Deep Adversarial Network For Sketch Retrieval

Content-based image retrieval(CBIR)is the current hot research topic in the field of computer vision.Sketch-based image retrieval is a branch of CBIR,which uses free-hand sketches as the query data.With the development of touch-screen devices,the demand of sketch retrieval will also rapidly increase.Free-hand sketches are mainly composed of abstract lines,and also contain geometric distortion.These make them quite different from inage data.Therefore,sketch retrieval is a specific cross-domain retrieval task.The core of sketch retrieval is to learn a comman feature subspace,where the features of sketches and images can be both discriminative and domain-invariant.In this paper,we present a deep adversarial network for sketch retrieval.Adversarial training is implemented as an interplay between feature extractor and domain classifier.We optimize domain classifier to distinguish the features of sketches and images,Besides,we optimize feature extractor to generate domain invariant features in order to confuse the domain classifier.With adversarial training,we can promote the feature fusion of sketches and images,and make network learn a more effective common subspace.We further impose similarity constraints on the feature extractor,in order to minimize the feature gap of sketches and images with same semantic labels,while maxmizing the distances among different semantic labels.Through the similarity constraints,the underlying semantic structure of data is better preserved when this data is projected into the common subspace.Experimental results on Fickr15 k show that our method can improve the precision and recall of retrieval.From the visual point of view,it also satisfies users' retrieval needs.