Research On Image Retrieval Technology Based On Deep Learning

With the development of Internet and multimedia technology, digital images people can get are massively increasing, how to extract useful images from the image database becomes a problem needed to be solved. Always, the common way of image retrieval is based on text, namely search by the keywords. However, as the number of images increased, the image retrieval method based on text comes out problems hard to be solved. In contrast, for content-based image retrieval technology, the system will analyze and understand the content of the image automatically, and extract the visual features for image retrieval. These visual feature includes color, texture, shape, and the key points and so on. However, the feature descriptors designed by people often do not have the characteristics of generality and universality, one descriptor may have a good expression for a particular object, but do not do well for others. In order to solve this problem, this paper proposed the use of deep learning techniques to extract the image feature, trying to get a robust feature extraction methods, the main work is as follows:(1) Studied the key technology of content-based image retrieval, analyzed and compared the advantages and disadvantages of these classical techniques; studied some key technology of deep learning, and grasped the deep learning methods of constructing network, training algorithm and optimization method.(2) Constructed a stacked auto-encode neural network model, proposed a learning algorithm and feature extraction method based on autoencoder neural network; designed an image retrieval system based on feature leaning, verified the effectiveness of the proposed network model through the experiments.(3) Constructed a convolution neural network model by different convolution network level, proposed a learning algorithm and feature extraction method based on deep convolution neural network; designed an image retrieval system based on deep convolved feature, verified the effectiveness of the proposed network model througth a large number of experiments.This paper hopes to build deep structure of neural networks to learn useful features from the image data, and apply them to the image retrieval system.