The Research Of Multi-instance Multi-label On Natural Scene Image Classification

With the advent of the information age,digital cameras and mobile phones almost become an indispensable part of people's lives.People can take an image in anytime and at anywhere,which led to the explosive growth trend of the number of images.So,the problem urgently need to solve in front of people is how to manage these images according to their class.Based on this,scholars first put forward the traditional framework of supervision learning,but such as more complex images like natural scenes,the framework does not solve the problem well,then scholars have put forward a multi-instances(MISL)learning framework and multi-labels(SIML)learning framework and multi-instances multi-labels(MIML)learning framework.In this thesis,the multi-instances multi-labels learning framework is applied to the classification of natural scene images.The main work is as follows:(1)Several classic natural scene image classification algorithms are introduced,including the MIMLBOOST algorithm and MIMLSVM algorithm which ignore the correlations between the instances and the labels,and the M~3MIML algorithm and the MIML-KNN algorithm which consider the correlations between the instances and the labels.Then,combined with five algorithmic evaluation indicators,the classical natural scene image classification algorithms are evaluated by experiment.The experiment shows that the algorithms which consider the correlations between the instances and the labels are better.(2)The MIMLRBF neural network model developed from multi-instances multi-labels framework is studied.In this thesis,because of the problem that MIMLRBF contains expressions of non-valid examples and isolated points,the algorithm of spectral clustering and an improved distance metric of two packages are introduced into MIMLRBF,and an improved algorithm based on MIMLRBF natural scene image classification is proposed to solve the influences of non-valid examples and isolated points on the classification effect.Then combined with five algorithmic evaluation indicators,the improved algorithm is evaluated with the original algorithm and the classical natural scene image classification algorithms.The experimental results show that the improved algorithm has better classification effect.(3)An improved multi-instances multi-labels image classification method based on sparse coding and deep neural network is proposed.The coding process of sparse coding technology and its application and the process of natural scene image classification are analyzed,and the development of deep neural network and its model design are described.As the process of sparse coding may lead to the incompleteness of the example,the sparse residuals are added to the coding results to form an example with high dimensional vector,which effectively solves the effect of the incomplete coding after sparse coding on the classification effect.Then combined with five algorithmic evaluation indicators,the improved algorithm is evaluated with the original algorithm and the classical natural scene image classification algorithms.The experimental results show that the improved algorithm has better classification effect.