Streamlined Feature Representation For Content-based Image Retrieval

Recently,with the vigorous growth of digital devices embedded with cameras and the rapid development of Internet technology,billions of people are projected to the sharing and browsing photos.The omnipresent access to digital photos and the Internet bright light on many rising applications based on image search.Content-based image retrieval(CBIR)has been a longstanding research topic in the computer vision society.In the early 1990 s,the study of CBIR truly started.Among all the research,the studies base on Scale-Invariant Feature Transform(SIFT)feature and Bag-of-Words(BoW)feature make great achievement and lay the foundation for the application of CBIR on massive data sets.Recently,with the prosper in Deep Neural Network(DNN),deep feature based on DNN models meet the success in many fields.Among them,the deep Convolutional Neural Network(CNN)has achieved most advanced performance in CBIR applications due to its powerful feature representation capabilities.Semantic gap is a fundamental challenge in CBIR.Semantic gap refers to the model's difficulty in describing highlevel semantic concepts with low-level visual features.In order to reduce the semantic gap,the academic community proposes the concept of metric learning(DML).DML refers to learning a measure of similarity model.In the feature space,the descriptor distance of semantically related objects should be small,and vice versa.Combining with the Siamese network and semantic-independent sample mining strategies,DML improves the network's semantic feature expression ability through loss functions,such as triplet loss and contrastive loss,thereby improving retrieval performance.The rank-based contrastive loss,proposed in this paper,extracts the global descriptors by fine-tuning the pre-trained model whose FC layers is removed and combined with a pooling layer at the bottom.Siamese network training structure is built by constructing a sample queue by semantically independent sample mining strategies.The CNN model is used to perform deep feature extraction on the samples,and the similarity ranking of the sample queue is performed based on the Learn to Rank(L2R)strategy,besides,the learning difficulty of the sample queue is evaluated.Then we propose dynamic margin in the contrastive loss according to the rank result above.And the model is updated by the proposed loss function.After the model is obtained,it is tested on the Oxford,Paris,and Holidays datasets.This paper proposes a sort-based adaptive contrast loss algorithm with better retrieval results.