Concept-Aware Web Image Compression Based on Crowdsourced Salient Object Detection

Abstract

Reduced output quality and being unaware of content are among major issues with traditional image compression techniques. Such issues cause some critical problems when it comes to quality-intensive applications, including object/face detection and recognition, Web-based image viewers and management systems, etc. On the other side, efficiency of Web-based image search engines and retrieval systems in terms of user experience and usability could be affected. In order to cope with these challenges, a novel image compression method is proposed that takes advantages of collective human cognitive intelligence to detect the salient object(s) based on the recognized key concept(s). Then, other less-important regions/objects will be subject to the safe compression. Such an approach, besides preserving semantic aspects of the images that will result in smart (concept-aware) compression, could provide some crowdsourced labels for more efficient indexing and annotating of images. In this regard, two birds could be beaten with one stone: compressing Web images with respect to their content/concept and annotating them with crowd-suggested labels. The experimental results as well as user acceptance evaluation proved the efficacy of the introduced method.