Underwater Acoustic Target Recognition Based on Adaptive Sample Reweighting and Feature Fusion

Recently, deep learning has made groundbreaking advancements in underwater acoustic target recognition (UATR), significantly advancing the field. Despite these successes, it is observed that classifiers tend to overfit biased training sets, especially when the samples exhibit classes imbalance. This tendency severely undermines the classifiers’ performance on the test set, leading to inadequate generalization capabilities. To tackle this issue, this paper introduces a target recognition method based on adaptive sample reweighting and feature fusion. This method employs ResNet32, which has demonstrated excellent performance in the field of deep learning image classification, as the baseline classification network, and utilizes both static and dynamic features as its input. Additionally, the meta set guides the training set to learn decision-making rules under unbiased conditions by updating the parameters of the meta weight network (MW-Net). Experiments demonstrate that the proposed approach can effectively enhance the generalization capabilities of the classification network to unknown datasets and optimize the performance of UATR in conditions of class imbalance.