Research On Optimization Method Of Multi-classification Recognition Model Based On Hadamard Error Correction Code

With the improvement of the degree of social informatization,the scale of data on the network is continuously increasing all the time,and the hot issues discussed by people are also constantly changing.In order to effectively process and recognize these new data and make use of the implied valuable information,people increasingly need a more complete classification method.In the field of supervised learning,the traditional multi-classification model based on probability statistics is mostly limited to a predetermined data set.In order to enable the trained model to recognize a new class,all parameters of the model need to be retrained,which is unreasonable for the practical application of the algorithm.Therefore,a multiclassification model that can add new categories at any time is needed by people.Under the premise of supervised learning,it can make certain changes according to the characteristics of new samples so as to learn and complete the classification of the information contained in new samples.In addition,the cost of such changes should be lower than that of retraining the whole model.This learning mode is similar to the process of human self-learning,so it has high research significance and future.In this paper,a multi-classification optimization method is proposed in the field of image classification.This article USES the constraint condition of the iris image classification as many experimental data,using the idea of divide and conquer a classification problem into a more complex than simple multiple binary classification problems,thereby improving model of flexibility,in order to ensure the overall performance of image classification model combines the artificial neural network powerful ability of feature extraction.Specifically,this paper takes a Hadamard error-correcting output coding method as the framework of "manyto-many" problem decomposition,uses a self-designed convolutional neural network to train the basic classifier to ensure the performance of each iris dichotomy,and solves the multi-classification problem by combining the two.The model is called Hadamard_ECOC&CNN.Hadamard error correction output coding can be used to construct a decomposition framework applicable to any number of categories,and it also has a good correction effect in the case of insufficient samples,incomplete features or defects in the algorithm itself.The local perception,Shared parameters and other characteristics of the convolutional neural network make the model more suitable for image classification and recognition without the need to manually select features.For iris images,it can also ensure a certain degree of invariance of displacement and shape.And,more importantly,through the combination of using error correcting output codes do not need all of the code word completely corresponding to category this characteristic can make the model in the case of only retrained parameters to add a new category,in addition,in this paper,through the comparison test sample code and standard code,can be classified according to the displacement track directly to the error of the base classifier,corrected by trimming technology training to optimize their parameters so that the overall precision of the model can be further promoted.In this paper,the CASIA-IRIS-LAMP-V4 data set published by the institute of automation of the Chinese academy of sciences and the JLU-4.0 IRIS data set collected by the biometrics and information security technology laboratory of jilin university were used for the experiment,and the recognition accuracy of the test set was used as the performance evaluation index.The experimental results show that the model achieves the accuracy of 98.19% and 96.35% respectively on the two data sets,and the classification performance is better than some traditional multi-classification methods.Also through to the correct training and add new classes of these two new function come to a conclusion,experiment analysis shows that correct training can be restricted to certain improve the overall precision of the model,and by setting a reasonable threshold can be retrained only part of the base classifier can realize the function of adding new classes.