Cloud based content classification with global-connected net (GC-Net)

Abstract

Real-time classification of video content has been envisioned to revolutionize human lives. The paper introduces a cloud-based video classification system that is able to perform lightweight video classification on real-time captured video content. An Internet of things (IoT) device, called Content Classification Box (CCB), is defined as an add-on to one or a number of cameras in vicinity for content classification. The CCB will communicate with the cloud server once any interested content/event (such as abnormality) is identified, in which the corresponding video content is transported to the cloud server for further inspection. To achieve the lightweight and intelligent video content classification at the CCB, a novel convolutional neural network (CNN) framework, namely Global-Connected Net (GC-Net), is introduced. GC-Net is featured by a novel deep learning architecture for exploitation of all the earlier hidden layer neurons, as well as an activation function that has the potential to approximate complexity functions. We will show that the proposed CNN framework can achieve similar performance in a number of object recognition benchmark tasks, namely MNIST and CIFAR-10/100, under significantly less number of parameters, thus being able to apply to low-computation and low-memory scenarios.