| Sketch Based Image Retrieval(short for SBIR)is to retrieve images through hand-drawn line drawings.Compared with other image retrieval methods(based on labels and images),it is more convenient and efficient.And with the popularity of touch screen devices,the application scenarios of SBIR are increasingly widespread.However,there are still many difficulties in the field of SBIR,among which cross-domain retrieval is the central issue.Image is made up of pixels,while sketch is made up of lines.There is a huge difference between the two domains,which makes SBIR more challenging.The most common solution to the problem is to transform the image into edge,so as to narrow the difference from the sketch.However,the quality of the edge generated by the existing edge extraction algorithm is poor,and it is easy to lose important information in the image,which greatly affects the accuracy of retrieval.In terms of this issue,Pseudo-supervised Multi-scale Edge Extraction algorithm(short for PS-MSEE)is proposed.In this paper,the pseudo supervision strategy of unsupervised object location is introduced into the edge detection algorithm,thus the unsupervised edge extraction based on deep learning algorithm is realized for the first time,and the edge with better quality can be learned.The network adds skip connections on the convolution module of VGG-16 to realize multi-scale fusion of feature map,and then uses bottom-up feature concatenating and multi-scale feature supervision to realize multi-scale supervision of features.In order to further reduce the missing edge,this paper creatively proposes a multi-level feature enhancement module based on color scale.The multi-level enhanced image is obtained by enhancing the color scale contrast,and then the features extracted from the enhanced image are fused to better detect the weak contrast edge.The edge obtained by the PS-MSEE algorithm improves the retrieval accuracy of both the Triplet SN and DSSA algorithms by 2% to 5% on Qmul-shoe and Qmul-chair datasets.The experimental results show that the edge obtained by PS-MSEE algorithm can better retain the key information of the image,and finally achieve the purpose of further improving the retrieval accuracy of SBIR on the basis of the existing retrieval algorithms.The main contents of this paper are as follows1.Investigate and analyze the existing problems in the field of SBIR,and systematically elaborate and summarize the theoretical basis and implementation methods in the field of SBIR.Research and analysis show that edge extraction is the best choice to solve the problem of differences between domains in SBIR.2.Investigate and analyze the existing edge extraction algorithms,compare the advantages and disadvantages of each algorithm,and analyze the reasons for the poor performance of edge extraction algorithms in SBIR.3.A pseudo supervised multi-scale edge extraction algorithm(PS-MSEE)is proposed to obtain edge images suitable for SBIR.4.Experiments are carried out on the commonly used SBIR datasets,and the retrieval accuracy is better than the original algorithm by using the edge generated by PS-MSEE.It is verified that the edge extracted by PS-MSEE is helpful to improve the retrieval accuracy. |
