Research On Image Retrieval Based On Multi-feature Fusion Coding Of CNN

In recent years,with the popularity of the Internet and the development of the social network,the image retrieval technology has been rapidly developed.At present,the feature extractions of image retrieval mainly include hand-craft feature extraction and convolutional neural network(CNN)feature extraction.CNN feature based image retrieval,to some extent,has solved the problem of low retrieval efficiency,and depend on which the speed and precision of image retrieval are improved dramatically.Therefore,the CNN feature based image retrieval technology has become one of the most commonly used image retrieval technology.However,there are still three problems that need to be solved:(1)The "semantic gap" between image features and images perceived by human beings is still exists.(2)The image retrieval efficiency becomes lower when the length of the dimension of image coding grows higher;(3)The feature extraction methods corresponding to different datasets are not robust.Based on the existing work,this thesis studies the following two issues to further solve the problems above:(1)To reduce the "semantic gap" and solve the problem of the efficiency reducing of image retrieval,this thesis proposes an image retrieval technology based on fusion hidden layer feature coding(FLFE).By modifying the structure of convolutional neural network,FLFE achieves the fusion of features extracted from local hidden layer and global hidden layer,making the image coding contains not only local feature information but also global feature information,which reduces the "semantic gap" and improves the accuracy of image retrieval.At the same time,the fixed-length encoding in hidden layer is used in FLFE to reduce the dimension of features,thereby solving the problem of retrieval efficiency.(2)For solving the problem that the existing feature extraction methods are usually not robust,this thesis proposes Compressed Joint-layer Feature Encoding(CJFE)and CJFE based image retrieval technology.CJFE fuses features combine multi-layer features of CNN,which reduces the "semantic gap" and improves the accuracy of image retrieval.At the same time,CJFE makes full use of the sparsity of deep of the encoding of convolutional network,and reduces the dimension of the feature by using the LDA reduction algorithm without losing retrieval accuracy.Finally,the retrieval accuracy of CJFE exceed that of the existing high-dimensional feature codes and greatly improves the image retrieval efficiency with a encoding length of only 256 bits.The researches in this thesis are able to be applied to large-scale image retrieval and image data storage.That is,the images are stored in the computer using the corresponding image codes,which can not only reduce the storage size of image data,but also speed up the process of image retrieval.