Research On Few Shot Recognition Of Remote Sensing Image Based On Incremental Learning

Remote sensing image recognition algorithms have important applications in military and civilian fields.When acquiring image information of a new target in an intelligence reconnaissance mission,the new target needs to be identified on the basis of the original model.Without using the original training data,relying on a small amount of new target data for incremental learning of the model has important research significance.The classic remote sensing image recognition algorithm based on deep learning cannot adapt to the needs of few-shot recognition tasks,nor can it solve the catastrophic forgetting problem caused by incremental learning.Therefore,this thesis focuses on the remote sensing image few-shot recognition algorithm and the remote sensing image few-shot recognition algorithm based on incremental learning.Aiming at the problem of insufficient training samples in the few-shot learning task,a few-shot image recognition algorithm based on transfer learning（TFN）is proposed,which uses transfer learning on the feature extraction network,and then performs few-shot learning on the classifier to make up for the lack of training data.Aiming at the defect that traditional data enhancement methods cannot perform semantic conversion,a few-shot image recognition algorithm based on implicit semantic data enhancement（ISDA-TFN）is proposed,which realizes semantic data enhancement by sampling feature vectors in a specific direction in the feature space and improves the generalization ability of the algorithm.Experiments on the public remote sensing image data set have verified that ISDA-TFN has higher recognition accuracy and better generalization ability.Aiming at the problem of incremental learning under the condition of few samples,this thesis designs a few-shot recognition algorithm based on incremental learning（ITFN）based on ISDA-TFN,and realizes incremental learning by merging classifiers.Firstly,a few-shot recognition algorithm based on incremental learning（ITFN）is proposed on the basis of ISDA-TFN,and incremental learning is realized by merging classifiers.In order to solve the problem of low model recognition accuracy due to the large difference in the data distribution of the new and old categories,a few-shot incremental learning algorithm based on cosine similarity（CTFN）is proposed.By using the cosine similarity to calculate the classification score,the influence of the classification weight vector modulus is eliminated,and the recognition accuracy is improved.Aiming at the problem that direct learning of new categories leads to catastrophic forgetting of old categories,a few-shot incremental learning algorithm based on attention attracting network（Att-CTFN）is proposed.The gap between the classification weights of the new category and the old category is reduced,and the accuracy loss of the old category is reduced.Experiments on the public remote sensing image data set have verified that Att-CTFN has higher recognition accuracy and stronger anti-forgetting ability.