A Semantic Image Retrieval Technique Through Concept Co-occurrence Based Database Organization and DeepLab Segmentation

Abstract

In this paper, a semantic image retrieval technique that efficiently depicts users' perspective is proposed. It primarily aims in the representation of contextual diversity of the user through a high level semantic segmentation technique called DeepLab-V3+. An online user interactive step is also included during the retrieval process. The significance of intra-concept variation in image retrieval is clearly presented in this paper. An efficient database organization, which forms the essence of the retrieval methodology, based on concept co-occurrence and inter-concept distance is also proposed. ResNet-101 CNN features extracted from the regions are utilized in classification and retrieval tasks. The simulation results and performance analysis conducted on PASCAL VOC2012 and SUN '09 datasets depict the superiority of the proposed technique over other approaches.