Research And Application Of Image Classification Based On Fast Fisher Discriminant Dictionary

Recently,with the decrease of data storage cost and the improvement of computer computing performance,huge volumes of data have emerged,which makes the algorithm based on deep learning model yield unusually brilliant results in many application fields.Especially in the field of image classification,image classifier based on depth learning model has been widely concerned because of its high classification accuracy.Although deep learning has shown strong capability of feature extraction and expression in this field for its complex network structure,the insufficient number of training samples has seriously affected the performance of these methods.Moreover,the neural network model is difficult to explain,which seriously limits the application and development of deep learning,such as in the case of high requirements for the interpretability of the model.Therefore,how to estimate the accuracy of sample classification efficiently and accuately on small-scale data sets remains to be solved.At this point,the dictionary learning model is a good choice.However,most dictionary learning models mostly use samples with fixed dimensions,and few studies works on the learning of sample dimensionality reduction matrix.To solve these problems,this paper proposes a Fisher dictionary learning algorithm for fast image classification and an image dimensionality reduction and classification algorithm based on discrimination dictionary.For image classification based on small-scale data sets,dictionary learning technology is used to quickly optimize the coding coefficient and discrimination dictionary by collaborative representation and calculating the closed solution of the coding coefficient one by one sample;Then,according to the learned dictionary and representation residuals,the calculated coding coefficients are also discriminative;Finally,the two discriminant information are used for classification.For the problem that the sample dimension reduction matrix is fixed,the dimension reduction matrix is added to the training process to further improve the classification accuracy of the proposed algorithm.This paper proposes a better classification model by studying collaborative representation,dictionary learning and dimensionality reduction.The main contributions of this paper are as follows:(1)A Fisher dictionary learning algorithm for fast image classification,called CRFDDL,is proposed.The FDDL algorithm and its related practical applications can’t be efficiently calculated for the use of iterative solutions.First of all,2-norm are used to replace 1-norm and the coding coefficients are solved one by one,so the closed solution of the coding coefficients obtained,which provides a novel update method for CRFDDL.Then,classification is carried out by combining discriminant dictionary and discriminative coding coefficient.The experimental results show that the proposed algorithm can update the model efficiently and has higher recognition rate compared with the classical Fisher discrimination dictionary learning and other deep learning algorithms.(2)An image dimension reduction and classification algorithm based on discrimination dictionary is proposed.High-dimensional data is often encountered in image processing.Most dictionary learning methods regard dimensionality reduction as two independent processes.First,dimensionality reduction is performed,and then dictionary learning approach are used for classification.However,this processing method will lose part of the discrimination information in the sample,reduce the interaction with the learned dictionary,and affect the classification results.To solve this problem,first of all,the learning of dimensionality reduction matrix,projection matrix and discrimination dictionary are added in the training process,so that they can match better;Then,according to the learned dictionary,projection matrix and representation residual,the calculated coding coefficient is also discriminative;Finally,the three discriminant information is used for classification.The experiment proves that ADRDDL can better save the discrimination information of sample than CRFDDL and FDDL,and has a better recognition rate.(3)An image classification system based on Fisher discriminant dictionary is proposed.The system is the application and visual display of the models mentioned in the previous two chapters,mainly including three modules:basic module,dictionary learning module and classification test module.The system can intuitively display the process,classification results,training and classification time of the proposed algorithm,which is helpful for comparative analysis and performance comparison with existing image classification algorithms.