Research On Structure Learning Of Sum-product Networks And Its Application In Image Classification

Sum-Product Networks(SPNs)are a new class of deep learning structure with double meanings of deep neural network and probabilistic graphical model,which is beneficial to accurate and rapid probabilistic inference.In recent years,SPNs structure has been widely used in many fields,such as image restoration,behavior recognition,speech recognition,fault location and other research fields,which has attracted much attention.The propose of this thesis is to study the structure learning algorithm of SPNs and apply it to the classification of flower images.The specific research work is as follows:Firstly,the thesis analyzes the basic principles and properties of SPNs structure,and studies the training and inference process of different structure learning algorithms.On this basis,we propose a new SLSPN structure learning algorithm.The algorithm abandons the idea of establishing the initial SPN structure and directly divides the variable set and the instance set from the dataset.Taking into account the characteristics of the structure and dataset,the algorithm adopts square correlation measure and spectral clustering method in the division process,constrains the end condition of the algorithm and adjusts the order of the partition properly.In the experimental,the algorithm is performed on 19 high-dimensional datasets.By comparing with the experimental results of the other two algorithms,it is shown that the algorithm has certain advantages in terms of likelihood score and execution time.Secondly,the method of flower image classification is studied by using SPNs structure as classifier.Because the captured images are easily affected by the clutter background,occlusion,individual morphological differences and shooting angles of the same kind of targets,a certain number of pixel blocks are extracted from the image.The pixel blocks are processed with zero mean,unit variance,and whitening.Then the PCA clustering algorithm and the pooling operation are used to reduce the dimension of the features and extract effective eigenvectors.These feature vectors are used as input of SPNs structure learning algorithm,and the classifier is trained.The simulation results of Oxford 17 Flower and Oxford 102 Flower datasets show that the proposed method has better classification performance in the task of flower image classification.Finally,the work of the thesis is summarized,which shows that the SPN structure is an effective tool to solve the task of image classification and has a fast and accurate probability reasoning mechanism.