Research On Sketch-based Image Retrieval Via Deep Supervised Hashing

With the rapid development of the Internet,image has become an important intermediary for people's daily communication.Image retrieval related topic has attracted much attention in research.Image can represent more accurate semantic contents than text.Besides,the appearance of touch-screen mobile devices nowadays brings greater convenience of free-hand sketch drawing.Therefore,sketch-based image retrieval(SBIR)has caused serious concern in recent years.SBIR can solve the situation when people want to convey query information without reference image.Due to the ambiguity and sparsity of sketches,as well as the limit of cross-domain retrieval,SBIR is more challenging to cope with than conventional content-based problem.Existing approaches usually adopt high-dimensional features which require highcomputational cost.Furthermore,they often use edge detection and parameter-sharing networks which may lose important information in training.In this study,we propose an efficient deep hashing based algorithm well suitable for large-scale image retrieval,getting high accuracy,fast retrieval time and low computation cost.Our contributions include:(1)We propose a point-wise binary codes deep learning strategy using prototype hash codes.By leveraging high semantic prototype hash codes,we embed different domains input(sketch and photo)into a common comparable feature space.(2)We present an efficient system with two separate networks specific to sketches and real photos.Thus,we can learn very compact features in the shared Hamming space,and solve the cross-domain challenge.(3)We design an algorithm for semantic extension retrieval situation when one image contains more than one objects and complex semantic relationship between them.We propose our corresponding prototype codes and network architecture for it.On our self made semantic extension dataset,our semantic extension model performs a good result.(4)We conduct our method on TU-Berlin Extension and Sketchy Extension,which are the largest dataset for SBIR.By fair comparison with several baseline methods,our method achieves state-of-the-art results in accuracy,retrieval time and memory cost on two standard large-scale datasets.Especially,we improve mAP by about 3% on TU-Berlin Extension and 4% on Sketchy Extension.