Research On Image Retrieval And Image Semantic Feature Based On Deep Learning

With the rapid development of big data and artificial intelligence,the number of large image data sets has grown and image retrieval technology has also developed rapidly.How to make users find the required image accurately from the massive image database has become a key problem in the field of image retrieval.To solve this problem,the feature extraction and aggregation technology based on deep learning is studied,multi-region strategy is proposed to generalize mean pooling,and at the same time convolutional features and traditional features is fused to generate feature vectors for image retrieval combing the properties of traditional low-level semantic features.The main research in this thesis are as follows:(1)As the GeM algorithm fails to effectively increase the influence of the target region when performing convolution feature aggregation,a multi-region average pooling algorithm RGeM is proposed.The algorithm first extracts the convolutional visual features of the input image to obtain feature maps,then divides these feature maps into multiple regions,then performs generalized mean pooling on each region to obtain feature vectors,and finally generates dimensionality reduction and other operations on the feature vectors.The RGeM algorithm applies a multi-region strategy to generalized mean pooling,pooling each region and increasing the impact of the target region.Through experiments on the Oxford5 k and Paris6 k data sets,the results show that the average accuracy of the RGeM algorithm for image retrieval are better than the commonly used image retrieval algorithms such as RMAC and GeM,respectively reaches 84.2% m AP value and 75.3% m AP value on the two data sets.(2)For large-scale image data,convolutional visual feature maps can't fully represent image features.Therefore,the characteristics of traditional image features is used to fuse convolutional visual features with traditional image features,and a multi-feature fusion algorithm T-RGeM is proposed.The T-RGeM algorithm fuses RGeM features with SIFT,FV and Bo VW features.First,the VGG16 network model is used to extract the convolutional visual features,then the traditional image features are extracted,then the two features are cascaded,and finally the dimension reduction is performed.The experiment was conducted on Oxford5 k and Paris6 k data sets.The results show that when the fused features are reduced to the same dimension as the convolutional features,the averageretrieval accuracy of the T-RGeM algorithm is still better than the RGeM algorithm.(3)In information retrieval,query expansion is an important operation.In the experiment of this thesis,firstly a certain enhancement operation was performed on the image,and then the query expansion operation was performed on the experiment to test the performance of the two algorithms RGeM and T-RGeM.The final experimental results show that the use of query expansion has improved the accuracy of image retrieval to a certain extent.It is also shown that RGeM algorithm and T-RGeM algorithm have certain advantages in image retrieval.