PF-Face: A Parallel Framework for Face Classification and Search from Massive Videos Based on Spark

Abstract

With the development of massive amounts of video data with human-centric, how to efficiently recognize and search these human face from big video data becomes a bottleneck in current video processing. Distributed computing platforms and GPU-based acceleration provide mainstream solutions for this challenge. However, integrated parallel framework and algorithms for massive face analysis are lacking in the current literature, to our best knowledge. To offer a complete and efficient solution for face big data analysis, this paper presents a parallel framework based on Spark. First, it is based on the Hadoop and Spark cloud computing platform to paralleled process enormous video data distributed storage. Specially, parallel deep learning model is supported by the CaffeOnSpark in our framework. Second, we implement parallel algorithms for face classification and search according to the hierarchical feature of video. These algorithms are based on Spark to achieve quickly analysis for large number of videos. At last, performance evaluation experiments of face classification and character search on a cloud cluster demonstrate the efficiency of approach to analyze massive video data.