PERIA-Framework: A Prediction Extension Revision Image Annotation Framework

Abstract

Image annotation is a procedure to interpret the semantic concepts associated with image objects and represent them as their textual descriptions. Automatic and manual techniques have been extensively discussed in recent years to annotate the image objects, but are not without limitations. Automatic image annotation techniques mainly consider a single classifier and a descriptor type to annotate the image objects. Furthermore, thesaurus based extensions and human-centered revisions of the annotations are usually not possible. The fine-tuning of classifiers is generally not supported. In contrast to this, manual image annotation improves the accuracy, but tedious to annotate huge collections of image objects. Alternatively, semi-automatic image annotation techniques are human-centered, enhances the efficiency, and also speed-up the annotation process by machine intervention. In this research, a semi-automatic image annotation framework is proposed to address limitations in automatic and manual image annotation techniques. Our image annotation framework considers multiple descriptors and artificial neural networks to annotate the image objects. Along with that, a voting mechanism is provided to recommend the suitable annotations extendible by thesaurus and human revisions. Revised and extended annotations employed further to fine-tune the classifiers. Image annotation framework is instantiated and tested on a real dataset by implementing an image annotation tool.