图像识别技术在腐蚀防护应用领域的现状
Current Status of Image Recognition Technology in the Field of Corrosion Protection Applications

摘要
Abstract

腐蚀每年给经济带来严重损失。因此，各种腐蚀防护和检测技术在大型金属工程结构的日常维护中得到广泛应用。
Corrosion brings serious losses to the economy annually. Therefore, various corrosion protection and detection techniques are widely used in the daily maintenance of large metal engineering structures.
图像识别技术的出现为无损检测带来了更方便、更快捷的方式。
The emergence of image recognition technology has brought a more convenient and faster way for nondestructive testing.
现有图像识别技术可根据算法分为两类：传统图像识别技术和基于深度学习的图像识别技术。
Existing image recognition technology can be divided into two categories according to the algorithm: traditional image recognition technology and image recognition technology based on deep learning.
这两种技术已广泛应用于金属、涂层和电化学数据图像这三个领域。
These two types of technologies have been widely used in the three fields of metal, coating, and electrochemical data images.
已开展大量工作来识别金属和涂层中的缺陷，基于深度学习的方法在识别电化学数据图像方面也显示出潜力。
A large amount of work has been carried out to identify defects in metals and coatings, and deep learning-based methods also show potential for identifying electrochemical data images.
将电化学图像与缺陷形态检测相匹配，将为金属和涂层的图像识别技术带来更深入的理解。
Matching electrochemical images with the detection of defect morphology will bring a deeper understanding of image recognition techniques for metals and coatings.
一个累积形态和电化学参数的数据库将使得使用图像识别技术预测钢和涂层的寿命成为可能。
A database of accumulated morphology and electrochemical parameters will make it possible to predict the life of steel and coatings using image recognition techniques.

1. 引言
1. Introduction

腐蚀是一种普遍存在的自然现象，因其隐蔽性和破坏性而常被称为“安静的破坏”。2014 年，中国金属腐蚀造成的损失占国内生产总值的 3.34%，约合 3100 亿美元[1]。然而，通过采取适当措施，可以防止 25%至 40%的腐蚀损失[2]。海洋蕴藏着大量的石油、天然气和可燃冰，需要大量的海上设备进行资源开采。海上石油钻井平台[3]、运输船[4、5]和输油管道[6]特别容易受到腐蚀，因为海洋具有高温、高湿度和高盐度的环境[7、8、9]。为了减轻腐蚀，科学家们探索了耐腐蚀材料[10、11]、腐蚀抑制剂[12]、防护涂层[13、14、15]和电化学阴极保护[16、17]。同时，工程师们还采用了无损检测（NDT）技术进行金属腐蚀检测。无损检测技术不会损坏检测对象，广泛用于识别腐蚀缺陷[18、19、20、21]。 常见的无损检测方法包括目视检测、超声检测（UT）[22,23]、脉冲涡流检测（ECT）[24,25,26]和射线检测（RT）[27,28]。
Corrosion, a pervasive natural phenomenon, is often referred to as “quiet destruction” due to its insidious and destructive nature. In 2014, the cost of metal corrosion in China accounted for 3.34% of the gross domestic product, approximately US $310 billion [1]. However, with appropriate measures, 25% to 40% of corrosion losses can be prevented [2]. The ocean, which harbors vast amounts of oil, natural gas, and combustible ice, necessitates extensive offshore equipment for resource extraction. Offshore oil drilling platforms [3], transportation ships [4,5], and oil pipelines [6] are particularly susceptible to corrosion due to the high-temperature, high-humidity, and high-salinity environment of the ocean [7,8,9]. To mitigate corrosion, scientists have explored corrosion-resistant materials [10,11], corrosion inhibitors [12], protective coatings [13,14,15], and electrochemical cathodic protection [16,17]. Concurrently, engineers have employed non-destructive testing (NDT) techniques for metal corrosion detection. NDT techniques, which do not damage the inspection object, are extensively used to identify corrosion defects [18,19,20,21]. Common NDT methods include visual inspection, ultrasonic testing (UT) [22,23], pulsed eddy current testing (ECT) [24,25,26], and radiographic testing (RT) [27,28].

视觉检查是最简单、最方便的检查方法，因此被广泛用于检测船舶、桥梁、管道和其他设施的缺陷[29,30]。工程师通过目视识别裂缝和腐蚀等缺陷。然而，这种方法需要大量的劳动力、材料和资金[31]。此外，识别结果可能因工人而异。人工智能（AI）的出现为腐蚀检测带来了新的方法[32,33,34,35]，包括图像识别作为一种无损检测技术，有可能取代视觉检查[36]。工程师使用无人机和机器人等智能设备拍摄大型钢结构的腐蚀图像，通过计算机视觉实现对腐蚀区域的高效、准确识别。
Visual inspection is the simplest and most convenient inspection method, making it widely used for detecting defects in ships, bridges, pipelines, and other facilities [29,30]. Engineers identify defects such as cracks and corrosion by sight. However, this method demands significant labor, materials, and financial resources [31]. Additionally, the identification results can vary between workers. The advent of Artificial Intelligence (AI) has introduced new approaches to corrosion detection [32,33,34,35], including image recognition as a non-destructive testing technique with the potential to replace visual inspection [36]. Engineers employ intelligent devices like drones and robots to capture corrosion images of large steel structures, enabling the efficient and accurate identification of corroded areas through computer vision.
这种方法不会损坏被检测对象，并且可以初步区分腐蚀的类型和程度。
This method does not damage the inspection object and can preliminarily distinguish the type and degree of corrosion.
重要的是，它的评估更客观，因为图像识别技术不仅可以复制人眼捕捉颜色、形状、纹理和其他特征的能力，还可以对这些数据进行量化，以便根据预设标准做出更准确的判断。
Importantly, its evaluation is more objective, as image recognition technology can not only replicate the human eye in capturing color, shape, texture, and other features, but also quantify these data to make more accurate judgments based on predefined standards.

图像识别技术在腐蚀防护中分为传统图像识别算法和基于深度学习的图像识别技术[37,38,39,40,41,42,43,44]。传统图像识别技术可以准确识别腐蚀区域，已被广泛应用[42,43]。然而，它们依赖于手动设计的特征提取器，使得调整过程复杂，并降低了泛化能力和鲁棒性。
Image recognition techniques in corrosion protection are categorized into traditional image recognition algorithms and deep learning-based image recognition techniques [37,38,39,40,41,42,43,44]. Traditional image recognition techniques can accurately identify corrosion areas and have been widely used [42,43]. However, they rely on manually designed feature extractors, making the tuning process complex and reducing the generalization ability and robustness.
相比之下，深度学习图像识别技术通过从大量数据中提取特征并揭示数据集之间更深层次的关联，更精确地表达数据集[37,38,39,40,41]。这种方法需要大量数据集和强大的计算资源。
By contrast, deep learning image recognition technology more precisely expresses datasets by extracting features from vast amounts of data and uncovering deeper associations between datasets [37,38,39,40,41]. This approach requires large datasets and powerful computing resources.

本文主要总结了腐蚀防护中图像识别技术的三个方面：金属、涂层和电化学成像。
This paper primarily summarizes three aspects of image recognition technology in corrosion protection: metal, coating, and electrochemical imaging.
金属和涂层的识别侧重于其形态图像特征，以区分腐蚀类型和特征。
The recognition of metals and coatings focuses on their morphological image features to distinguish between corrosion types and characteristics.
电化学数据图像的识别涉及到理解腐蚀机制，并基于数据图像进行预测。图 1 提供了图像识别技术在腐蚀防护中的应用概述。
The recognition of electrochemical data images involves understanding the corrosion mechanism and making predictions based on data images. Figure 1 provides an overview of the application of image recognition technology in corrosion protection.

图 1. 图像识别技术在腐蚀防护领域的应用[24,45,46,47,48,49]。
Figure 1. Application of image recognition technology in the field of corrosion protection [24,45,46,47,48,49].

2. 图像识别技术的一般步骤
2. General Steps in Image Recognition Techniques

传统的图像识别方法通常包括以下步骤：图像采集、图像预处理、特征提取和图像分类。基于深度学习的图像识别算法则有很大的不同。
Traditional image recognition methods typically involve the following steps: image acquisition, image preprocessing, feature extraction, and image classification. Deep learning-based image recognition algorithms differ significantly.
它们需要大量的标注数据，并且它们的算法架构更加复杂。图 2 展示了这两种方法的流程图。
They require a large amount of labeled data, and their algorithmic architecture is more complex. Figure 2 illustrates the flowcharts of both methods.

图 2. 传统图像识别技术和基于深度学习的图像识别技术的流程图。
Figure 2. Flowcharts of traditional image recognition techniques and deep learning-based image recognition techniques.

2.1. 传统图像识别步骤
2.1. Traditional Image Recognition Steps

2.1.1. 图像采集集合
2.1.1. Image Acquisition Collection

图像识别技术中使用的三种主要捕获设备是：摄像机、数码相机和扫描仪器（如扫描电子显微镜和扫描声学显微镜）。
Three main types of capture devices are used in image recognition technology: video cameras, digital cameras, and scanning instruments (such as scanning electron microscopes and scanning acoustic microscopes).
这些设备可以通过采样来提取数字图像（包括动态图像），并将图片的颜色、形状、纹理和外部声音存储在计算机设备中。
These devices can extract digital images (including moving images) through sampling, and storing the color, shape, texture, and external sounds of the pictures in computer equipment.
图像采集是图像识别的第一步，采集到的图像质量会影响识别精度。
Image acquisition is the first step in image recognition, and the quality of the acquired image affects recognition accuracy.

图像的质量通常与捕获设备的像素、分辨率和 DPI（每英寸点数）有关。像素是最小的单位，只能显示一种颜色。每个像素越大，每个单位面积可以显示的像素越少。
The quality of an image is generally related to the pixels, resolution, and DPI (dots per inch) of the capture device. Pixels, the smallest units, can only display one color. The larger each pixel, the fewer pixels can be displayed per unit area.
分辨率是指图像中包含的像素数量，而 DPI 则表示沿图像长度每英寸（1 英寸=25.4 毫米）包含的像素数量。因此，较高的分辨率和 DPI 通常会导致更高的图像质量，提供更多的信息。
Resolution refers to the number of pixels contained in an image, while DPI measures the number of pixels per inch (1 inch = 25.4 mm) along the length of the image. Therefore, a higher resolution and DPI typically result in higher image quality, providing more information.

除了图像质量和清晰度外，图像类型也会影响识别方法的选择。腐蚀识别常用的图像类型包括彩色图像、灰度图像和黑白图像。
In addition to image quality and clarity, the type of image also affects the choice of recognition method. Common image types for corrosion recognition include color images, grayscale images, and black-and-white images.
彩色图像是最常用的，可以应用于传统和深度学习图像识别算法中。灰度图像和黑白图像通常是从彩色图像中派生出来的，尽管一些扫描仪器，如扫描电子显微镜，可以直接捕获灰度图像。
Color images are the most commonly used and can be employed in both traditional and deep learning image recognition algorithms. Grayscale and black-and-white images are often derived from color images, although some scanning instruments, such as scanning electron microscopes, can directly capture grayscale images.

2.1.2. 图像预处理
2.1.2. Image Pre-Processing

图像预处理旨在分割图片中的对象，消除不必要的信息，并专注于研究所需的数据。这一步在特征提取之前至关重要。
Image preprocessing aims to segment objects in the picture, eliminating unnecessary information and focusing on the data required for research. This step is crucial before feature extraction.
该任务可以使用 Matlab R2024a 或 OpenCV 4.0 来完成，这两个工具都提供了专门为图像识别设计的各种功能工具包。
The task can be accomplished by utilizing either Matlab R2024a or OpenCV 4.0, both of which offer a variety of functional toolkits specifically designed for image recognition.
常见的预处理方法包括图像二值化、降噪、增强和几何变换。
Common preprocessing methods include image binarization, noise reduction, enhancement, and geometric transformation.
图像二值化对于传统的图像识别算法至关重要，它使用特定的阈值将图像的灰度值转换为二进制（0 和 255），便于提取关键信息。
Image binarization is essential for traditional image recognition algorithms, converting the grayscale values of an image into binary (0 and 255) using a specific threshold, facilitating key information extraction.
降噪可最大程度减少干扰，而增强则突出关键特征以提高图像质量。降噪通常通过滤波方法实现，包括平均、中值、高通和低通滤波。
Noise reduction minimizes interference, while enhancement highlights critical features to improve image quality. Noise reduction is typically achieved through filtering methods, including average, median, high-pass, and low-pass filtering.
图像增强技术多种多样，例如对比度调整、锐化、颜色调整、腐蚀和膨胀。几何变换，如翻转、平移、旋转和缩放，不会改变图像信息，但有助于计算机更好地理解图片。
Image enhancement techniques vary widely, such as contrast adjustment, sharpening, color adjustment, erosion, and dilation. Geometric transformations, like flipping, panning, rotating, and scaling, do not alter image information but help computers better understand the picture.
这些变换还增加了图像数据量，以便更有效地进行深度学习训练。
These transformations also increase the amount of image data for more effective deep learning training.

2.1.3. 特征提取
2.1.3. Feature Extraction

计算机将图像处理为数字，因此特征提取涉及通过特定算法获取这些数字或向量数据的信息。可以提取颜色、亮度、边缘和纹理等特征。
Computers process images as numbers, so feature extraction involves obtaining information about these numerical or vector data through specific algorithms. Features such as color, brightness, edges, and texture can be extracted.
各种图像特征，包括边缘、角、脊和区域，用于识别不同类型的腐蚀。特征提取方法根据处理方法进行分类，包括基于纹理的方法[50,51]、基于颜色的方法[52]、基于边缘的方法[44]、基于灰度的方法[45]和基于模型变换的方法[23,53]。
Various image features, including edges, corners, ridges, and regions, are used to recognize different types of corrosion. Feature extraction methods are categorized based on the processing approach, including texture-based methods [50,51], color-based methods [52], edge-based methods [44], grayscale-based methods [45], and model transformation-based methods [23,53].

2.1.4. 图像分类
2.1.4. Image Classification

图像分类的核心任务是为图片分配相应的标签。处理后，通过分类器使用标签对图像特征数据或向量进行分类。
The core task of image classification is to assign a corresponding label to the picture. Once processed, the image feature data or vectors are classified using labels through a classifier.
传统图像识别算法通常使用分类器，如支持向量机（SVM）和随机森林（RF）[46]。SVM 分类器可以进行二分类，是一种用于确定区域是否存在腐蚀等任务的线性判断方法。
Traditional image recognition algorithms commonly use classifiers such as Support Vector Machines (SVM) and Random Forests (RF) [46]. The SVM classifier can perform binary classification and is a linear judgment method used for tasks such as determining whether there is corrosion in an area.
随机森林是一种更复杂的判别系统，可以通过判断多个决策树预测的结果来判断非常复杂的问题。
Random forests is a more complex discrimination system, which can judge very complex problems by judging the results predicted by multiple decision trees.
作为应用最广泛的分类器，SVM 特别适合于中小规模复杂数据集的分类，因此其效果优于传统算法。
As the most widely used classifier, SVM is especially suitable for the classification of small- and medium-sized complex datasets, so its effect is better than traditional algorithms.

2.2. 基于深度学习的图像识别
2.2. Deep Learning-Based Image Recognition

基于深度学习的图像识别涉及四个主要步骤：图像采集、预处理、模型训练和应用。它利用卷积神经网络（CNN）从图像中提取高级特征，需要具有准确标注图像的大型数据集。
Deep learning-based image recognition involves four main steps: image acquisition, preprocessing, model training, and application. It utilizes convolutional neural networks (CNNs) to extract high-level features from images, requiring large datasets with accurately labeled images.
深度学习的常见分类架构包括 AlexNet、VGG 和 ResNet。深度学习可以解决简单的腐蚀问题，例如钢板和涂层的缺陷，以及腐蚀数据的复杂图像分析[47,54]。
The common classification architectures for deep learning include AlexNet, VGG, and ResNet. Deep learning can address both simple corrosion problems, such as defects in steel plates and coatings, and the complex image analysis of corrosion data [47,54].

3. 金属的鉴定
3. Identification of Metals

3.1. 传统方法
3.1. Traditional Methods

传统图像处理算法严重依赖于人工设计的特征提取器，这些特征提取器可以满足各种需求并具有成熟的应用技术，但它们也容易受到干扰和噪声的影响[55]。相比之下，深度学习可以在训练过程中自动学习图像特征，而无需人工特征设计，具有更强的泛化能力。
Traditional image processing algorithms rely heavily on manually designed feature extractors, which can meet various needs and have mature application techniques, but they are also susceptible to interference and noise [55]. By contrast, deep learning can automatically learn image features during the training process without manual feature design, offering stronger generalization capabilities.
常见的传统方法包括基于小波变换的方法、基于分形理论的方法和基于灰度图像的方法。表 1 总结了金属识别的传统算法，而表 2 则列出了它们的优缺点。
Common traditional methods include wavelet variation-based methods, fractal theory-based methods, and grayscale image-based methods. Table 1 summarizes traditional algorithms for metal recognition, while Table 2 outlines their advantages and disadvantages.

表 1. 基于传统算法的图像识别总结。
Table 1. Summary of image recognition based on conventional algorithms.

方法 Method	材料 Materials	主要内容 Main Content	作者/参考文献 Author/Ref.	年 Year
基于小波变换的方法 Method based on wavelet transform	7075-T76 铝合金 Al 7075-T76	对电化学噪声信号进行分析，以详细揭示电位信号的特征，并监测局部腐蚀的发生 The electrochemical noise signal was analyzed to reveal the characteristics of the potential signal in detail and monitor the occurrence of local corrosion	刘[23] Liu [23]	2001
	航空铝合金 Aero-aluminum alloy	腐蚀坑的形成和生长进行了建模和模拟。利用基于小波的图像处理对腐蚀板进行了评估。 The formation and growth of corrosion pits were modeled and simulated. The corroded panels were evaluated by wavelet-based image processing	皮达帕蒂[53] Pidaparti [53]	2007
	耐候钢 Weathering steel	小波变换与 PSO-SVM 技术相结合对耐候钢的腐蚀状态进行评估 The corrosion state of weathering-resistant steel was evaluated by combining wavelet transform with the PSO-SVM technique	严[50] Yan [50]	2014
基于分形理论的方法 Methods based on fractal theory	因科镍合金 600 Inconel alloy 600	蚀坑的分形维数几乎不随温度变化 The fractal dimension of pitting pits hardly changes with temperature	公园[56] Park [56]	2003
	钢筋混凝土（RC） Reinforced Concrete (RC)	建立了裂缝分形维数、腐蚀速率与钢筋直径之间的关系，以评估退化钢筋混凝土结构中钢筋的腐蚀行为 The relationship between the crack fractal dimension, corrosion rate and steel bar diameter was established to evaluate the corrosion behavior of steel bars in degraded reinforced concrete structures	李[57] Li [57]	2022
	X80 钢 X80 steel	腐蚀坑的二维/三维分形维数随交流密度的增加而增加，分别呈线性和指数关系 The two-dimensional/three-dimensional fractal dimension of corrosion pit increased with the increase of AC density, showing linear and exponential relationships, respectively	傅 Fu [58]	2019
基于灰度的方法 Method based on grayscale	镀锌高强度钢 Galvanized high strength steel	腐蚀坑的二值图像的椭圆参数被提取出来，并且发现腐蚀坑的椭圆参数在所有方向上都满足高斯分布 The elliptic parameters of the binary image of the corrosion pit were extracted, and it was found that the elliptic parameters of the corrosion pit satisfy the Gaussian distribution along all directions	徐[59] Xu [59]	2016

表 2. 基于传统算法的图像识别的优缺点。
Table 2. Advantages and disadvantages of image recognition based on traditional algorithms.

方法 Method	优势 Advantage	劣势 Disadvantage
基于小波变化的方法 Methods based on wavelet variations	低频区域响应图像的颜色空间分布和亮度差异，而高频区域能够响应腐蚀的局部特征。 The low-frequency region responds to the color spatial distribution and brightness differences of the image, while the high-frequency region is able to respond to the local features of corrosion.	小波能量摘计算方法的不同可能会导致腐蚀图像处理结果的差异 Differences in the calculation of wavelet energy entropy may lead to differences in corrosion image processing
基于类型论的方法 Approach based on typing theory	它具有识别图像中非欧几里得几何特征的能力，并能够区分图像信息中粗糙度的差异。 It has the ability to recognize features of non-traditional Euclidean geometry in an image and to distinguish differences in roughness in image information.	分形维数可能对于不同的图像是相同的，不能单独用作判断图像的标准 The fractal dimension may be the same for different images and cannot be used alone as a criterion for judging images
基于灰度图像的方法 Gray scale image based approach	它具有降低复杂性、提高计算速度和提高细节清晰度的能力。 It has the ability to reduce complexity, increase computing speed, and improve detail clarity.	信息不完整表示和无法处理具有颜色特征的腐蚀 Incomplete representation of information and inability to process corrosion with color characteristics

3.1.1. 基于小波变换的方法
3.1.1. Wavelet Transform-Based Approach

小波变换是一种图像处理算法，通过将图像的时域转换为频域，将图像分解为多个子图像，从而实现更好的特征提取。电化学噪声信号可以通过金属腐蚀纹理[50,51]、小波变换[23]、腐蚀坑[53]等进行研究。Yan[50]使用腐蚀纹理图像的小波变化评估了风化钢桥梁的腐蚀状态。Jahanshahi[51]评估了颜色空间、颜色通道和子图像块大小等参数对基于颜色小波的纹理分析算法检测腐蚀特性的影响。Liu[23]使用小波变换将电化学噪声信号分解为时间-频率域中的一系列子信号，以研究局部腐蚀。Pidaparti[53]使用小波变换处理评估航空航天材料的腐蚀损伤，使用人工神经网络（ANNs）研究材料损失和剩余强度预测。
Wavelet transform, an image processing algorithm, decomposes an image into multiple sub-images by transforming the image’s time domain into the frequency domain, enabling better feature extraction. Electrochemical noise signals can be studied using metal corrosion textures [50,51], wavelet transforms [23], corrosion pits [53], and more. Yan [50] assessed the corrosion state of weathering steel bridges using wavelet variations on corrosion texture images. Jahanshahi [51] evaluated the effect of parameters such as the color space, color channel, and sub-image block size on the color wavelet-based texture analysis algorithm to detect corrosion properties. Liu [23] used wavelet transform to decompose electrochemical noise signals into a series of sub-signals in the time–frequency domain to study localized corrosion. Pidaparti [53] evaluated the corrosion damage of aerospace materials using wavelet transform processing, investigating material loss and residual strength prediction with Artificial Neural Networks (ANNs).
基于此，还完成了基于 ANN 的腐蚀疲劳条件下的寿命预测。
Based on this, ANN-based life prediction under corrosion fatigue conditions was also accomplished.

小波变换可以有效地提取腐蚀图像时频图的特征。
Wavelet transform can effectively extract the features of time–frequency maps of corrosion images.
低频区域对应于图像的颜色空间分布和亮度差异，而高频区域对应于腐蚀的局部特征。
The low-frequency region corresponds to the color spatial distribution and brightness differences of the image, while the high-frequency region responds to the local features of corrosion.
然而，小波能量熵的计算差异可能会导致腐蚀图像处理的变化。此外，小波方法获取腐蚀图像特征往往受到数据集大小的影响。
However, differences in the calculation of wavelet energy entropy may lead to variations in corrosion image processing. Additionally, the acquisition of corrosion image features by wavelet methods is often affected by the dataset size.

3.1.2. 基于分形理论的方法
3.1.2. Approaches Based on Fractal Theory

腐蚀图像中的结构特征通常不符合传统的欧几里得几何，分形理论可以有效地描述这些不规则形状。因此，可以使用分形理论提取每个局部腐蚀区域的特征。
The structural features in corrosion images often do not conform to traditional Euclidean geometry, and fractal theory can describe these irregular shapes effectively. Therefore, the features of each localized corrosion region can be extracted using fractal theory.
通过研究腐蚀图像的分形维数，研究人员可以分析钢的腐蚀发展模式、腐蚀坑的生长变化以及钢筋混凝土的腐蚀速率[56,57,58]。Park[56]使用分形理论研究了腐蚀坑的 SEM（扫描电子显微镜）图像的分形维数，发现分形维数随溶液温度的升高而增加。
By studying the fractal dimension of corrosion images, researchers can analyze the corrosion development pattern of steel, the growth change of corrosion pits, and the corrosion rate of reinforced concrete [56,57,58]. Park [56] studied the fractal dimension of SEM (scanning electron microscope) images of corrosion pits using fractal theory and found that the fractal dimension increases with increasing solution temperatures.
腐蚀坑在不同溶液温度下的生长满足一定的分形几何特征。李[57]使用钢筋混凝土表面裂缝形状的分形维数来研究裂缝纹理与钢筋混凝土棱柱体的腐蚀速率和直径尺寸之间的关系。他发现分形维数与腐蚀速率和钢筋直径密切相关。
The growth of corrosion pits at different solution temperatures satisfies certain fractal geometric features. Li [57] used the fractal dimension of the cracked shape of reinforced concrete surfaces to study the relationship between the cracked texture and the corrosion rate and diameter size of reinforced concrete prisms. He found that the fractal dimension was closely related to the corrosion rate and the diameter of the reinforcement.
傅[58]研究了不同交流电流密度对 X80 钢在沿海土壤溶液中腐蚀形貌发展的影响。
Fu [58] studied the effect of different alternating current densities on the development of the corrosion morphology of X80 steel in a coastal soil solution.
他发现，在低交流电流密度下，X80 钢主要是均匀腐蚀；当交流电流密度达到 150 A/m2 时，腐蚀形态逐渐转变为不规则点蚀腐蚀，二维/三维分形维数分别随交流电流密度的增加呈线性和指数增长。
He found that at low alternating current densities, X80 steel was mainly corroded uniformly; when the alternating current density reached 150 A/m², the corrosion morphology gradually transformed into irregular pitting corrosion, and the two-dimensional/three-dimensional fractal dimensions increased linearly and exponentially, respectively, with the increase in alternating current density.
分形理论可以识别图像中非传统欧几里得几何的特征，并检测图像信息中的粗糙度差异。
Fractal theory can identify the features of non-traditional Euclidean geometry in images and detect roughness differences in image information.
然而，它通常不能单独用作判断图像的标准，因为不同的图像可能具有相同的分形维数。
However, it often cannot be used alone as a criterion for judging images, because the fractal dimension may be the same for different images.

3.1.3. 基于灰度图像的方法
3.1.3. Methods Based on Grayscale Images

基于灰度的图像处理也是一种常见的方法，利用图像的灰度值差异来确定最佳灰度阈值并创建二值图像。Zhu 等人[24]使用扫描声学显微镜（SAM）在 C 模式下结合层析声学微成像（TAMI）来确定 7050 铝合金腐蚀坑的形貌和深度。
Grayscale-based image processing is also a common method, utilizing differences in the gray values of images to determine the optimal gray threshold and create a binary image. Zhu et al. [24] used scanning acoustic microscopy (SAM) in C-mode with tomographic acoustic microimaging (TAMI) to determine the morphology and depth of corrosion pits on 7050 aluminum alloy.
他们使用光学显微镜检查结果，并使用二值图像计算点蚀面积。图 3 显示了二值计算的结果。此外，他们还绘制了去除腐蚀产物后的铝合金样品的 3D 形貌（图 4），以直观地观察腐蚀深度分布的差异。通过二值图像可以清晰地提取缺陷区域（如裂缝）的面积份额。徐等人[59]提出了一种基于图像处理的方法，通过对蚀坑深度的图像进行二值化，并提取腐蚀椭圆的参数，对高强度钢丝表面的蚀坑进行可视化建模。
They examined the results using an optical microscope and calculated the pitting area using a binary image. Figure 3 shows the results of the binary calculation. Additionally, they plotted the 3D morphology of an aluminum alloy sample after removing the corrosion products (Figure 4) to visualize the difference in the distribution of the corrosion depth. The area share of defective areas (such as cracks) can be clearly extracted using the binarized image. Xu et al. [59] proposed an image processing-based method for the visual modeling of pits on the surface of high-strength steel wires by binarizing the image of the pit depth and extracting the parameters of corrosion ellipses.
他们通过统计发现，腐蚀表面坑的长轴方向与圆周方向一致，平面几何特征满足高斯分布。钱等。[60]研究了 AerMet100 钢在不同腐蚀阶段的腐蚀形貌图像差异。他们使用中值滤波、灰度变化和模糊增强对原始图像进行了预处理。
They statistically found that the direction of the long axis of the pits on the corroded surface is consistent with the circumferential direction and that the planar geometric features satisfy the Gaussian distribution. Qian et al. [60] investigated the image differences in corrosion morphology of AerMet100 Steel at different corrosion stages. They preprocessed the original images using median filtering, grayscale variation, and fuzzy enhancement.
在确定图像二值化的最优阈值后，他们基于灰度阈值分割理论计算方法提取图像特征来计算腐蚀程度。
After determining the optimal threshold for image binarization, they extracted image features to calculate the degree of corrosion based on the separation theory grayscale threshold calculation method.
发现 AerMet100 钢（一种高强度马氏体合金钢）在加速试验中最初表现出点蚀腐蚀，随后逐渐发展为均匀腐蚀。
It was found that AerMet100 Steel (a high-strength martensitic alloy steel) initially showed pitting corrosion in the accelerated test, which then gradually developed into uniform corrosion.

图 3. TAMI 扫描下 7050 铝合金 SAM 图像的二值图像处理[24]（将亮斑的所有灰度值设置为 0，将其他点的所有灰度值设置为 1）；（a-g）分别为深度为 51.0、76.5、102.0、127.5、153.0、178.5 和 204.0 μm 的分层图像。
Figure 3. Binary image processing of SAM image of 7050 aluminum alloy under TAMI scanning [24] (setting all the gray values of the bright spots to 0 and all the gray values of the other points to 1); (a–g) are the layered images with depths of 51.0, 76.5, 102.0, 127.5, 153.0, 178.5, and 204.0 μm, respectively.

图 4. 去除腐蚀产物后的 7050 铝合金样品的 3D 图像[24]。(a)左上角区域；(b)右下角区域。
Figure 4. 3D image of 7050 aluminum alloy sample after removal of corrosion products [24]. (a) Upper-left region; (b) lower-right region.

3.2. 基于深度学习的方法
3.2. Deep Learning-Based Approach

金属图像识别中使用的算法通常是卷积神经网络（CNN）的增强版本。这些优化模型能够有效地识别和检测金属表面的裂纹缺陷。例如，Zhao [61] 通过重新配置网络结构并引入多尺度融合和可变形卷积网络来改进 Faster R-CNN 算法。
Algorithms used for metal image recognition are often enhanced versions of Convolutional Neural Networks (CNNs). These optimized models effectively identify and detect crack defects on metal surfaces. For instance, Zhao [61] improved the Faster R-CNN algorithm by reconfiguring the network structure and introducing multi-scale fusion and deformable convolutional networks.
这种增强将钢表面缺陷的检测准确率提高到平均 0.752，比原始算法高 0.128。同样，肖等人[54]通过将 Transformer 模型与 YOLO-v5 骨干特征提取网络相结合，检测镀锌钢表面的锌花缺陷。图 5a-c 展示了这种混合算法 YOLOv5-TB 的架构、识别性能和结果。
This enhancement increased the detection accuracy of steel surface defects to an average of 0.752, which is 0.128 higher than the original algorithm. Similarly, Xiao et al. [54] detected zinc bloom defects on galvanized steel surfaces by combining the Transformer model with the YOLO-v5 backbone feature extraction network. Figure 5a–c illustrate the architecture, recognition performance, and results of this hybrid algorithm, YOLOv5-TB.
通过使用加权双向特征金字塔网络（Bi-FPN）进行多尺度特征融合，YOLOv5-TB 模型在检测精度和效率方面超过了大多数现有的主流目标检测算法。
By employing a weighted bidirectional feature pyramid network (Bi-FPN) for multi-scale feature fusion, the YOLOv5-TB model surpasses most existing mainstream target-detection algorithms in both detection accuracy and efficiency.

图 5.（a）YOLOv5-TB 算法架构；（b）YOLOv5-TB 算法缺陷的实验数据；（c）该算法对角缺陷检测的有效性[54]。
Figure 5. (a) YOLOv5-TB algorithm architecture; (b) experimental data of YOLOv5-TB algorithm defects; (c) effectiveness of this algorithm for detecting diagonal defects [54].

4. 腐蚀数据图像的识别与挖掘
4. Recognition and Mining of Corrosion Data Images

涂层的腐蚀检测，是一种受多种因素影响的现象，可以通过多种方法来研究。
Corrosion detection of coatings, a phenomenon influenced by various factors, can be approached through multiple methods.
电化学阻抗谱（EIS）是一种理想的无损检测方法，可捕获涂层损伤引起的电化学反应信息。
Electrochemical Impedance Spectroscopy (EIS) is an ideal non-destructive testing method that captures information about the electrochemical reactions caused by coating damage.
模态值的大小反映了涂层的防护性能，并支持涂层寿命预测。机器学习方法可以有效地分析 EIS 谱。Ma 等人[46]提出了一种使用原位 EIS 数据预测多层 Cr/GLC 涂层寿命的方法，结合了一个将涂层阻抗与寿命联系起来的已建立方程。
The magnitude of the modal value reflects the protective performance of the coating and supports coating life prediction. Machine learning methods can analyze EIS spectra effectively. Ma et al. [46] proposed a method for predicting the lifetime of multilayer Cr/GLC coatings using in situ EIS data, combined with an established equation linking the coating impedance and lifetime.
他们改进的机械-经验模型表明，低频阻抗和多层 Cr/GLC 涂层在不同静水压力下的暴露时间遵循线性经验关系。
Their modified mechanistic–empirical model demonstrated that low-frequency impedance and the exposure time of multilayer Cr/GLC coatings at different hydrostatic pressures follow a linear empirical relationship.
这种机械-经验方法和机械学习模型的组合方法有效地区分了各种类型的电化学阻抗谱，将涂层性能与涂层寿命相关联。该模型的总体结构如图 6 所示。对多个模型的比较分析表明，ANN + RF 集成机器学习模型对涂层性能和寿命的综合预测准确率最高，达到 97.9%，如图 7 所示。
This combined approach of mechanistic–empirical methods and mechanistic-learning models effectively differentiates various types of electrochemical impedance spectra, correlating the coating performance with coating life. The overall structure of this model is shown in Figure 6. A comparative analysis of multiple models revealed that the ANN + RF integrated machine learning model achieved the highest combined prediction accuracy of 97.9% for coating performance and lifetime, as illustrated in Figure 7.

图 6.（a）合成机制的经验方法和机器学习模型的过程示意图；（b）BP-ANN 算法的示意图；（c）RF 算法的示意图；（d）柱塞子表面下的多层 Cr/GLC 涂层；（e）多层 Cr/GLC 涂层的表面形貌；和（f）[46]的横截面形貌。
Figure 6. (a) Schematic of the process of synthesizing the empirical method of the mechanism and a machine learning model; (b) schematic of the BP-ANN algorithm; (c) schematic of the RF algorithm; (d) multi-layer Cr/GLC coatings on the sub-surface of a plunger; (e) the surface morphology of the multilayered Cr/GLC coatings; and (f) cross-sectional morphology [46].

图 7. 机械模型、ANN + 机械模型、RF + 机械模型和 ANN + RF + 机械模型的经验模型的评估指标 R2 [46]。0.1 MPa、5 MPa、10 MPa 和 15 MPa 表示不同的“静水压力条件”。R2 代表四个模型的性能指标，接近 1 的值表示性能更好。
Figure 7. Evaluation metrics for the empirical model of the mechanistic model, ANN + mechanistic model, RF + mechanistic model, and ANN + RF + mechanistic model R² [46]. 0.1 MPa, 5 MPa, 10 MPa and 15 MPa indicate that they are different “hydrostatic pressure conditions”. R² represents the performance indicators of the four models, with a value close to 1 indicating better performance.

EIS 分析具有挑战性，因为不同的腐蚀类型对应不同的等效电路。它需要深入了解腐蚀反应，以构建适当的等效电路图，并使用计算软件拟合相应的参数。
EIS analysis is challenging because different corrosion types correspond to different equivalent circuits. It requires a deep understanding of corrosion reactions to build appropriate equivalent circuit diagrams and fit corresponding parameters using computational software.
一些研究尝试使用机器学习模型分析电化学阻抗谱，以识别其相应的等效电路。 Bongiorno 等人[47]研究了数据集大小对机器学习模型的解析性能的影响，考虑了分类和拟合情况。
Some studies have attempted to analyze electrochemical impedance spectra with machine learning models to identify their corresponding equivalent circuits. Bongiorno et al. [47] investigated the effect of the dataset size on the parsing performance of machine learning models, considering classification and fitting cases.
训练和测试数据集使用 Labview™编程语言进行数值模拟，并使用内部开发的程序生成了大量的 EIS 光谱。
The training and test datasets were numerically simulated using the Labview™ programming language and a large number of EIS spectra were generated using an in-house developed program.
首先确定一个等效电路，然后给出拟合参数值的范围，随机选择数据并分配给电路元件，然后模拟其等效电路。
First, an equivalent circuit was determined, then the range of fitting parameter values was given, the data were randomly selected and allocated to the circuit components, and then their equivalent circuit was simulated.
他们发现，一个包含 200 个样本的数据集足以训练具有多达五个等效电路元件的模型。在用于拟合时，该模型的准确率达到 95%，且具有较高的召回率。
They found that a dataset of 200 samples is sufficient for training models with up to five equivalent circuit components. When used for fitting, the model achieved 95% accuracy with high recall.
用于分类时，准确率约为 75%，但不同电路的召回率差异较大。
For classification, the accuracy was around 75%, though recall varied significantly across different circuits.

5. 涂层的识别
5. Identification of Coatings

5.1. 传统方法
5.1. Traditional Methods

利用计算机视觉系统从涂层状况的图像中提取特定特征，使工程师能够更快速地评估涂层腐蚀的程度。
Using a computer vision system to extract specific characteristics from images of coating conditions allows engineers to assess the extent of coating corrosion more rapidly.
该系统可以根据锈蚀程度分析导致腐蚀的主要因素，并确定是否需要重新涂覆。Momber[52]利用基于颜色的数字图像处理技术来评估涂层恶化的程度。图 8a 展示了海上风能设备上一些涂层区域的图像识别结果，而图 8b 显示了涂层的颜色变化，表明设备涂层恶化。
This system can analyze the primary factors causing corrosion based on the degree of rusting and determine if recoating is necessary. Momber [52] utilized color-based digital image processing techniques to evaluate the degree of coating deterioration. Figure 8a illustrates the image recognition results for some coating areas on an offshore wind energy device, while Figure 8b shows the color changes in the coating, indicating equipment coating deterioration.
在该图中，标签 1 和 2 表示一种新涂层的颜色分布，而标签 3 和 4 超出了可接受范围，表明该涂层不符合规格。图 8c 展示了腐蚀区域的 HSV（色调-饱和度-值）直方图，这些特征的分布差异区分了重锈和轻锈。对每种图像类型的损坏原因进行分析表明，大多数涂层损坏是由于结构设计不当和机械载荷造成的。
In this figure, labels 1 and 2 represent the color distribution of a new coating, whereas labels 3 and 4 fall outside the acceptable range, signaling that the coating does not meet the specifications. Figure 8c presents the HSV (Hue–Saturation–Value) histogram of the corroded area, where the distribution differences in these features distinguish between heavy and light rust. The analysis of the damage causes, corresponding to each image type, revealed that most coating damage is due to inappropriate structural design and mechanical loading.

图 8. 涂层评估对海上风能安装（OWEA）的有效性[52]。（a）使用数字图像处理技术评估涂层状况的示例；图片下方的数字表示涂层恶化程度。（b）涂层颜色图表。（c）两种钢腐蚀阶段的 HSV 直方图。
Figure 8. Effectiveness of coating evaluation on Offshore Wind Energy Installation (OWEA) [52]. (a) Example of evaluating the coating condition with digital image processing techniques; the numbers below the picture indicate the degree of coating deterioration. (b) Coating color chart. (c) Histogram of HSV for two steel corrosion stages.

监测涂层质量是另一个关键应用。卢等人[44]开发了一种混合算法框架，能够有效地识别三种不同的热障涂层（TBC）图像。
Monitoring coating quality is another critical application. Lu et al. [44] developed a hybrid algorithmic framework that effectively recognizes three different thermal barrier coating (TBC) images.
通过将图像分析技术与统计方法相结合，他们的模型能够准确识别 TCL 层的上下边界，并为 TBC 图像生成掩码。该方法在识别这些边界方面的分类准确率达到了 98%。
By integrating image analysis techniques with statistical methods, their model accurately identifies the upper and lower boundaries of the TCL layer and generates masks for TBC images. This method achieved a classification accuracy of 98% for identifying these boundaries.
准确区分涂层边界有助于计算涂层的孔隙率。与他们之前基于自适应局部阈值的孔隙率测量方法[62]相比，新算法扩展了 TBC 物种识别的范围。
Accurately differentiating the coating boundary helps calculate the coating’s porosity. Compared to their previous porosity measurement method based on adaptive local thresholding [62], the new algorithm extends the range of TBC species recognized.

检测涂层表面的损伤形态也具有重要意义。起泡是一种常见的涂层缺陷，主要通过目视检查来检测，这是一项劳动密集型且耗时的工作。
Detecting the damage morphology on coating surfaces is also significant for evaluation. Blistering, a common coating defect, is primarily detected through visual inspection, which is labor-intensive and time-consuming.
机器学算法使计算机能够有效地检测破损的涂层层。ISO 4628-2[63]提供了使用图像方法对起泡大小和频率进行分级的标准图像，为使用人工智能监测涂层表面老化提供了基础。Nadia Moradi 等人[45]通过分析起泡区域反射光的差异，创建了起泡涂层像素梯度变化的三维图，如图 9a 所示。他们的方法分别检测大泡和小泡，以准确评估各种泡大小。然后，计算整个图像中凸起的频率，并将其与标准图像频率进行比较，以确定其评级。图 9b 显示了标准图像的计算频率，图 9c 显示了使用该方法的涂层评级结果与手动评级的比较。该算法对直径大于 5 毫米的泡的识别准确率达到 95%，超过了手动视觉监测方法。
Machine learning algorithms enable computers to detect broken coating layers efficiently. ISO 4628-2 [63] provides standard images for grading the size and frequency of blistering using image methods, forming a basis for monitoring coating surface aging with artificial intelligence. Nadia Moradi et al. [45] created a three-dimensional graph of pixel gradient variations in blistered coatings by analyzing the reflected light differences from blistered areas, as shown in Figure 9a. Their method separately detects large and small blisters to accurately assess various blister sizes. Subsequently, the frequency of bulges in the entire image is calculated and compared with standard image frequencies to determine its rating. Figure 9b displays the calculated frequency for the standard image, and Figure 9c shows the coating rating results using this method compared with manual ratings. The algorithm achieves a 95% identification accuracy for blisters larger than 5 mm in diameter, surpassing manual visual monitoring methods.
它能有效地识别白色和浅灰色涂层起泡，但对深灰色和黑色涂层的效果较差。
It effectively identifies white and light gray coating blisters but is less effective for dark gray and black coatings.

图 9. 涂层气泡的识别[45]。(a) 气泡图像及其梯度模值的 3D 图。由于不同尺寸气泡反射光的强度不同，因此可以根据其灰度梯度变化绘制图像。(b) 算法用于根据 ISO 4628 对标准化图像进行分级的频率准则值。它们是通过计算标准图像中不同尺寸气泡的比例得出的，该标准图像是在[63]中提出的，占图像面积的比例。(c) 检查员尺寸评估的平均值与样本尺寸评估算法的比较。误差条表示检查员结果的标准偏差。
Figure 9. Identification of coating bubbles [45]. (a) 3D graph of the blister image and its gradient magnitude. Because of the differences in the intensity of the reflected light from the bubbles of different sizes, the image can be drawn based on their grayscale gradient changes. (b) Frequency criterion values used by the algorithm for grading standardized images according to ISO 4628. They are found by calculating the proportion of differently sized bubbles in the standard image proposed in [63] to the image area. (c) Comparison of the mean value of the inspector size assessment with the sample size assessment algorithm. Error bars indicate the standard deviation of the inspector results.

图像识别还可以同时检测钢板的涂层缺陷和腐蚀。陈柏翰等人[64]实施了图像识别技术来评估钢桥涂层的有效性。他们使用多分辨率模式分类（MPC）来识别铁锈并计算缺陷百分比，以确定涂层质量是否符合验收标准。
Image recognition can also simultaneously detect coating defects and corrosion in steel plates. Po-Han Chen et al. [64] implemented image recognition technology to evaluate the effectiveness of steel bridge coatings. They used multi-resolution pattern classification (MPC) to identify rust and calculate defect percentages, determining whether the coating quality meets acceptance criteria.
这种方法提供了一种更准确的检测手段。
This method offers a more accurate means of detection.

5.2. 深度学习图像识别算法
5.2. Deep Learning Image Recognition Algorithm

基于深度学习的方法需要对涂层图像进行明确的标记，以准确确定涂层质量和损坏类型。胡等人[65]通过基于 SSEResNet101 骨干的特征融合方法改进了 SSEResNet101 回归模型，通过图像预测表面粗糙度。
Deep learning-based methods require clear labeling of coating images to accurately determine the coating quality and types of damage. Hu et al. [65] improved the SSEResNet101 regression model to predict surface roughness through images using feature fusion methods based on the SSEResNet101 backbone.
此外，他们还开发了一种增强型级联 R-CNN 模型，该模型能够有效地识别飞机涂层激光清洗过程中产生的火焰变化，从而能够精确评估清洗质量。
Additionally, they developed an enhanced Cascade R-CNN model that effectively identifies flame changes produced during the laser cleaning of aircraft coatings, allowing for the precise evaluation of cleaning quality.
SSEResNet101 模型训练过程中的均方误差（MSE）损失为 0.0249，平均绝对误差（MAE）为 0.278μm。
The mean square error (MSE) loss during the SSEResNet101 model training was 0.0249, and the mean absolute error (MAE) was 0.278 μm.
改进后的级联 R-CNN 模型在交并比（IoU）为 0.6 时的平均准确率（mAP）值达到 93.6%。一些研究已经利用 CNN 网络对涂层的形态特征进行多层提取。Liu 等人[66]开发了一个智能评估系统，该系统包括基于区域的卷积神经网络（FAST R-CNN）和深度迁移学习 Vgg19 模型，能够有效地识别表面、边缘和焊缝上的涂层失效和腐蚀（CBC）。Holm[67]比较了不同卷积神经网络（CNNs）在自动分类桥梁结构腐蚀和涂层损伤图像方面的性能。
The improved Cascade R-CNN model achieved a mean accuracy (mAP) value of 93.6% at an intersection over union (IoU) of 0.6. Several studies have utilized CNN networks for the multi-layer extraction of the morphological features of coatings. Liu et al. [66] developed an intelligent evaluation system that includes a region-based convolutional neural network (FAST R-CNN) and a deep migratory learning Vgg19 model, which efficiently identifies coating breakdown and corrosion (CBC) on surfaces, edges, and welds. Holm [67] compared the performance of different convolutional neural networks (CNNs) for automatically classifying bridge structural corrosion and coating damage in images.
VGG-16 训练的卷积神经网络表现出了最佳的整体性能，其召回率、准确率、精确率和 F1 得分分别为 95.45%、95.61%、97.74%和 95.53%。
The VGG-16-trained convolutional neural network demonstrated the best overall performance, with recall, precision, accuracy, and F1 scores of 95.45%, 95.61%, 97.74%, and 95.53%, respectively.

卷积神经网络擅长提取涂层的表面形貌和厚度特征，可以与电化学测试数据进行比较。Samide 等人[68]在 0.1 mol-L-1 HCl 溶液中对 PVA 和 nAg/PVA 涂层进行了电化学测试和扫描电子显微镜（SEM）观察。
Convolutional neural networks are adept at extracting the surface topography and thickness features of coatings, which can be compared with data from electrochemical tests. Samide et al. [68] conducted electrochemical tests with scanning electron microscopy (SEM) observations on PVA and nAg/PVA coatings immersed in 0.1 mol-L-1 HCl solution.
他们的图像特征使用卷积神经网络进行提取，以评估 PVA 和 nAg/PVA 在延缓铜腐蚀方面的性能。
Their image features were extracted using convolutional neural networks to evaluate the performance of PVA and nAg/PVA in retarding copper corrosion.
CNN 数据与电化学测量和 SEM 结果进行了比较，结果表明在 PVA（0.94）和 nAg/PVA（0.98）存在的情况下表面覆盖率很高。Schmitz 等人[69]结合 FEA 数值模拟和深度学习方法通过色散曲线来表征涂层厚度和均匀性。
The CNN data were compared with results from electrochemical measurements and SEM, showing high surface coverage in the presence of PVA (0.94) and nAg/PVA (0.98). Schmitz et al. [69] combined FEA numerical simulations with a deep learning approach to characterize coating thickness and uniformity through dispersion curves.
特征通过将瞬态导波转换为频散图来提取，表明其对分类涂层厚度具有可行性。于等。[70]提出了一种结合深度卷积神经网络（CNN）和基于 Dempster-Shafer（D-S）理论的改进数据融合方法的新方法，用于评估煤炭输送和洗选厂的腐蚀和涂层缺陷。
Features were extracted by transforming transient guided waves into dispersion plots, showing feasibility for classifying coating thickness. Yu et al. [70] proposed a novel method combining a deep convolutional neural network (CNN) with an improved data fusion approach based on Dempster–Shafer (D-S) theory to evaluate corrosion and coating defects in coal transmission and washing plants.
这种方法能够准确识别涂层缺陷，并且对各种类型和强度的噪声干扰具有鲁棒性。
This method accurately recognizes coating defects and is robust against various types and intensities of noise interference.

这些深度学习算法的进步显著提高了检测和分析涂层缺陷和腐蚀的能力，为维护结构完整性提供了更精确和高效的工具。表 3 总结了涂层识别的传统和深度学习方法。
These advancements in deep learning algorithms significantly enhance the ability to detect and analyze coating defects and corrosion, providing more precise and efficient tools for maintaining structural integrity. Table 3 summarizes traditional and deep learning methods for coating identification.

表 3. 涂层图像识别方法综述。
Table 3. Summary of image recognition methods of coatings.

方法 Method	具体算法 Concrete Algorithm	材料 Materials	主要内容 Main Content	参考文献 Ref.
传统方法 Traditional method	基于灰度的方法 Method based on grayscale	醇酸漆 Alkyd paint	根据灰度图像确定气泡大小和面积比，从而确定气泡等级 The bubble size and area ratio are determined based on the gray image, so as to determine the bubble grade	[45] [45]
	基于颜色空间的方法 Method based on color space	海上风力发电平台涂层 Coatings on offshore wind platforms	定量记录和评定涂层劣化过程 Quantitative recording and rating of coating deterioration processes	[52] [52]
	基于灰度的方法 Method based on grayscale	钢结构桥梁涂层 Coatings on steel bridge	多分辨率模式分类（MPC）被用于分析钢桥的表面涂层图像，并计算灰度图像中涂层和铁锈的比例 Multiresolution pattern classification (MPC) was used to analyze the surface coating images of steel bridges, and the proportion of coating and rust in gray images was calculated	[64] [64] [64]
基于深度学习的方法 Methods based on deep learning	深度迁移学习技术 Deep transfer learning techniques	压载舱涂层 The coating on the ballast tank	基于人工智能的 CBC 评估系统（A-CAS）被开发出来，以识别不同类型的涂层损伤和腐蚀 An AI-based aided CBC evaluation system (A-CAS) was developed to identify different types of coating damage and corrosion	[66] [66]
	卷积神经网络（CNN） Convolutional neural networks (CNN)	聚醋酸乙烯银纳米粒子（nAg/PVA） PVAsilver nanoparticles (nAg/PVA)	CNN 数据与电化学测量和扫描电子显微镜（SEM）数据进行了比较 The CNN data were compared with electrochemical measurements and scanning electron microscopy (SEM) data	[69] [69]

6. 未来展望
6. Future Outlook

当前腐蚀领域的图像识别方法主要集中在识别钢板和涂层的缺陷上。
Current image recognition methods in the field of corrosion predominantly focus on identifying defects in steel plates and coatings.
虽然形态分析有助于识别和分类各种缺陷特征和损坏特征，但它不能提供更深入的腐蚀电化学信息。因此，腐蚀数据的分析至关重要。Li [71]提出了一种“腐蚀大数据”方法，主张建立一个标准化的腐蚀数据仓库，同时进行数据建模，并使用这些数据进行腐蚀过程模拟和实验验证。
While morphological analysis helps recognize and classify various defect features and damage characteristics, it does not provide deeper insights into corrosion electrochemical information. Therefore, the analysis of corrosion data is crucial. Li [71] proposed a “corrosion big data” approach, advocating for the establishment of a standardized data warehouse for corrosion, along with data modeling and the use of this data for corrosion process simulation and experimental validation.
这种方法旨在模拟腐蚀过程并进行实验验证，同时还建议更直观地可视化和显示腐蚀数据。
This approach aims to simulate the corrosion process and validate it experimentally, and also suggests that corrosion data be visualized and displayed more intuitively.

深度学习在数据图像解析方面具有重要潜力。通过测试获得的电化学信息可以以图像的形式可视化和显示，然后可以使用卷积神经网络对各种腐蚀图像进行分析以提取特征。
Deep learning has significant potential for data image parsing. Electrochemical information obtained through testing can be visualized and displayed in images, which can then be analyzed using convolutional neural networks to extract features from various corrosion images.
将提取的特征信息与钢或涂层的形态变化联系起来，可以创建一个受控数据库，从而增强对腐蚀图像的理解。
Linking this extracted feature information with the morphological changes in steel or coatings allows for the creation of a controlled database, enhancing the understanding of corrosion images.
图像识别技术的应用可以加速腐蚀数据和图像的积累，从而拓宽其工程应用。
The application of image recognition technology can accelerate the accumulation of corrosion data and images, thereby broadening its engineering applications.
通过不断将各种缺陷图像与电化学图像进行关联，将有可能建立一个包含电化学信息的“腐蚀大数据”存储库，这为预测钢材或涂层的使用寿命提供了巨大的潜力。
By continuously linking various defect images with electrochemical images, it will be possible to establish a “corrosion big data” repository containing electrochemical information, offering great potential for predicting the lifespan of steel or coatings.

7. 结论
7. Conclusions

传统的检测方法主要依赖于人工目视检查，这种方法效率低下且成本高昂，需要大量的人力、物力和财力。
Traditional detection methods mainly rely on manual visual inspection, which is inefficient and costly, requiring significant human, material, and financial resources.
相比之下，图像处理技术可以提高识别金属或涂层缺陷的速度和准确性。本文总结了图像识别技术在金属、腐蚀数据和涂层中的应用：
By contrast, image processing technology can improve the speed and accuracy of identifying metal or coating defects. This paper summarizes the application of image recognition technology in metals, corrosion data, and coatings:

• 传统方法已被有效用于检测钢结构中的裂纹、麻点和其他缺陷。
  Traditional methods have been effectively used for detecting cracks, pitting, and other defects in steel structures.
  通过对腐蚀钢材的采集图像进行分析，提取腐蚀特征，并运用数学方法挖掘灰度图像的深层特征，从而能够区分各种腐蚀形貌。
  By analyzing collected images of corroded steel, extracting corrosion features, and applying mathematical methods to mine the deep features of grayscale images, it is possible to distinguish between various corrosion morphologies.
  然而，这些模型通常缺乏泛化能力，这意味着环境和识别对象的变化会极大地影响准确性。此外，选择合适的模型和提取特征可能很复杂。
  However, these models often lack generalization, meaning that changes in the environment and the recognition object can greatly affect the accuracy. Additionally, selecting the right model and extracting features can be complex.
  深度学习方法需要大量的钢材表面形貌图像，提供了一种解决方案，不同的应用场景可以使用相同的模型，尽管使用的是不同的数据集。
  Deep learning methods, which require a large number of steel surface topography images, offer a solution where different application scenarios can use the same model, albeit with different datasets.
• 通过分析金属腐蚀过程中电化学参数的变化图像，可以提取更深入的腐蚀信息，例如确定等效电路图和预测涂层的使用寿命。
  By analyzing images of changes in electrochemical parameters during metal corrosion, deeper corrosion information can be extracted, such as determining equivalent circuit diagrams and predicting the lifetime of coatings.
• 传统的涂层图像识别方法与金属的类似。然而，深度学习图像识别算法在识别涂层的边界、厚度和均匀性方面显示出更大的潜力。
  Traditional image recognition methods for coatings are similar to those used for metals. However, deep learning image recognition algorithms show more potential in recognizing the boundary, thickness, and uniformity of coatings.

将电化学测试数据与图像识别技术相结合是一个有前途的研究方向。这种集成使得图像识别的结果可以得到电化学信息的支持。
Combining data from electrochemical testing with image recognition technology is a promising research direction. This integration allows the results of image recognition to be supported by electrochemical information.
虽然需要大量的数据积累和工作量，但它提供了一种更快的腐蚀检测和评级方法。
Although it requires a substantial amount of data accumulation and workload, it offers a faster method for corrosion detection and rating.