Negative Correlation Incremental Integration Classification Method for Underwater Target Recognition

Abstract

In this paper, an incremental learning algorithm based on negative correlation learning (NCL) is used as an identification classifier for underwater targets. Based on Selective negative incremental learning SNCL (Selective NCL) algorithm in the process of training, there are numbers of hidden layer nodes that are difficult to determine training time. Problems such as over fitting analysis arise. The algorithm combined with Bagging makes the difference between individual network further increase, and ensures the generalization performance of the whole. On the basis of this method, the use of the selective integration method based on clustering and a new proposed algorithm called SANCLBag, combined with the convolution of underwater target recognition neural network shows that the proposed integration approach can make the difference between individual network in the classification process further increase, and ensure the whole generalization performance. The model has higher identification accuracy, and can effectively solve the problem of incremental learning. . Growing NCL (GNCL). In the former, the number of learner is fixed, and when new data training sets are added, the neural network integration trained before would be used for training. Different from the former, in GNCL, the number of learners in integration would increase as the new data training sets are added. When there are new training sets, a new ANN is constructed and the NCL algorithm is used to train in new training sets, which would be added into the previous integration after training. Selective negative incremental learning based on clusteringThis paper has referred to the selective negative learning framework of Minlong Lin and modified the model training and model selection of it. The selective negative incremental learning framework is shown in Figure 1.