Automatic Image Annotation using Semantically Modified HITS on ConceptNet
Madhat Alsoos · International journal of imaging and robotics · 2019
Image annotation is now a requirement for many computer applications in various fields. Many researches in image annotation field try only to classify scene images into limited pre-known classes. In this paper, we provide a novel system to automatically annotate images with semantically related labels from ConceptNet common sense knowledge base concepts. We represent each scene image by a bag of visual words to find relevant labels from a manually annotated image dataset. Then we apply our SMHITS algorithm on these labels to extract semantically associated high-level labels from ConceptNet concepts. We test our proposed system on more than 105 images from Corel5k, SUN and LabelME datasets. Results compare our system to current state of the art annotation methods and show its superiority in extracting high-level labels even when image dataset has only visual objects as labels. Results also show that our system performance is remarkably increased when image dataset size is increased.