Research And Implementation Of Deep Learning Based Augmented Class Learning Method

Learning with Augmented Class(LAC)is an important method to deal with the problem of open and dynamic environment with which most practical applications are faced.Therefore,the research of LAC methods has important practical significance.Previous LAC algorithms are based on traditional machine learning algorithms,and their features are usually artificially designed,often not abstract enough and complete,which will lead to limitations in the use of features.Because deep learning algorithms have great advantages in extracting abstract and complete features,they are often used as feature extraction front ends of traditional machine learning algorithms to improve algorithm performance.Therefore,deep learning is supposed to be used to help solve the LAC problem.Deep Learning-based Label Confidence Propagation(DL-LCP)algorithm was proposed to improve the classification performance,which was adapted from the original LCP algorithm.Compared with the LCP algorithm,the DL-LCP algorithm uses more abstract and complete features for label confidence initialization and label propagation,which greatly reduces the noise accumulation of the LCP algorithm during the model training process,thereby improving the performance of the algorithm.Specifically,a Scatter Loss function was proposed to make full use of all training data(including labeled data and unlabeled data)to learn a feature representation with "make samples belonging to the same class closer and samples belonging to the different classes more separate" property.This feature representation enables samples of known classes to be clustered and samples of augmented class to be spread over a wider open space,thereby enhancing the discriminability of the features.The experimental results show that the DL-LCP algorithm has a significant improvement on the different scales datasets compared to the LCP algorithm and other algorithms in the same field.In addition,in order to further reflect the engineering practice value of LAC method and verify the practical effect of DL-LCP algorithm in real-world applications,the image classification and annotation tool "OpenLabel" with DL-LCP as the core algorithm was designed and implemented.The function and performance of the annotation tool was fully tested and the results show that it presents good class prediction performance and relatively high labeling efficiency in the open dynamic reality environment.