An efficient ir approach based semantic segmentation

Content Based Image Retrieval (CBIR) is the task of finding similar images from a query one. The state of the art mentions two main methods to solve the retrieval problem: (1) Methods dependent on visual description, for example, bag of visual words model (BoVW), Vector of Locally Aggregated Descriptors (VLAD) (2) Methods dependent on deep learning approaches in particular convolutional neural networks (CNN). In this article, we attempt to improve the CBIR algorithms with the proposition of two image signatures based on deep learning. In the first, we build a fast binary signature by utilizing a CNN based semantic segmentation. In the second, we combine the visual information with the semantic information to get a discriminative image signature denoted semantic bag of visual phrase. We study the performance of the proposed approach on six different public datasets: Wang, Corel 10k, GHIM-10K, MSRC-V1,MSRC-V2, Linnaeus. We significantly improve the mean of average precision scores (MAP) between 10% and 25% on almost all the datasets compared to state-of-the-art methods. Several experiments achieved on public datasets show that our proposal leads to increase the CBIR accuracy.