Study Of Image Classification Method Based On Non-Gaussian Probability Model

Image classification is one of the important research issues in the field of computer vision.It has been successfully applied in many fields and has attracted more and more attention from scholars at home and abroad.Image classification refers to the process of automatically classifying images into a set of predefined categories according to certain classification rules.Image classification has been extensively studied,and a large number of image classification algorithms have been proposed one after another.Probabilistic mixture models have received extensive attention due to their strong expressive ability and have been successfully applied to classification problems,including image classification.In the classification method based on the probabilistic mixture model,choosing an appropriate probability distribution as the basic distribution to describe the potential distribution of the data has a crucial impact on the classification performance.Conventional commonly used classification model is based on Gaussian mixture method,the data analysis and processing on this basis is constructed.However,many data generated in practical applications,such as images,text,audio data and video data,etc.They have characteristics that cannot be described by Gaussian distribution,such as asymmetry,boundedness,and nulti-tailed.Therefore,this assumption based on Gaussian distribution often leads to unreasonable model construction,resulting in unsatisfactory classification results.The core of the related application of non-Gaussian mixture model lies in the reasoning and learning of the model.Mainstream learning methods need to be calculated for the entire training set in each iteration of the optimization process.This belongs to batch learning,so it cannot handle big data or dynamic streaming data,nor can it be promoted in practical applications containing large-scale data or streaming data.In response to the above problems,this article will study image classification algorithms based on non-Gaussian mixture models.First of all,in view of the problem that the Bayesian estimation of the non-Gaussian mixture model represented by the Dirichlet mixture model cannot derive the analytical solution expression,then a complete set of stochastic variational inference algorithms is proposed.Find the evidence lower bound of the initial Variational inference objective function of the sampled data,through strategies such as convex function properties and Taylor series expansion.Using stochastic optimization and natural gradient descent algorithms to obtain the analytical solution expression of the variational posterior distribution,solving the problem of parameter estimation and model selection within the same framework;Secondly,use artificially synthesized data sets to test the accuracy and computational efficiency of the algorithm proposed in this article;A large number of experiments show that the algorithm can estimate model parameters and the number of mixed components more accurately.But it has a faster convergence speed and lower calculation cost than the batch learning method;Finally,the algorithm is applied to solve the image classification problem.A large number of experimental results on the open source image data set verify the effectiveness and feasibility of the algorithm.