Research On Content-based Image Retrieval Based On Deep Learning

The development progress of the network makes the image sources more and more widely,the use of text search image can not satisfy people's demands which about the fast browsing,searching and querying images.Content based image retrieval system is in the process of image database,find the similar image content according to user needs automatically,and return the results quickly to the users.The traditional retrieval methods are mostly based on image texture,shape,color and other simple features,It is necessary to specify the condition of image feature retrieval,which has some disadvantages such as single condition and limited searching range.How to effectively extract features,how to realize image retrieval is the key point of Content-based image retrieval to realize image retrieval.This is an image retrieval method based on deep neural network structure.Deep learning neural network simulates the biological nervous system,and has good retrieval effect in image retrieval.The depth of deep learning neural network using multi hidden layer structure,learning image data structure and classify the features automatically.Training face image data samples using depth belief network,fusion of local binary pattern to extract stable features,an unsupervised greedy training method is used to limit the Boltzmann machine by layer by layer,the training weights,bias and other parameters are used to predict the test samples,then,calculate the cost function,and fine-tuning by error back-propagation.Because of the depth of the neural network training sample data of the huge demand,therefore,the combination of local texture features as the input characteristics of deep belief networks,contribute to the deep understanding of the deep belief network distribution image,extract features' parameters,reduce the probability of network learning to unfavorable characteristics.using the ORL face database,and add a random sampling data of face recognition,discussed the relation between the number of iterations and the result of recognition.As a comparative experiment on the shallow BP neural network learning the algorithm of face retrieval,The recognition rate of BP network is lower than that of deepnetwork,the experimental results show that the deep learning face image retrieval accuracy is satisfactory.