The Research Of Image Retrieval Technology Based On Deep Learning

With the popularity of mobile devices such as mobile phone and the rise of the shopping website and social network, online image has seen an explosive increasing. It has always been one of the hot issues of content-based image retrieval to retrieve relevant information, while the features selected directly affect the performance of retrieval. Deep learning has recently advanced the state-of-the-art in image classification by extracting features layer by layer. In order to adapt features extracted from classification model to image retrieval, we do the following work:First of all, the basic structure and learning methods of deep neural networks are studied and also a comprehensive of hidden layers of convolution neural network, a detailed analysis of Le Net and AlexNet network model, and the research on current mainstream deep learning framework.Secondly, use Transfer learning to Fine-tune neural network model so that new datasets can use trained models on other datasets. Through the analysis of model training process, it confirms that model fine-tuned have better accuracy and training time than directly trained model. It also demonstrate that features extracted from classification model can achieve better retrieval performance than other features.Finally, in order to build efficient indexing system of image retrieval, we first use PCA to reduce dimension of features extracted from model, experiments shows that features throw dimension reduction have better performance; then we use hash technology to binary features and do the experiments respectively use PCA binary and group binary, the experiments show that features binarized still have better retrieval performance; the performance was test on ImageNet dataset at the end.