Underwater Acoustic Target Feature Fusion Recognition Using Multitask Learning

This paper introduces a feature fusion recognition method for underwater targets based on multitask learning (MTL). Initially, to address the limitations of single-feature input in radiated noise spectrum recognition algorithms, this paper proposes an improved Vision Transformer (ViT) for depth learning underwater target classification. Building upon this, the problem of multifeature fusion recognition is reformulated into an MTL framework, and an optimized feature fusion recognition model is proposed using the multiple gradient descent algorithm (MGDA). The performance advantages of the proposed algorithm is validated on the underwater open source dataset DeepShip, offering a novel approach to underwater target feature fusion recognition.