The Fine-Grained Retrieval Of Sketches Based On Deep Learning And Related Research

With the rapid development of Internet,the content of the Internet gradually changed from the original text information to various rich media data,the main one of which is image.However,with the rapid growth of image data,it becomes more difficult for the user to retrieve the image needed,and fine-grained retrieval is the inevitable trend of future image retrieval.In addition,the diversity of images has led multiple fields'application of cross-domain image retrieval,such as retrieving the street view image from aerial image,retrieving the pictures of the day from night scene image and so on.This paper focuses on the cross-domain fine-grained retrieval from hand-drawn sketches to images.So the core problem we need to solve is the three points:first is cross-domain retrieval.Sketches and pictures has obvious differences in the features,the features of the sketch and image belong to different feature space,the paper needs to map different domains of image to the same feature space;Secondly,to achieve fine-grained retrieval.That is,the retrieval system has very sensitive recognition effects on small differences;Finally,the efficiency of the sketch retrieval.The large-scale development of the data,the image retrieval faces the problems of storage space and calculation speed.This paper studies the above problems gradually.By studying the relevant papers and a large number of experiments,we design and build a kind of heterogeneous network,with two different network branch structure for adapting to feature extraction of sketch domain and image domain.Because there is no data set to train alone the whole deep network model,so we took a step by step training way.On the training set,we train gradually from coarse-grained to fine-grained,using contrast loss and triplet ranking loss to limit gradually two domain feature similarity.finally we got very good effect on fine-grained retrieval,and achieve end-to-end image retrieval.After the feature extraction of sketches and images,we continue to study the efficiency of sketch retrieval in large-scale data.Based on the existing heterogeneous network model,we designed a deep hash model based on heterogeneous network.In the last layer of the heterogeneous network,the hash map layer is designed and the discrete binary hash codes are mapped.In order to solve the unconductance of binary hash code similarity constraints,we have studied a lot of papers and methods,and finally completed the derivability of the loss function through relaxation.In the end,we have achieved great improvement in retrieval accuracy and efficiency.