Image Feature Representation And Application Based On Generative Adversarial Networks

Content-based image classification,retrieval,description,and other applications depend on the feature representation of the image.In this paper,we solve image feature extraction and hash representation by using SGANs(Stacked Generative Adversarial Networks)model in order to improve the retrieval accuracy and time efficiency of image retrieval applications:extracting the semantic features of image hierarchy and improving the accuracy image retrieval applications;the semantic hashing of the image is obtained by discretization,thereby improving the time efficiency of the image retrieval application.On the basis of unsupervised SGANs,some semantic annotation information is introduced to guide the hash semantic features of similar images to have better clustering characteristics and further improve the retrieval accuracy.The specific work is summarized as follows:(1)Unsupervised hierarchical structure semantic feature representation learning: using SGANs to decompose image feature representation learning into a layer-by-layer stacked generation confrontation model,gradually extracting images from low-level contour features to high-level abstract semantic features and through hashing techniques get a low-dimensional hierarchical hash feature vector,increase the chance of gathering images of the same type,and improve accuracy.(2)Semi-supervised depth semantic hash features represent learning: image annotation information not only provides image category information but also implies similarity between images.On the basis of unsupervised SGANs,some annotation semantic information is introduced.Using the relative similarity between images,the depth hash optimization target for image retrieval is designed,and the hash semantic features of similar images are guided to have better aggregation characteristics and the accuracy of the search further improved.(3)Experimental verification and retrieval application system prototype realization:validity of model validation and retrieval accuracy on standard data sets and large-scale e-commerce commodity images.Experiments show that the hash feature of the image obtained by the stacked model against the network has good clustering characteristics.The semi-supervised SGANs deep hashing method has significantly better retrieval performance than the unsupervised SGANs method.Compared to the manual feature-based hashing method and supervised deep learning hashing methods,image retrieval performance based on semi-supervised SGANs deep hashing method has also been significantly improved.Stacked models are deployed in the retrieval application to implement the system prototype of the retrieval application.