Research And Implementation Of Image Retrieval Based On Multi-feature Fusion

The number of images on the Internet has reached a mass state.It is difficult to search out the desired images in so many images.Therefore,the technology of efficient and fast image search has been put on the agenda.Image search technology can be divided into two categories: text-oriented and content-oriented.The former requires experienced experts to describe images.For lots of images,it takes a lot of manpower and time,and different experts have different understanding of images,which leads to a lower accuracy of the retrieval.The latter is a search method that uses the content of the image itself,and uses its own visual information features to express the content of the image,which is more suitable for the query of a large number of images,which is also the main direction of the development of image search technology.In the early stage,the technology which used image content to search was mainly applied to image color,image texture,image shape and other underlying features.In recent years,with the popularity of deep learning,more and more researchers use the convolutional neural network(CNN)in deep learning to extract image features to conduct research on image search,and achieve good results.Therefore,the following points are mainly studies in this thesis:1.The traditional features and depth features are used for image search,and the two search results are compared and analyzed,and an image search method combining multiple different features is proposed.2.An adaptive weighted image search method based on multi-feature fusion is proposed,which can be applied to both supervised and unsupervised cases.3.An adaptive weighting strategy for multi-feature fusion is proposed in both supervised and unsupervised cases.An image search method based on multi-feature adaptive weighted fusion is designed considering the similarity measure fusion at measurement level.Through experiments,the mAPs of this method on the Holidays and Oxford5 k image sets are 88.6% and 58.44%,and the N-S score on the Ukbench image set is 3.906.This method achieves good retrieval accuracy and improves the retrieval performance,and proves the effectiveness of this method.