Scholarlink

<p>Corrosion poses a persistent challenge to marine infrastructure, resulting in structural degradation, increased maintenance costs, and safety risks. Although deep learning has shown promise in automating corrosion detection, comprehensive benchmark studies using large-scale, high-resolution datasets remain scarce. This study presents a binary classification framework employing four state-of-the-art convolutional neural networks (CNNs)—VGG16, VGG19, InceptionV3, and ResNet50—to distinguish between corroded and healthy marine surfaces. Experiments were conducted using the publicly available Marine Corrosion Dataset, consisting of 9,000 unaltered images (512 × 512 pixels). Eight corrosion categories were aggregated into a single All forms of corrosion class and evaluated against Healthy structures under standardized training conditions on the KAI software platform. The dataset was partitioned to preserve class balance (8000 corrosion and 1000 healthy samples), and geometric augmentations were applied exclusively to the training subset to prevent data leakage. Among the evaluated architectures, ResNet50 achieved near-perfect classification performance (accuracy, precision, recall, and F1-score ≈ 100%), followed by InceptionV3 (F1-score = 99%) and the VGG models (F1-scores = 96–97%). These results demonstrate the reliability of deep learning for binary corrosion detection and establish a controlled benchmark that supports future work on model generalization, cross-dataset validation, and type-specific corrosion identification.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Binary classification of marine corrosion using deep learning: a benchmark study on the full marine corrosion dataset

  • 期刊论文

  • 发表日期：

    2025年10月24日

查看全文
  • Jianxing Yu,
  • Benard Aijuka Mashaija,
  • Erick Phares Massami,
  • Castor Neleson Mwankefu

摘要

Corrosion poses a persistent challenge to marine infrastructure, resulting in structural degradation, increased maintenance costs, and safety risks. Although deep learning has shown promise in automating corrosion detection, comprehensive benchmark studies using large-scale, high-resolution datasets remain scarce. This study presents a binary classification framework employing four state-of-the-art convolutional neural networks (CNNs)—VGG16, VGG19, InceptionV3, and ResNet50—to distinguish between corroded and healthy marine surfaces. Experiments were conducted using the publicly available Marine Corrosion Dataset, consisting of 9,000 unaltered images (512 × 512 pixels). Eight corrosion categories were aggregated into a single All forms of corrosion class and evaluated against Healthy structures under standardized training conditions on the KAI software platform. The dataset was partitioned to preserve class balance (8000 corrosion and 1000 healthy samples), and geometric augmentations were applied exclusively to the training subset to prevent data leakage. Among the evaluated architectures, ResNet50 achieved near-perfect classification performance (accuracy, precision, recall, and F1-score ≈ 100%), followed by InceptionV3 (F1-score = 99%) and the VGG models (F1-scores = 96–97%). These results demonstrate the reliability of deep learning for binary corrosion detection and establish a controlled benchmark that supports future work on model generalization, cross-dataset validation, and type-specific corrosion identification.