Automatic Image Management Investigation Based On Image Content Analysis

With the rapid development of multimedia technology,more and more user generated data is produced at an ever-increasing rate.At the same time,the problem can not be neglected that how can we manage these resources effectively.There has been some research on this topic recently,while it’s not that systematic and problems arise applying them into daily use.This paper aims at managing images by image content understanding.Specifically,it includes scene classification,the design and optimization of search algorithms,the transfer from face recognition to face verification and the pursuit of efficient feature learning algorithms.In terms of scene classification,this thesis optimizes loss function which originally used for single class classification to adapt to the multi-label scenery,which suits more of daily life.In addition to this,the thesis adopts the Shufflenet architecture,which is mobile-device friendly,as the base model architecture.To improve classification accuracies and make the network more adapt to the dataset,Squeeze and Excitation Network has also been used,facilitating as attention across feature channels.Thus,the concern of data privacy,large space consumption due to huge model size and network delay for running model on the cloud can be alleviated.Apart from this,a balanced binary cross-entropy loss is designed to better facilitate the demand.An iOS real-time demo has been made to test the performance of the scene classifier.According to the result of the test,the design of such a model has proved to be efficient and effective.When it comes to face verification,a set of regularizations are proposed to better guide the training process of face recognition task.The regularization takes advantage of intra-class compactness and inter-class discrimination.Besides,it makes use of redundancy restriction,which regularizes non-diagonal parts of the covariance matrix.A radial basis function is introduced to tackle the problem of label noise,which is significant for large scale webly-collected dataset,while no explicit separate procedure for data cleaning is needed.Experiments show that although not directly optimized for the verification task,the deeply learned features are transferrable to the verification task,and the proposed regularization does not incur runtime overhead at test time.To the part of search,two approximate search algorithms are introduced,inverted file list and product quantization,which focus more on speed and memory usage,respectively.In this way,either algorithm could be chosen to better suit the demand of various conditions.Meanwhile,this paper refined several aspects where the original implementation lags,to further boost the speedup of search speed.