Deep Learning Class Discrimination Based on Prior Probability for Human Activity Recognition

Abstract

Convolutional neural network (CNN) depicts the transformation of the input image through a series of convolutions and other non-linear phases for recognition and classification purposes. This method has gained popularity for its significant contributions to computer vision applications and the improvement of the state-of-the-art. With CNNs, the softmax loss is used as the traditional loss function. This loss function allows deep features of distinct classes to be separated and promote the effective training of deep neural networks. An improvement on CNNs' discriminative power for face recognition was recently reported, where softmax and center loss were jointly used as supervisory a loss signal. In this paper, it is shown that such a supervisory loss function is not optimal in human activity recognition, and hence a new likelihood regularization term aimed at improving the feature discriminative power of the CNN models. This regularization term is modeled from Bayesian distribution for the posterior estimation of class probability density. The regularization term is shown to improve different class discrimination, and it is capable of maximizing the distance between different classes and minimizing distances within the same class in human activity recognition. The results obtained on the KTH and Weizmann datasets were encouraging.