Research On Image Content Analysis And Storage Based On Machine Learning

With the popularity of computers and the development of application platform,the explosive growth of image data poses a huge challenge for data analysis and storage.The time delay duing to data access for large scale image analysis application is a tricky problem.The semantic information asymmetry between image applications and storage is the prime culprit.In recent years,the intelligent algorithms for large-scale image data are dominating the research field.Machine learning is one of the most commonly used methods in intelligent algorithms with characteristics of offline learning and online generation.It can solve the problem of association,classification,annotation and Hash mapping by extracting image content semantic,and improve the capability of image semantic analysis and reduce online time consumption.Machine learning for image is divided into shallow learning and deep learning by proceeding.The former consists of two segments.Firstly,extract features according to traditional methods,and then use the shallow machine learning to complete the task.The latter is an artificial neural network that can be end-to-end feedback.In the network,the convolution neural network is used for feature extraction and the task network is composed of multi-layer perceptron and a loss function.Because of independence of the feature extraction and task,shallow learning is hard to get better results.Deep learning is just to solve this problem.But because of the poor design of network parameters and the dependence on the labels,so that it is hard to apply to the actual circumstance.In view of characteristics,I research on image annotation,image Hash and image storage by shallow learning and deep learning respectively.Image annotation is an important topic in image semantic analysis,which can establish the qualitative relationship between image content semantics and labels.In order to settle problem of instability of single view,I apply ensemble learning,which is a kind of shallow machine learning,to carry on the multi-view multi-label image annotation research.At first,I extract features by BOF,VLAD and PHOG for different views.Then,using Bagging and Boosting to train SVM.At last,to learn the basic classifier weights on validation set innovatively to complete annotation classifier.The results on Label Me and VOC2012 show that single view is unstable and MVML is always better.Especially,I try different order structures in this process and find that the effect of the difference between the basic learner with the ability to get better results.In contrast to the MKL-SVM algorithm,the MVML annotation accuracy is 3.5% higher at most of the time.The content semantic information generated by image annotation technology has limitations in expressive ability and alignment ability in large-scale environment.Similarity Hash is considered to be an effective method because of lightweight expression and convenient comparison.Deep learning hashing can promote relavance of features and task.But the problem of without label is challenge to impose it.In view of the above problems,I present the deep self-taught similar Hash framework for image content-based semantics(DSTH).DSTH is composed of hash label generating stage and hash function learning stage.In hash label generating stage,I use the deep and shallow mixed learning.Firstly,I get image features by finetuned with CNN and existing model.After that,graph model of features is constructed using KNN algorithm and mapped using LE algorithm.Then I apply binarization to map results to hash codes as hash labels.In hash function learning stage,I apply deep learning framework with shallower network to learn hash function through the hash labels generated in hash label generating stage.In this processing,I will select Slice network as task network because of its superiority on Hash.Moreover,I will combine with Batch Norm as the activation function and Euclidean distance loss function to complete the whole learning task.I config different task networks and different activation functions on different datasets for our framework and compare with each other.At last,I find configuration of Slice with Batch Norm under feature of Goog Le Net is better.Based on the results of 32-bit Hash on CIFAR-10 and 48-bit Hash on STL-10 with classification label and without label,DSTH is better than others on precision recall and number of return image,and sometimes better than others on code analysis.The results prove that DSTH can impose the deep learning in unlabeled environment under the premise of guaranteeing accuracy.In addition,I test generalization ability of DSTH on CIFAR-100 by different classification for same data.It is proved that DSTH has strong generalization ability and can be used in environment with variable image classification.In order to solve the problem of delay time,a scheme of content-based image stor-age system is proposed.I introduce conception of graph meta-data and carry on research of intelligent image content-based semantic storage system.I propose an innovative image storage system architecture that is compatible with deep learning techniques and construct metadata organization in the form of graphs using similarity hash codes based on semantics of image content.First of all,I design a new storage system architecture that can repeat the learning and generate semantic content,and give a solution to update learning and semantic metadata.Then,I give the scheme of using similarity Hash as semantic metadata to add metadata,and how to construct and manage semantic metadata with the structure of graphs.In addition,I propose Sem Rank promoted form Page Rank with character of Hash graph.Experiment results show the effectiveness and reliability.For the application of image retrival,the Sem Rank can calculate the global influence of the node and locate hot data.Finally,experiments show that the design scheme can be applied to large-scale environments through the insertion time trend of nodes in the graph database.Moreover it can build bridge between application of image analysis and storage with semantic meta-data.