Uncertain Handwritten Digit Recognition Research Based On Belief Random Forests

With the development of information technology,massive data has flooded into social production,and the uncertain information of data from different channels has gradually increased.For example,in the process of image acquisition and storage of handwritten digit recognition,the image is often missing or fouled due to various factors,and then the uncertainty is introduced.In practice,the processing methods in the practice are mostly to complete the indeterminate reasoning by manual labeling,and then based on the determined data.The set performs learning modeling,which is inefficient,costly,and poorly stable.How to efficiently process uncertain information and complete the learning and construction of classifiers based on uncertain datasets has become a new challenge in handwritten digital recognition research.Belief function theory provides a new idea and method for uncertain data processing,which has attracted wide attention in recent years.In order to solve the problem of learning and modeling uncertain handwritten digital datasets,this paper proposes an improved belief random forests algorithm based on machine learning decision tree theory.By introducing belief function theory and ensemble learning idea,a new attempt to effectively model uncertain datasets and complete handwritten digital recognition is presented.To solve the uncertainty of class label missing from training sample output,the improved algorithm introduces the idea and method of belief function based on evidence theory,by constructing the mapping between sample spacing and basic belief allocation function,the basic belief allocation function value of uncertain samples is calculated,and the prediction and reasoning of class label of sample output is completed by using evidence combination rules.For the uncertain attribute values of training samples,by adjusting the weight of samples and optimizing relevant calculation,the optimal partition attribute selection and the classification of missing samples are completed,and random attribute subset selection rules are introduced to reduce the risk of "over-fitting" of decision tree in the process of partitioning nodes of decision tree,which realizes the construction of belief random decision tree.In order to improves the generalization performance of learners,the idea of ensemble learning is used to combine several base belief random decision trees to generate belief random forests,the learning and modeling of uncertain handwritten digital datasets are completed.In order to evaluate the belief of random forests algorithm modeling uncertain datasets and realize the actual generalization performance of handwritten digit recognition,this paper carries out experimental verification research.Firstly,binarization and refinement of dataset images,the handwritten digital experimental dataset is generated.Then,the training set data samples are extracted from statistical and geometric levels,and the input feature vector of handwritten digital training set are constructed.Finally,the belief random forests algorithm is programmed by Python programming language,and the algorithm is run on the uncertain handwritten digital dataset to verify the experiment,the model parameters are adjusted,optimized and implemented.The research and analysis of the experimental results demonstrate the feasibility and efficiency of belief random forests algorithm in modeling uncertain data sets and realizing handwritten digit recognition.