Research On Image Retrieval Algorithm Based On Feature Fusion

With the rapid development of the Internet and the upgrading of mobile devices,a large amount of picture data and life have become closely connected.How to accurately and efficiently search in massive pictures has always been the focus and difficulty in the field of image retrieval.In the early days,image retrieval mainly added tags to images by manual means,and realized image matching based on text.With the surge in the number of images,the increase in labor costs and the decrease in matching accuracy,text-based image retrieval cannot meet the needs.In the 1990 s,the visual information of images could be extracted.Researchers used the visual features of images to achieve better results in image retrieval.Content-based image retrieval technology gradually occupied the mainstream.However,visual information is a low-level feature of the image,and the semantic information of the image cannot be understood,and there are still many shortcomings.In recent years,deep learning technologies have developed rapidly,especially convolutional neural networks,which can extract high-level semantic features of images,and have shown better performance than traditional methods in the field of image retrieval.However,whether it is low-level features or high-level semantic features,the expressive power of a single feature is not comprehensive enough.Using multiple features to fuse and describe the image content from different angles through complementary features can overcome this problem.Therefore,this paper proposes a method of image retrieval combining multiple features to improve the accuracy of image retrieval.The main research contents of this article are as follows:(1)Aiming at the problem of insufficient expression of the image content by a single feature,this paper proposes a fusion feature combining the features of the convolutional neural network and the SIFT feature.The RMAC feature extracted based on VGG16 is selected as the high-level semantic feature,and the VLAD feature based on the SIFT descriptor is selected as low-level features.The fused feature contains both the low-level information of the image and the high-level semantic information of the image,which can express the content of the image more comprehensively.At the same time,two schemes of feature layer fusion and score layer fusion are designed,and the weights in the fusion scheme are determined through experiments.Finally,the effect of fusion features was verified on three datasets: Holidays,Oxfords5 k and Paris6 k.Experiments show that compared with a single feature,the fusion feature can achieve better results,proving that the two features are complementary.The improvement of the scoring layer fusion scheme is more obvious.For these two features,the scoring layer fusion is more suitable and effective than the feature layer fusion.(2)In order to give full play to the role of features,an adaptive weighted multifeature fusion method based on the change rate of the distance score is proposed.The same feature has different effects on different images,and the size of the weight occupied by different image features during feature fusion should also change.Based on the scoring layer fusion scheme,the scores between the images are calculated based on the Euclidean distance,the distribution characteristics of the scores in the neighborhood space of the query picture under a certain feature are analyzed,and the change rate of the distance score of the neighborhood space under the certain feature is obtained to determine weights.Experiments show that compared with the fixed weight method,the weighting method based on the change rate of the distance score can achieve better results and fully play the role of different features.