Research And Implementation Of Image Retrieval Technology Based On Deep Learning Feature

With the popularity of e-commerce and social networks,multimedia data,such as images,are growing at a high speed in the Internet.The accurate search for resources from massive image data relies on image retrieval technology.Traditional image retrieval technology mainly uses handcraft features while handcraft features have limitations in the expression of high-level semantics.With the rise of deep learning,more and more researchers turn their attention to this special model that can better express the high-level semantics.Based on deep learning,this thesis mainly studies two problems:feature extraction and feature hashing.Research on the former is devoted to finding ways to improve the ability of feature expression,research on the latter is devoted to generating efficient and semantic-containing hash codes.The main contents are as follows:First,study single layer feature of convolution neural network.A variety of methods to improve the expressive ability of single layer feature are explored.Network fine-tuning and dimensionality reduction can improve the features of dense layers.Feature aggregation can improve the features of convolution layers.Further experiment shows that the feature aggregation method proposed makes the convolution layer feature surpass the dense layer feature.Second,construct fusion feature based on convolution neural network.Two types of feature fusion methods are designed to break through the limitation of single layer feature.One is the dense layer feature fusion which connect two dense layer features with different weights.The other is convolution layer feature fusion which designs a network level fusion model with efficient feature extraction branch and local connection layer.Experimental results show that the fusion model works very well.Third,build an image hash model based on deep learning.We first implement an image hashing model based on the stacked autoencoder and analyze its semantic and efficiency defects.In order to overcome these defects,we propose an image hashing model based on convolution neural network and validates the validity of the model experimentally.Fourth,optimize the image hash model based on deep learning.We propose corresponding solutions to the problems of insufficient utilization of convolutional neural network model information and the fact that hash codes are not restricted by rules.We use the bypass connection and improved loss function to optimize the structure of the neural network,and finally verify the improvement of the network hash before and after the improvement.