Research On Large-Scale Image Retrieval Based On Adaptive Feature Fusion

Content-based image retrieval is a cross-cutting topic in the fields of Applied Mathematics,Computer Vision,etc.In recent years,although it has achieved rapid development,there is still room for improvement in retrieval accuracy and retrieval efficiency.Based on the research of existing methods,the paper mainly does some work as follows:Firstly,considering that the appropriate image descriptors can effectively improve the retrieval performance of the image retrieval system,the paper uses weighted tandem original image features and grayscale features to describe the image.At the same time,due to the weakness of the generalization performance of image retrieval system based on single feature,this paper combines multiple CNN features for the image retrieval.The experimental results show that the image description method of weighted tandem original image features and grayscale features has better retrieval performance than the original image feature description method.And the fusing deep features of multiple networks has achieved better retrieval performance than the single feature.Secondly,considering that the image retrieval system based on feature fusion is more computationally intensive and less efficient than the single feature-based image retrieval system,this paper proposes a hierarchical index structure based on K-means clustering.Specifically,when the users enter the query image,we firstly match the similar class,and then find the similar image in the matched similar class.Because of the similar images have high similarity,the matched similar images returned to the user also have high similarity with each other,which can effectively reduce the loss of retrieval accuracy while great improving retrieval efficiency.The experimental results show that,under the condition of a small loss of retrieval precision,the preprocessing of the dataset based on the hierarchical structure proposed in this paper can effectively improve the retrieval efficiency.Finally,in the image retrieval system based on feature fusion,distinguishing between good and bad features can effectively improve retrieval performance for different queries.In this paper,an adaptive weighting method is proposed for feature selection.Specifically,in each query,the weight of a single feature is dynamically updated with the query image,so that the retrieval system can make full use of the feature performance.Additionally,the proposed method can be used in both supervised and unsupervised image retrieval.The experimental results show that the performance of the proposed method is better than the related feedback method in supervised image retrieval.In unsupervised image retrieval,the performance is better than the global average weight.The proposed method is validated on four benchmark image retrieval datasets which are Corel-1k,Holidays,UC Merced Land Use,and RSSCN7.And the value of precision of Top 20 on Corel-lk,UC Merced Land Use,and RSSCN7 are 0.9709,0.8896,and 0.9442,respectively.The value of mAP on Holidays is 0.8845.