An efficient image aesthetic analysis system using Hadoop

Abstract

Assessing aesthetic appeal of images is a highly subjective task and has attracted a lot of research interests recently. Prior researchers have developed several aesthetic analysis systems on standalone computers. However, it is challenging to run the algorithms on mobile devices since the process of aesthetic analysis is quite complicated and time-consuming, especially for large amount of images. Hadoop is a popular technology for big data processing on cloud to offload computing burden from terminals. However it has NOT been used on image aesthetic yet. In this paper, we present an image aesthetic analysis system based on Hadoop framework to provide an efficiency solution and better user experience. We address several major problems: (1) adapt MapReduce for image data format and aesthetic analysis algorithms; (2) improve computing performance for large amount of small image files; (3) design a dynamic scheduling mechanism to optimize concurrent multiple users׳ requests; (4) design an effective commutation service between cloud and terminals. Experimental results demonstrate significant performance improvements with our system. At the same time, the system efficiency increases linearly with the expansion of the slaves in Hadoop.