Research And Application On Autoencoder Based Feature Learning Model Of Neural Network

With the rapid development of multimedia and Internet,images have played important role in our daily life because of their property of intuitive expressions.Plentiful semantic information is contained in images.There has been developed a large challenge that how we could seize image people required from large scale image database.Generally speaking,excellent feature expression can not only represent category,but also capture semantic information of images.To address above problems,this paper make researches on feature expression of neural network based on autoencoder.Spatial relation information is contributed to the feature learning besides class labels to boost representation of neural network.The content of this paper is as follows:Firstly,this paper analyzes the research status of image retrieval.By summarizing the key factor of this field is feature expression.And then,we review the development of the deep learning,finding the combination of deep learning and image retrieval through analyzing the feature learning process in neural network.After that,we can find that the feature expression of the neural network by learning more specific target information catch more excellent feature expression ability in image retrieval.Secondly,a neural network learning model based on Autoencoder is proposed in this paper.We use the latent information corresponding to the image to train the neural network,in which the features learned by the neural network have better ability of feature expression.The latent information is obtained though Autoencoder and is seen as the target of neural network to produce boost in case of multi-object representation.Thirdly,this paper also proposed a neural network learning model based on Variational Bayes Autoencoder.Unsupervised model generates the characteristics of the sample data by studying the true nature of raw data.Variational Bayes Autoencoder is the formalization of this problem in the framework of the probability graph model.We maximize the lower bound of the logarithm of the data,so as to achieve our goal of generation in this probability graph model.This part uses this generation feature instead of the traditional Autoencoder feature to enhance the image description ability.Finally,our proposed model is evaluated on PASCAL VOC 2012 segmentation dataset and Microsoft COCO dataset in retrieval task.And then,experiments on kmeans clustering,feature visualization and sparsity show that our proposed model has better feature expression ability.