Research On Sketch-based Image Retrieval

With the development of Internet technology and the popularity of multimedia devices,the number of digital images has increased dramatically.Effective retrieval technique for digital image has become an urgent need.Most traditional image retrieval techniques are text-based image retrieval.However,many images could not be accurately described by words.Therefore,text-based image retrieval often fail to meet people's retrieval intensions.Hand-drawn sketch is an input modality that intuitively represents the user intent.In recent years,with the popularity of the touch screen,drawing sketches becomes more simple and convenient.Therefore,sketchbased image retrieval technology has become the current research focus.This paper introduces the background and history of image retrieval,and elaborates the retrieval process and key technologies of sketch retrieval.Through the deep research and analysis of the key technologies for sketch retrieval,this paper aims to improve the effectiveness of sketch feature.The training data,initialization,and network loss design are all considered on the research of sketch-based image retrieval algorithms.In this paper,an image aided sketch-based image retrieval method is proposed,which improves the effectiveness of the feature and the accuracy of matching.Taking into account the lack of sketch data,large-scale image data is used for pre-training.Then,the pre-trained network is used as the initialization network.In order to reduce the domain gap between images and sketches,a more effective contour extraction method is used to extract the sketch approximation.In addition,considering that the shape information is the key information of sketch-based image retrieval,a shape regression loss is introduced on the basis of the original classification loss.The two losses are jointly used to optimize the network.Experimental results on the Flickr15 k dataset show that the proposed method outperforms the state-of-the-art results.