Deep Learning-Based Decision Support System for Automatic Detection and Grading of Surface Corrosion on Galvanized Steel Sheets

Abstract

Corrosion in the sheets produced leads to significant material losses, including the loss of resources, capital, labor, energy and knowledge. Corrosion control is significant for sheets produced and sent to customers in iron and steel factories. Surface corrosion testing of produced sheets and the accurate detection of corrosion levels are of great importance. The corrosion detection process for sheets in steel factories is performed visually with the naked eye. This is a subjective and time-consuming method. Identifying corrosion damage by visual detection and accurately determining the type and extent of corrosion requires expertise. Wrong decisions at this stage lead to losses during the production phase. Therefore, there is a need for systems that can automate this process and make it human-independent. In this study, a decision support system was designed to automatically detect the level of corrosion in galvanized sheets using convolutional neural networks. The average accuracy of the system is 97.5%, the average precision is 0.98, the average recall is 1 and the average F1 score is 0.99. The results we obtained indicate that a successful system has been developed for the detection and determination of corrosion levels. The high performance of the convolutional neural network models used for corrosion detection supports the practical applicability of the developed system. This system will increase the reliability and efficiency of industrial processes by enabling the accurate and automatic classification of corrosion. This system, which meets a significant need in this area for industrial organizations, reduces production costs and also makes the corrosion detection process more consistent and faster.

1. Introduction

Metals are found in nature mixed and balanced with chemicals in the form of oxides, sulphates and carbonates. Metal is separated from these mixtures and purified for industrial use. When purified in this way, the metal tends to revert to its natural state, i.e., the state in which it was originally found in nature, when exposed to atmospheric conditions, heat, humidity, pressure or contact with other metals or materials [1]. For this reason, metals undergo a reaction due to the chemical or electrochemical effects of their environment, and, without requiring any additional energy, they first become ionic and then combine with other elements in the environment to form low-energy compounds. This reaction, which causes the metal to lose its metallic properties, is defined as corrosion [2]. Corrosion is a very important natural phenomenon that occurs in all sectors of industry. Tanks, reservoirs, pylons, vehicles, machinery, metal buildings, underground pipelines, ships, piers and anything else exposed to the elements are all subject to the process of corrosion. The unexpected collapse of structures that have lost strength due to corrosion can cause industrial accidents. Wasted expenditure due to corrosion also increases production costs [3]. Corrosion and wear, or the combined effects of these damage mechanisms on metallic materials, cause billions of dollars in losses each year in industrialized economies. In particular, corrosion damage is the most common damage mechanism for metallic materials. Considering that nearly 7% of the nation’s Gross Domestic Product (GDP) is attributed to wear and corrosion damage, the importance of this issue becomes clear. While the total cost of 52 major natural disasters in the United States over the past 22 years was USD 380 billion, a 2001 corrosion cost study found that the direct cost of corrosion to metallic materials was USD 276 billion per year. This figure is equivalent to 3.1% of the US GDP [4]. For all of these reasons, corrosion is a major concern in the steel industry.

In steel factories, it is very important to carry out corrosion control on the sheets produced and sent to customers and to investigate the causes of corrosion on the corroded sheets. The excessive level of corrosion on the produced sheets poses a serious threat to all living beings, especially humans, and also leads to significant financial losses, such as the loss of materials, capital, labor, energy, and information. For these reasons, it is crucial to perform corrosion tests on the produced sheets and to accurately determine the amount and degree of corrosion. ASTM B 117 is a corrosion assessment standard used in steel factories that relies on visual examples to compare corrosion levels and is based on visual analysis [5]. Visual analysis to detect corrosion damage requires expertise to accurately determine the types and rates of corrosion. The operators use their expertise to visually assess the corrosion on the sheet surface [6]. This is a subjective and time-consuming method. Therefore, there is a need for systems that can automate this process and perform it independently of human intervention.

Production in all industries is characterized by increasing flexibility and adaptability. Due to the variability of products, requirements for quality characteristics may change frequently, and therefore quality assessments cannot be easily repeated. The increasing customization of assessments increases the already high number of manual visual inspections [7]. In addition, visual assessment by workers is subjective [8]. The reliability of visual assessment depends on the experience of the operator. Therefore, the objectivity and reliability of visual inspectors depend on the expertise acquired through training. It is not always possible to find skilled workers at every stage; sometimes it may be necessary to authorize unskilled workers. Current approaches to automating visual inspection are designed to replace manual visual inspection [9].

There are many automated visual inspection applications used in industry [10]. Specifically, in an automated visual inspection system, there is a need to automate subjective quality criteria and eliminate subjectivity. In practice, deep learning offers great potential to objectively assess the condition of a component. In recent years, deep-learning methods have been widely used in computer-vision and image-processing applications [11]. These methods are highly advantageous because they can extract features directly from images [12]. Deep-learning models are artificial neural networks with multiple hidden layers [13]. Depending on the characteristics of the dataset, such as the number of data points and the requirements of the model (e.g., computation time), an appropriate deep-learning model is selected. Recent studies have shown that deep-learning methods such as CNN and U-Net have been applied to image processing and recognition in various fields, from the steel industry [14,15] to textiles [16] and construction [17,18], and has demonstrated a high performance. In particular, texture analysis performed with convolutional neural networks (CNNs) can be easily integrated into inspection and characterization solutions based on quality control [19]. CNNs are widely used in the steel industry for defect detection in manufactured goods [20,21] and for product quality assessment [15,22]. Deep-learning models are divided into two categories: the multi-layer perceptron (MLP) and the CNN. To achieve better results in image recognition, deep-learning models are constantly being developed or improved. CNN models such as ResNet, GoogleNet and AlexNet have been shown to give good results [23].

A decision support system (DSS) is a computerized program used to support the identification, judgement and action plans within an organization or business. A DSS examines and analyzes large amounts of data by compiling comprehensive information that can be used to solve problems and make decisions [24]. A decision support system collects and analyzes data and synthesizes them to produce comprehensive information reports. In this way, as an information application, a DSS differs from a standard transactional application, which only functions to collect data [25]. A DSS may be fully computerized or human-assisted. In some cases, it may be a combination of both. They make decisions on behalf of the user, allowing more informed decisions to be made faster than human users [26]. The use of decision support systems has gained significant importance in industry and production environments over the last twenty years [27]. Decision making, in this context, is a crucial technique that helps to generate new ideas and accurate guidance in the smart industry [28]. Intelligent decision support system methods are used to extract meaningful information from data. Given the large numbers of data generated by industrial processes and the different scenarios considered for decision making, decision support systems have become increasingly important for the entire production chain [29].

In this study, a decision support system for the steel industry was designed to automatically detect corrosion and the level of corrosion on galvanized sheets using CNN, independent of human intervention. The designed system can automatically and accurately detect the level of corrosion at four different levels, using images of corroded galvanized sheets.

The main contributions and innovations of this paper are as follows:

(1)

A novel and unique system was developed that detects the level of corrosion on galvanized sheets using CNN. This system can automatically and accurately detect the level of corrosion at four different levels.

(2)

This study is the first of its kind in the literature and fills a significant gap and critical need in the field. By developing a system to accurately detect and classify corrosion levels on galvanized steel sheets using deep learning, this research contributes valuable insights and tools for corrosion assessment, paving the way for future advances in the field.

(3)

A highly useful application was developed that outperforms traditional visual analysis methods for corrosion detection in the steel industry. This system is faster, more stable and more consistent across different users, eliminating the variability and limitations associated with human intervention. By automating the corrosion detection process, it provides a more reliable and efficient solution for the industry.

(4)

The developed system has a structure that can be practically used in many facilities within the steel industry and will support smart manufacturing

This paper is organized as follows: Section 2 provides an overview of the corrosion detection process on galvanized sheets and the designed decision support system, including the definition and acquisition of the dataset, CNN, and system performance evaluation. The test results of the system and the evaluation of these results are presented in Section 3. The Conclusion section (Section 4) contains interpretations of the main findings of the study.

2. Materials and Methods

In this study, a decision support system was developed to automatically detect corrosion and the level of corrosion on the surfaces of galvanized sheets using deep learning, independent of human intervention. The block diagram of the decision support system developed in this project is shown in Figure 1.

2.1. Description of the Dataset

Steel, a metallic alloy product, is subject to rapid corrosion due to environmental degradation, resulting in the formation of white and brown rust on its surface. Various methods can be used to mitigate the damage caused by corrosion. Among these methods, galvanizing—coating the surface of steel with zinc—is one of the primary techniques. The ASTM B 117 salt-spray test standard is widely used for comparative corrosion testing of galvanized steel to evaluate the corrosion resistance of metal and metal-coated surfaces. ASTM B117 is carried out in special chambers constructed to the specifications of the standard. A 5% NaCl salt solution, dissolved in pure water at 35 ± 2 °C, is continuously sprayed onto the surfaces of bare sheets and painted/coated materials placed inside the chamber to promote the corrosion of metals and aging of painted/coated materials. The sheets are placed in the chamber at an angle of 30° or 40°.

The standard requires samples to be exposed for 72 h. ASTM B 117 is a corrosion evaluation standard that uses visual examples as a reference for comparing corrosion levels and is therefore based on visual analysis [5]. The device used for the ASTM B 117 test, which is the current test method in the MMK laboratory, is the Sheen FMC1000 salt-spray fog chamber (Sheen Instruments, Cambridge, UK) shown in Figure 2.

The image of the 150 × 200 mm² standard sample coming out of the salt fog bath cabinet is shown in Figure 3.

MMK Metallurgy Steel Co. Ltd. (Hatay, Turkey), has an internal test method based on the ASM B 117 standard for determining the corrosion of galvanized steel sheets. At MMK Metallurgy Steel Co., Ltd. (Hatay, Turkey) the white rust rates of galvanized steel sheets are specified in 4 classes. These classes are 1st class, 1–25%; 2nd class, 26–50%; 3rd class, 51–75%; and 4th class, 76–100%. In this method, 150 mm × 200 mm samples are cut from the produced galvanized sheets and exposed to ASTM B 117 standard for 72 h. The corrosion result of the sample removed afterward is expected to be 25% and below.

To determine white rust, the operator uses a 150 mm × 200 mm transparent plastic template which is divided into 100 equal parts and drawn. After the test, the transparent plastic template is placed on the samples taken from the machine, the cells with white rust are calculated and the white rust is determined. The operator calculates the % cut by determining where white rust occurs on the sample. The determination of where the rust occurs is entirely based on the operator’s knowledge and experience. Two different operators can give different results for the same sheet. This process is subjective and varies from person to person.

For this study, 132 samples were obtained under the above conditions, and images of these samples were captured using a Canon IXUS (IXY Digital in Japan), camera, resulting in a database consisting of 132 images of corroded galvanized sheets.

In this study, the level of corrosion on the sheets was analyzed according to four different classes. Class 1 images represent corrosion levels of 0–25%, Class 2 images represent 26–50%, Class 3 images represent 51–75% and Class 4 images represent 76–100% corrosion levels.

The images obtained were shown to the 3 most senior and experienced operators at MMK Metallurgy Steel Co., Ltd., and they were asked to classify these images according to the abovementioned standard. The images were categorized according to these results. The number of images in each group is shown in

Table 1. Classes of corroded sheet metal and the number of images for each class.

Classes of Corroded Sheet Metal	Numbers of Images
Class 1 (1–25%)	33
Class 2 (26–50%)	33
Class 3 (51–75%)	33
Class 4 (76–100%)	33

.

It is critical to the performance of the designed classifier that all images have the same resolution, lighting and posing standards. To ensure that the images used are of the same standard, a cabinet was designed, as shown in Figure 4. This cabinet has a compartment where the corroded sheet is placed. At the top of the cabinet, there is space for the camera. Two lights are installed inside the cabinet for illumination.

Corroded sheet samples, exposed to corrosion for 72 h, were placed in the appropriate compartment of the cabinet shown in Figure 4, and their images were taken. In this way, 132 images of the same standard were obtained for each group. A sample image for each group can be seen in Figure 5.

The data to be used in the CNN contain undesirable noises and insufficient and unbalanced contrast distribution. The data to be used after the elimination of such problems will give more accurate results for CNN. In order to obtain more accurate results from the designed CNN, the images were pre-processed. Images were pre-processed as follows: RGB to grayscale converted, resized, contrast modified by processing pixel intensities. The images were first converted to grayscale. The images in the dataset have a resolution of 1350 × 2050 pixels. For the CNN models used in this study, the images were resized to resolutions of 227 × 227 and 224 × 224 pixels. In contrast modification, because the density of pixels in an image varies from image to image, the contrast process treats each pixel separately and emphasizes the important points in the image. In corroded images, the parts with rust and the parts without rust have different reflections and different pixel intensities under light. This process has highlighted the differences in rust in the images.

2.2. Convolutional Neural Network

Machine learning is a branch of artificial intelligence and computer science that develops methods to enable a system to make decisions about a specific problem by using data and algorithms that imitate the way humans learn. Artificial neural networks (ANNs) are a technique inspired by the way the human brain processes information and can solve problems in a mathematical environment. Deep learning involves the use of multi-layered “deep” neural networks to create systems capable of recognizing features from large numbers of labeled training data. Deep learning is an approach to machine learning. CNN is an architecture inspired by artificial neural networks that can learn from collected information from start to finish. Thanks to its intelligent model structure, CNN delivers highly successful results when working with large numbers of data. However, the disadvantages of CNN include a lengthy training process and the potential to get stuck in a local solution during the training phase.

CNN is a forward-processing architecture of artificial neural networks and a deep-learning approach that, unlike traditional neural networks, includes a feature-extraction layer [30]. Compared to other neural network models, such as the multi-layer perceptron (MLP), CNN is designed to take multiple sequences as input and then process the input using a convolution operator within a local region, imitating the way the eye perceives images. For this reason, CNN performs exceptionally well in solving computer vision problems such as image classification, recognition and understanding [31,32].

CNN is based on the 5-layer LeNet-5 architecture (excluding the input and subsampling layers), as shown in Figure 6. It consists of a convolutional layer, a subsampling layer and a fully connected layer. To handle more complex datasets and problems, different CNN architectures have been developed, such as AlexNet (8 layers), GoogLeNet (22 layers), VGG-16 (16 layers) and ResNet. Among these different architectures, CNNs have four key features: weight sharing, local connectivity, pooling, and the use of multiple layers. Typically, the input layer is followed by a convolutional layer. The convolutional layer is followed by a subsampling layer. This combination is repeated several times to increase the depth of the CNN. The fully connected layers are designed as the last layers to map the extracted features to labels.

CNNs perform feature extraction from images across layers. This process consists of several basic steps:

• Convolution Layer

The convolution layer uses small filters (kernels) to identify local features (such as edges, corners and color differences) in the image. This process allows the filters to recognize specific features in the image. The output is a set of feature maps. These maps represent important features in the image.

2.

Activation function

CNNs usually use the ReLU (Rectified Linear Unit) activation function. This function converts negative values to zero and leaves positive values as they are. This allows the network to learn non-linear features.

3.

Pooling Layer

The pooling layer is used to make the feature map smaller and more manageable. Max Pooling is often used. This shrinks the feature map by selecting the largest value in a given range, giving a more condensed set of information. This layer makes it more robust to the location of features and reduces the computational cost.

4.

Fully Connected Layer

Towards the end of the network, the extracted features are flattened and fed into one or more fully connected layers. These layers learn higher-level features for final tasks, such as classification or regression.

5.

Softmax Layer

The extracted features are then fed into a classification layer, such as Softmax, which assigns a probability to each class.

The CNN passes the input image through layers, extracting more complex features at each level. The first layers learn basic geometric features (edges and corners), and the subsequent layers learn more abstract and complex features. Convolutional filters are taught to extract the most meaningful features from the image, and these features are optimized as the network is trained. This structure gives CNNs a human-like ability to extract features and is particularly useful for tasks such as image classification and object recognition.

The layered structure of the CNN allows features to be determined at different levels within each layer. These layers operate at different levels of abstraction, from low level to high level, and extract meaningful features from the image. In the convolution layers of the CNN, the so-called low-level features are the detection of distinct edges in the image, the detection of points where different colors or tones merge, the determination of low-level texture patterns and the differences between pixel intensities. The first layers of the CNN work with small filters (e.g., 3 × 3 or 5 × 5 in size) and examine the relationships between local pixels. These filters begin to extract the basic features of the image by identifying simple structures in the image.

Convolution and Pooling layers are the layers from which mid-level features are extracted; the features determined in this layer are object parts, more complex patterns, and geometric shapes by combining several low-level features. Middle layers combine features from previous layers using larger filters (e.g., 7 × 7). These layers detect more complex object features and patterns. Pooling layers reduce the size of the image, resulting in a more concise representation of the features and allowing the model to be further abstracted.

Fully connected layers are the layers from which high-level features are extracted. The last layers are usually composed of fully connected neurons. These layers combine all the features from the previous layers and perform a final classification or recognition process. The features extracted at this stage provide general information about the whole image.

The output after the convolution stage is as shown in Equation (1).

$$A_j^l=f\left({\sum }_{i=1}^{M^{ı-1}}{A}_i^{l-1}{w}_{ij}^l+b_j^i\right)$$

(1)

where $M^{ı-1}$ is the number of feature maps in layer (l − 1), $w_{ij}^l$ denotes the kernel weights from feature map in layer 1 to feature map j in layer (l − 1) and $b_j^i$ is the bias parameter. Equation (2) is used to generate the feature maps.

$$G\left[m,n\right]=\left(f\ast h\left[m,n\right]\right)={\sum }_j{\sum }_{k}h\left[j,k\right]f[m-j,n-k]$$

(2)

In Equation (2), f is the input image, h is the filter and (m, n) is the size of the result matrix.

In this study, CNNs with deep-learning methods, such as AlexNet, GoogleNet and ResNet-50 models, were used. Stochastic Gradient Descent with Momentum (SGDM) [33] was used to optimize the CNN models used in this study.

SGDM is an improved version of the classical Stochastic Gradient Descent (SGD) algorithm. Its main aim is to speed up the learning process and overcome some of the weaknesses of SGD (e.g., slow approach to minima and getting stuck in false minima). The impulse term provides an additional acceleration effect to the optimization process, allowing the model to reach the optimum point faster and more reliably. Momentum increases acceleration in the gradient direction by incorporating information from previous steps, avoiding false minima and allowing the model to learn faster. The most appropriate optimization algorithm for specific datasets and problems may vary depending on the structure of the network and the training dataset. The SGDM algorithm is often used in CNNs because it provides a faster, more stable and more efficient learning process when optimizing deep and complex CNNs [34]. By adding momentum, the weaknesses of SGD are overcome, and a more reliable optimization is achieved. This enables CNN to work with higher accuracy in tasks such as classification and object recognition. Since classification is performed in this study, the SGDM algorithm is chosen because of these features.

In each iteration, the SGDM optimizer updates the weights and bias values of the network to minimize the loss function. Momentum is used to prevent oscillations along the steepest descent path [34]. SGDM is represented by Equation (3).

$${\theta }_{i+1}={\theta }_i\alpha \nabla E_R\left({\theta }_i\right)+\gamma ({\theta }_i-{\theta }_{i-1})$$

(3)

In Equation (3), θ represents the network parameter vector, i denotes the iteration number and α is the learning rate. For the different CNN models used in this study, α was set between 0.0001 and 0.001. $E_R$ is the loss funciton, while γ is the momentum term, which was set to 0.8 in this study. The categorical cross-entropy loss function shown in Equation (4) was used for the optimization process.

$$E_R\left(\theta \right)=-ln\left({\displaystyle \frac{e^{{\theta }_P}}{\sum _j^S{e}^{{\theta }_j}}}\right)$$

(4)

where θp is the CNN score for the correct class, j is the number of iterators and s is the number of classes.

This study used CNNs, including AlexNet, GoogleNet and ResNet-50 models, and compared the test results. To improve the performance of deep-learning models in accurately determining the level of corrosion on the surface of galvanized sheets, modifications were made to the architectures of AlexNet, GoogleNet and ResNet-50.

Determining the parameters in CNN models is a critical stage that directly affects the performance, efficiency and learning process of the model. Different strategies usually determine these parameters depending on the model design, dataset and objective. Accurate tuning of CNN parameters is usually an iterative process and is determined experimentally after many trial-and-error methods [35]. In this study, many different parameters were used and tested, and the parameters that gave the highest accuracy were selected. In other words, the parameters were experimentally determined as a result of many trial-and-error methods.

MATLAB R2023A was used in the CNN models used for the system developed in this study and the evaluation of the results.

2.2.1. AlexNet

AlexNet is the first architecture to use a GPU to improve training performance. It consists of 5 convolutional layers, 3 max-pooling layers, 2 normalized layers, 2 fully connected layers and 1 SoftMax layer. Each convolutional layer consists of a convolutional filter and a non-linear activation function called “ReLU”. Pooling layers are used to perform the maximum pooling function, and the input size is fixed due to the presence of fully connected layers. AlexNet has a large network structure with 60 million parameters and 650,000 neurons.

The activation function is used to provide the non-linearity property in neural networks. Therefore, traditional activation function options include the logistic function, tanh function, arctan function, etc. However, in deep models, these functions tend to encounter the problem of gradient vanishing because the gradient is a large value only when the input is in a small range around 0. To overcome this problem, a new activation function called the Rectified Linear Unit (ReLU) is used. The definition of ReLU is as follows:

ReLU(x) = max(x,0)

(5)

The gradient of ReLU is always 1 unless the input is less than 0. It has been shown that deep networks using ReLU as the activation function converge faster than the tanh activation function. This speed-up contributes greatly to training [36]. Convolution and pooling are used for automatic feature extraction and reduction.

In this study, the “relu” activation function with seven outputs is used for AlexNet. The number of epochs was set to 80, and the learning rate was set to 0.001. A stochastic gradient descent (SGDM) optimizer with 0.9 momentum was used.

2.2.2. GoogleNet

GoogLeNet is a 22-layer CNN architecture (27 including pooling layers) developed by researchers at Google [37]. This architecture uses a novel element called the initial module, which contains 1 × 1, 3 × 3, and 5 × 5 convolutional filters. This design allowed the number of parameters to be reduced from 60 million in AlexNet to 4 million.

In this study, the “relu” activation function was used to improve performance. The number of epochs was set to 65, the learning rate was set to 0.001 and the SGDM optimizer was used with a momentum of 0.9.

2.2.3. ResNet-50

ResNet is a CNN architecture that stands out for its solutions to the training difficulties in deep networks. The most distinctive feature of ResNet is the residual connections, which are direct connections that add the input of one layer directly to the output of the next layer. This allows the network to learn the difference between input and output, preventing the missing gradient problem and allowing the network to be deeper [38].

ResNet uses “identity mapping” to increase the learning capacity of the network, while preventing overfitting as the network deepens. The core idea of ResNet is to add “transformational connections” to the network. These connections are direct links from one layer to another within the network. This approach allows the network to become deeper, reducing the amount of information lost during training and allowing more effective gradient backpropagation.

One of the key benefits of ResNet is its ability to train deeper networks, resulting in higher accuracy. This is particularly effective for complex tasks such as image classification. ResNet also provides a suitable architecture for transfer learning; a pre-trained ResNet model can be easily adapted for different tasks [38]. Different versions and variants of ResNet are commonly used in deep-learning projects and are available in many deep-learning libraries.

In this study, the “relu” activation function was used for ResNet-50. The number of training epochs was set to 77, with a learning rate of 0.001. An SGDM optimizer with a momentum of 0.9 was used.

2.3. Evaluating the System’s Performance

The “confusion matrix” is a term used in fields such as statistics and machine learning [39]. It is used to assess how far a model’s predictions are from the actual values. It is typically used as part of a comparison matrix, which is a tool to measure the performance of a model.

The confusion matrix is used to evaluate the accuracy of a model’s predictions in a classification problem. The confusion matrix is shown in

Table 2. Confusion matrix.

Classification
Test negative	True negative (tn)	False positive (fp)
Test positive	False negative (fn)	True positive (tp)

.

The values in this table can be explained as follows:

True positive (tp): The number of true positives refers to cases where the model correctly predicts a class. In other words, it shows the cases where the model predicted “positive” and the actual value was indeed positive.

False positive (fp): The number of false positives refers to cases where the model incorrectly predicts a class as positive. The model incorrectly predicts instances that do not belong to a particular class as belonging to that class.

True negative (tn): The number of true negatives refers to cases where the model correctly predicts a class as negative. In other words, it shows the cases where the model predicted “negative” and the actual value was actually negative.

False negative (fn): The number of false negatives refers to cases where the model incorrectly predicts a class as negative. The model incorrectly predicts instances belonging to a particular class as not belonging to that class.

Using this matrix, performance metrics such as a model’s precision, recall and F1 score can be calculated. Precision measures the proportion of examples predicted as positive by the model that are actually positive, while recall indicates how many of the actual positives were correctly predicted. The F1 score combines precision and recall to provide an overall measure of the model’s performance.

To evaluate the effectiveness of the developed system, the accuracy, precision, sensitivity and F1 score were calculated as shown in Equations (6)–(9).

$$Accuracy={\displaystyle \frac{t_p+t_n}{t_p+t_n+f_p+f_n}}$$

(6)

$$Precision={\displaystyle \frac{t_p}{t_p+f_p}}$$

(7)

$$Recall={\displaystyle \frac{t_p}{t_p+f_n}}$$

(8)

$$F1 Score=2\times {\displaystyle \frac{precision\times recall}{precision+recall}}$$

(9)

3. Results and Discussion

In this study, a system was designed to automatically detect the amount of corrosion on corroded metal sheets. The designed system aims to classify the amount of corrosion in images of corroded sheets into one of four different classes. Class 1 images represent corrosion levels of 0–25%, Class 2 images represent 26–50%, Class 3 images represent 51–75% and Class 4 images represent 76–100% corrosion levels.

A total of 132 images with varying degrees of corrosion were used for system design and testing. Of these 132 images, 70% were used to train CNN models, and 30% were used to test the trained networks. Accordingly, 23 images from each class were used for training, and 10 images from each class were used for testing. The training and testing images from each class were randomly selected. For testing, 10 images from each class, making a total of 40 images, were not used during the CNN model-training phase.

AlexNet, GoogleNet, and ResNet models of CNN were trained with different parameters by making many trials. The parameters of the models (the details of which are given in Section 2.2.1, Section 2.2.2 and Section 2.2.3) with the best verification results are shown in

Table 3. The parameters and verification results for CNN models.

Model	Optimizer	Learning Rate	Mini Batch Size	Max Epochs	Validation Frequency	Validation Accuracy (%)
AlexNet	sgdm	0.001	128	80	3	92.86
GoogleNet	sgdm	0.001	128	65	3	89.29
ResNet-50	sgdm	0.001	128	77	3	82.15

.

As can be seen from Table 3, AlexNet was the model with the best validation result, with 92.86%. The accuracy rate for GoogleNet was 89.29%, and the accuracy rate was 82.15% for ResNet-50.

The test images were sequentially fed into the decision support system designed using AlexNet, GoogleNet and ResNet-50 to evaluate the system. Metrics such as accuracy, precision, recall and F1 score were calculated to evaluate the performance of the designed decision support system. The calculated values of accuracy, precision, recall and F1 score for AlexNet are shown in

Table 4. The system performance evaluation for AlexNet.

Class	Number of Tests	Correct Number of Grading	Accuracy (%)	Precision	Recall	F1 Score
Class 1	10	10	100	1	1	1
Class 2	10	9	90	0.90	1	0.95
Class 3	10	10	100	1	1	1
Class 4	10	10	100	1	1	1
Overall	40	39	97.5	0.98	1	0.99

.

As can be seen in Table 4, AlexNet gave only one incorrect result for Class 2. Otherwise, it correctly classified all the test images in the other groups. In all, 39 out of 40 test images are correctly classified. The average accuracy rate is 97.5%, the precision is 0.98, the recall is 1, and the average F1 score is 0.99.

The accuracy, precision, sensitivity and F1 score values calculated for GoogleNet are given in

Table 5. The system performance evaluation for GoogleNet.

Class	Number of Tests	Correct Number of Grading	Accuracy (%)	Precision	Recall	F1 Score
Class 1	10	10	100	1	1	1
Class 2	10	7	70	0.78	0.88	0.83
Class 3	10	9	90	0.90	1	0.95
Class 4	10	9	90	1	0.90	0.95
Overall	40	35	87.5	0.92	0.95	0.93

.

As can be seen from Table 5, the total number of correct tests obtained with GoogleNet is 35. Class 1 correctly classified all the test images. Class 2 correctly classified seven of the test images. Class 3 and 4 are correctly classified nine of the test images. The average accuracy rate is 87.5%, the precision is 0.92, the recall is 0.95 and the F1 score is 0.93.

The accuracy, precision, recall and F1 score obtained using ResNet-50 are shown in

Table 6. The system performance evaluation for ResNet-50.

Class	Number of Tests	Correct Number of Grading	Accuracy (%)	Precision	Recall	F1 Score
Class 1	10	10	100	1	1	1
Class 2	10	6	60	0.86	0.67	0.83
Class 3	10	4	40	0.50	0.67	0.57
Class 4	10	10	100	1	1	1
Overall	40	30	75	0.84	0.84	0.85

.

As shown in Table 6, the total number of correct tests obtained with ResNet-50 is 30. The system correctly classified all of the Class 1 test images. Six of the Class 2 test images are correctly classified. Four of the Class 3 test images, and all of the Class 4 test images are correctly classified. The average accuracy rate is 75%, the precision is 0.84, the recall is 0.84 and the F1 score is 0.85.

The results are consistent with the validation accuracy in Table 3. According to the validation accuracy in Table 3, the CNN model with the best result is AlexNet. Looking at the results in Table 4, Table 5 and Table 6, we can see that the model with the best classification parameters was AlexNet.

Figure 7 shows the performance evaluations of the decision support system obtained with AlexNet, GoogleNet and ResNet-50 for Class 1.

As can be seen from Figure 6, Class 1 corrosion level detection is successful for each model. The system correctly classified all test images in all three models.

Figure 8 shows the performance evaluations of the decision support system obtained with AlexNet, GoogleNet and ResNet-50 for Class 2.

As can be seen in Figure 7, the best performance among the CNN models used for Class 2 is obtained with AlexNet. The results obtained with AlexNet are quite high. The performance results obtained for GoogleNet and ResNet-50 are not sufficient.

Figure 9 shows the performance evaluations of the decision support system obtained with AlexNet, GoogleNet, and ResNet-50 for Class 3.

The best performance results of the Class 3 system were obtained with AlexNet. The results obtained with AlexNet are quite high. For AlexNet, the system correctly classified all the test images. The performance results obtained with GoogleNet are also quite high. However, the results obtained with ResNet are low.

Figure 10 shows the performance evaluations of the decision support system obtained with AlexNet, GoogleNet and ResNet-50 for Class 4.

The system performed well on all models used for Class 4. For AlexNet and ResNet-50, the system correctly classified all the test images. The results obtained with GoogleNet are quite high.

Figure 11 shows the average performance evaluations of the decision support system obtained with AlexNet, GoogleNet and ResNet-50 for four classes.

As shown in Figure 10, the CNN model that best classified the four different classes was AlexNet. The test results obtained with AlexNet show that the designed decision support system detects different levels of corrosion on corroded images with very high accuracy. The test results with GoogleNet show that although the system did not achieve the high accuracy of AlexNet, it still detected corrosion levels with high accuracy. The test results with ResNet-50 did not reach the expected high values. The results show that the developed system effectively determines the level of corrosion on galvanized sheets.

4. Conclusions

In this study, a decision support system was developed to automatically detect the level of corrosion on galvanized sheets using CNN, which is crucial for the iron and steel industry and operates independently of human intervention. A high-performance decision support system was designed to reduce production losses and address a significant need in industrial organizations. The study utilized images of corroded sheet samples obtained from MMK laboratories and applied CNN models, including AlexNet, GoogleNet and ResNet-50, to determine corrosion at four different levels. The performance of the models was compared.

The results are as follows:

• AlexNet: It achieved the highest performance, with an average accuracy of 97.5%, precision of 0.98, recall of 1 and average F1 score of 0.99.
• GoogleNet: It achieved an average accuracy of 87.5%, precision of 0.92, recall of 0.95 and F1 score of 0.93.
• ResNet-50: It achieved the lowest performance, with an average accuracy of 75%, precision of 0.84, recall of 0.84 and F1 score of 0.85.

The highest performance results were obtained with AlexNet, while ResNet-50 obtained the lowest. These results demonstrate that a successful system was developed for detecting and determining corrosion levels.

The developed system is faster, more stable and more consistent across different users, eliminating the variability and limitations associated with human intervention. By automating the corrosion detection process, the system provides a more reliable and efficient solution for the industry. This study fills a significant gap and critical need in the field, providing valuable insights and tools for corrosion assessment and paving the way for future advancements. The developed system is practically applicable across many facilities within the steel industry and will support smart manufacturing.

Future work may include automatic detection and grading of surface corrosion on pre-painted galvanized steel.