基于多专家动态协作学习的长尾声呐图像识别OA

Sonar image recognition plays a crucial role in the field of underwater environment detection.While existing sonar image recognition models based on deep neural network have improved classification accuracy,they often face the challenges of long-tailed distribution in practice,leading to insufficient identification of rare yet high-value targets.To remedy this,we propose a novel Multi-expert Dynamic Collaboration model to enhance recognition accuracy for long-tailed sonar image(MEDC-SI).Our model consists of multi-expert network and dynamic learning strategy.The multi-expert network contains a conventional branch for feature representation learning and two re-balancing branches for tail samples learning.And three experts collaborate to achieve imbalanced sonar image recognition.The dynamic learning strategy is designed to shift the focus of model training between the conventional branch and re-balancing branches to improve the feature learning and classifier learning simultaneously.Finally,extensive experi-mental results on three sonar image recognition datasets,KLSG,FLSMDD,and NKSID,demonstrate the effectiveness of the pro-posed model,achieving overall accuracies of 91.51%,99.74%,and 96.19%,respectively.