Research On Image Retrieval Method Based On Fusion Of Deep Features And Hand-crafted Features

In this thesis,an image retrieval method based on the fusion of deep features and traditional hand-crafted features is proposed,and the effectiveness of this method is verified on relevant image datasets.First,attention is paid to deep global features.In this thesis,the pre-trained VGG16 neural network is applied to extract bottleneck features and two fully connected layer features for global image representation.Furthermore,these features get optimized by some methods,which include dimension reduction with PCA and fusion of features extracted from different layers.Deep convolutional features are also extracted from the same neural network.And an aggregation method based on spatial weighting and channel weighting is designed to get the global representation for images.The spatial weight is computed on variances of different feature maps and the channel weight is based on the sparseness of different channels.Experiments on different datasets for image search prove that this method outperforms the current aggregation methods on pre-trained network.Next,traditional hand-crafted features are coded in this thesis.SIFT features are selected for image representation,and then VLAD method is used for global coding.In the process of acquiring codebook by clustering,FCM is used to replace the traditional K-means clustering algorithm,and its advantage is highlighted through experiments.Then,VLAD coding of image features is obtained by accumulating the residuals of feature vectors and cluster centers.Then,an image retrieval method is proposed,which combines the three kinds of features introduced above.The features are normalized and weighted fusion is performed.Dimension reduction and whitening are performed for the fusion features.When calculating the similarity of image features,query expansion is used to further improve the accuracy of image retrieval.Last,experiments are performed on different datasets to compare the retrieval accuracy between the fused features and some other methods.And the results verify the feasibility and effectiveness of the method proposed in this thesis.