Near-similar Image Retrieval Based On Feature Fusion

The development of image processing technology makes image data editing,transmission,storage increasingly convenient and low cost.Faced with the existence of a large number of near-duplicate images in the network,how to filter and screen these images has become a practical research topic.The traditional content-based image retrieval algorithm(CBIR)has achieved certain results but still has some problems.In recent years,many near-duplicate image retrieval algorithms have been proposed successively to improve retrieval accuracy by using more complex image local feature descriptions or training deep learning networks to obtain better feature representations.However,occlusion of objects,differences in image resolution,changes in illumination,and more similar image data sets make the retrieval of near-duplicate images a challenging problem.This paper proposes a near-duplicate image retrieval algorithm based on features fusion,which includes the deep features extracted by the convolutional neural network and the improved local features.The features are merged in the retrieval stage,so that the performance of the algorithm is not affected by the single feature property,and the retrieval accuracy is improved.This article specifically describes the following three contribution points:1)An interest domain extracted by a convolutional network(CNN)is proposed for improving the Speeded Up Robust Features(SURF).According to the results obtained by the CNN front-end pooling layer,the channel weight and the spatial weight are respectively calculated to obtain the overall response map.The response map on multiple pooling layers is superimposed,and the final interest domain is calculated by a similar filtering operation.The local feature points are deleted by using the interest domain to optimize the local feature description.2)A central weight is proposed for improving the local aggregate vector(VLAD)algorithm.The weight design idea is derived from the word frequency-inverse text frequency(TF-IDF)algorithm,which is applied to the image field.After calculating the center weight,the part added to the VLAD to quantify the feature is used to improve the effect of the highly significant central word and reduce the effect of the invalid vocabulary,so that the image representation obtained by VLAD quantization is more excellent in locality.3)An improved C-MI retrieval method is proposed.The feature fusion based on the global feature and the local feature proposed in this paper is combined with the graph-based re-ranking method to optimize the initial search results.Search results that more closely match the similarity requirements.