object recognition

Abstract

In this paper, we propose various approximations aimed at increasing the accuracy of the S1, C1 and S2 layers of the original Gray HMAX model of the visual cortex. At layer S1, an image is convolved with 64 separable gabor filters in the spatial domain after removing some irrelevant information such as illumination and expression variations. At layer C1, some of the minimum scales values are exploited in addition to the maximum ones in order to increase the model’s accuracy. By applying the embedding space in the additive domain, the advantage of some of the minimum scales values is taken by embedding them into their corresponding maximum ones based on a weight value between 0 and 1. At layer S2, we apply clustering, which is considered one the most interesting research areas in the field of data mining, in order to enhance the manner by which all the prototypes are selected during the feature learning stage. This is achieved by using the Partitioning Around Medoid (PAM) clustering algorithm. The impact of these approximations in terms of accuracy and computational complexity was evaluated on the Caltech101 dataset containing a total of 9,145 images split between 101 distinct object categories in addition to a background category, and compared with the baseline performance using support vector machine (SVM) and nearest neighbor (NN) classifiers. The results show that our model provides significant improvement in accuracy at the S1 layer by more than 10% where the computational complexity is also reduced. The accuracy is slightly increased for both approximations at the C1 and S2 layers.