NAS-MFF: NAS-Guided Multiscale Feature Fusion Network With Pareto Optimization for Sonar Images Classification
Yule Chen, Hong Liang, Shaohua Jiao
Abstract
Underwater target recognition technology based on sonar images has received considerable critical attention in recent years. However, the sonar sensors encounter disturbance from seafloor reverberation noise and a complicated background, resulting in notable difficulties for precise sonar target classification. On the other hand, traditional machine learning methods inevitably lose features relying on expert systems, and manual network creation is relatively inefficient with limited sonar data. To tackle these challenges, we propose a neural architecture search (NAS)-guided multiscale feature fusion (NAS-MFF) algorithm for sonar images classification based on the differentiable architecture search. Specifically, our approach consists of two stages: a search stage with the Pareto optimization, and a training stage using the optimal architecture. NAS-MFF begins by reconfiguring the search space based on the characteristics of sonar images, which includes the introduction of the MF Block <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">${k}$ </tex-math></inline-formula> with multiscale feature extraction ability. By synergizing a recognition-driven convolutional neural network (CNN) with Pareto optimization, it achieves a dual advantage in both accuracy and model efficiency using the available data. Extensive experiments on three sonar image datasets of different sizes and sources distinctly demonstrate that NAS-MFF outperforms several existing manual design methods and NAS approaches.