Research And Implementation Of Image Classification Algorithm Based On Discriminative Semi-supervised Dictionary Learning

In recent years,machine learning and artificial intelligence have been rapidly developed.Computer vision tasks such as face recognition and behavior detection have been used in practical scenarios.In traditional machine learning methods,supervised learning has always played the most important role since label information for training data.However,in most practical applications,labeled training data is usually very limited,and it is relatively easy to acquire large amounts of unlabeled training samples.Therefore,semi-supervised learning with only a few marked samples has attracted the attention of researchers.In all semi-supervised learning methods,semi-supervised dictionary learning is one of the most promising directions.The core of semi-supervised dictionary learning is how to learn a classifier dictionary with discriminative features through a small number of labeled samples.Different coding representation methods and adding different discriminative items will have an important influence on the classification ability of the last learned dictionary.This paper presents two different semi-supervised dictionary learning algorithms,which have achieved better performance in different image classification applications,such as autism medical image classification,face image recognition,digit recognition and object classification tasks.The specific contributions are as follows.Firstly,an improved label propagation algorithm combined with semi-supervised dictionary learning is proposed.This method adopts a common and specific joint coding method for training data based on the assumption that the samples with identical label have commonality and specificity.At the same time,the improved label propagation algorithm will more accurately estimate the probability of unlabeled samples belonging to each category.Experiments were conducted on four autism datasets collected from different universities.The experimental results show that our semi-supervised dictionary learning method has achieved a better recognition rate than the existing methods.Secondly,we propose a semi-supervised dictionary learning method with maximum entropy where the maximum entropy regularization term can effectively avoid over-estimation of unlabeled samples on uncertain classes.This allows unlabeled samples to continuously approach the real sample class space in dictionary update iterations.At the same time,the joint combination of dictionary optimization and probability updating of unlabeled sample make our approach more interpretive.Through the experimental results on Extended Yale B,LFW face database,USPS and MNIST digit database,and Texture25 target classification datasets,we can see that our proposed method outperforms the existing methods in many pattern recognition tasks.In summary,this paper proposes an image classification task based on semisupervised dictionary learning.Through the rational assumption,i.e.,commonality and particularity of the samples,and different discriminative terms,our method can more effectively explore the discriminative information hidden in unlabeled data which make the learnt dictionary classifier have stronger classification ability.