Target Recognition Method Based on Incremental Learning
Weize Qin · 2022
The traditional batch learning method is only suitable for the solidification recognition model with a fixed number of samples. When the number of new training samples is increasing, the existing and new training samples must be reintegrated, and the recognition model must be retrained to complete the model update. The time cost, computational cost and storage resources of each update are very high, which greatly reduces the efficiency of recognition. In this paper, a new target recognition algorithm based on incremental learning is proposed. By expanding the feature extraction network and classification network, the problem of catastrophic forgetting is effectively suppressed. At the same time, the decoupling small sample incremental learning model based on prototype registration uses meta-learning method in the training process to effectively improve the efficiency of model learning on small sample data. The results are verified on CIFA100 and Tianzhi Cup challenge data sets. After two incremental learning, the recognition accuracy of the algorithm is reduced by only 21 % and 13 %, which is higher than the existing best algorithm by 4.7 % / 19.6 %. The results show that the method based on sample playback and model expansion is compatible, which can improve the accuracy of existing methods.