Research On Technologies Of Image Retrieval Based On Hybrid Features

With the development of Internet technology,traditional text search can no longer meet the retrieval needs brought about by the growth of information,and the research on high-performance image retrieval is of great significance.For the current main research methods,the features extracted by traditional manual models represented by compact descriptor for visual search have good resistance to anti-invariance,and the features extracted by deep models represented by neural networks have strong representation.Since the anti-invariance and high representation of image features can form a good complement,better retrieval performance can be achieved.Therefore,this thesis studies the high-performance image retrieval technology based on hybrid features.However,the compact descriptor model is a fixed bit rate allocation mode,which brings a lot of unnecessary space waste;the deep model does not fully utilize the potential information of the training samples.In addition,the existing feature fusion technology simply combines the scores of the two and does not take advantage of the potential complementarity between the two.Therefore,this thesis constructs a high-performance manual feature extraction algorithm based on dynamic bit rate allocation;a high-performance deep feature extraction algorithm based on hard sample generation;a highperformance retrieval algorithm based on hybrid features.Specifically,this thesis first proposes a dynamic bitrate allocation scheme based on the compact descriptor model.Through the analysis of the difficulty of retrieval of the images in the database,more appropriate bitrates are assigned to pictures of different complexity,so as to achieve high performance.In addition,based on deep image retrieval technology,this thesis designs a two-stage hard sample generation framework,which uses an adversarial network to independently generate hard positive and negative samples for simple training samples.In the model training stage,this thesis proposes an end-to-end feature extraction architecture that uses generated features and original features to jointly train the network architecture to strengthen the use of potential information from simple samples.Since the above-mentioned research cannot meet the requirements for image resistance to invariance and high characterization at the same time,this thesis innovatively combines traditional and deep features.When training the deep model,this thesis proposes a hard sample selection scheme based on manual features to achieve more effective training sample selection.In this thesis,we use local descriptors of manual features to delete irrelevant targets from the top position of the search results to further improve the search results.This thesis starts with the mining of hard sample sets,the fusion of retrieval results,and the reordering of retrieval results.The hybrid of the two is analyzed and practiced theoretically from various angles.Sufficient experimental results show that the fusion of the two has a significant improvement in image retrieval performance for a single method.