Abstract

Protective coatings are the first line of defence against corrosion on naval, aerospace and land defence platforms, yet their condition is still judged largely by human visual inspection, which is subjective, difficult to audit and hard to scale across a fleet. Recent deep-learning studies automate corrosion and coating-defect detection, but the systems reviewed here mostly return generic labels such as “rust” rather than the graded severity ratings that quality systems expect under ISO 4628, SSPC-VIS and AMPP/NACE practice. This paper reviews ten works published between 2018 and 2025, covering supervised CNNs, post-hoc explainability, few-shot learning, edge prognostics and acoustic-emission classification, and compares them in a year-wise survey matrix. The matrix exposes four gaps: no standards-referenced severity grading, data scarcity in defence archives, little multi-modal fusion, and security constraints that rule out centralised training. In response, we formalise a defect taxonomy mapped to the three standards families, propose a severity-scoring formulation based on extent, density and trend, and outline a five-module conceptual framework that combines multi-modal fusion, few-shot learning, standards mapping, clause-linked explanation and federated edge deployment. The framework is not implemented or evaluated here; a protocol for its evaluation is proposed as future work.

Keywords

How to Cite This Article

Rights