Training of Neural Network-Based Cascade Classifiers

Abstract

Abstract—Deep Artificial Neural Networks (ANNs) achieve state-of-art performance in many computer vision tasks; however, their applicability in the industry is significantly hindered by their high computational complexity. In this paper we propose a model of ANN classifier with cascade architecture, which allows to lower the average computational complexity of the system by classifying simple input samples without performing full volume of calculations. We propose a method for joint optimization of all ANNs of the cascade. We introduce joint loss function that contains a term responsible for the complexity of the model and allows to control the ratio of the precision and speed of the resulting system. We train the model on CIFAR-10 dataset with the proposed method and show that the resulting model is a Pareto improvement (regarding to speed and precision) compared to the model trained in a traditional way.