PSI - Issue 70

Bagathi HarshaVardhan et al. / Procedia Structural Integrity 70 (2025) 447–452

450

VGGNet is known for its simplicity and effectiveness in image classification and object recognition tasks. The hallmark of VGGNet is its use of very small convolutional filters (3×3) stacked together to increase depth, enabling it to learn complex features. 2.1 Hardware and software configuration: In the present study, to train the deep learning model workstation featuring NVIDIA 2000 ada (memory 16GB) with software like python, tensor flow,Ubuntu etc. is used. 2.2 Preparation of Image Dataset: The imaginary dataset used in this work collected from diverged sources like online repositories and real time captured images, the dataset consists of a total of 300 images, categorized into three groups: 100 images representing no corrosion, 100 images depicting medium corrosion, and 100 images showing severe corrosion. To ensure effective model training and evaluation, the dataset is split into 70% for training, 10% for validation, and 20% for testing. Furthermore, image augmentation techniques such as rotation, flipping, and contrast adjustments are applied to enhance the model's ability to generalize across different lighting conditions and surface textures, ultimately improving its robustness in real-world applications. The dataset undergoes preprocessing where images are resized to 512x512 pixels to maintain uniformity. Normalization is applied using mean values of [0.485, 0.456, 0.406] and standard deviations of [0.229, 0.224, 0.225] to standardize input data. The dataset is loaded using PyTorch's ImageFolder and is divided into training, validation, and test sets for model evaluation. 2.3 Source of Corrosion Images: In the present work, the dataset of corrosion images was compiled from two main sources: publicly available online repositories and real-time images captured during laboratory inspections of steel structures at the Civil Engineering Department, VNR Vignana Jyothi Institute of Engineering and Technology. The online datasets comprised images of corroded rebars, structural steel members, and steel pipes, reflecting a variety of corrosion scenarios. The real-time images were captured in controlled laboratory conditions using a high-resolution camera to ensure clarity and consistency. All the images were carefully annotated and divided into three classes — no corrosion, medium corrosion, and severe corrosion — to facilitate accurate classification by the deep learning models.

Table 1. Details of training parameters

Parameters

Values

No of Classes

3

Optimiser

adam 0.001

Learning rate Batch size No:of epoch

32 50

3. Interpretation of results: The pre-trained ResNet50 model achieved an accuracy of 76.7% in classifying corrosion severity levels. However, the confusion matrix indicates minor misclassification instances, especially between 'Medium corrosion' and 'Severe corrosion." High validation accuracy indicates that the model has effectively learnt the features from the training and exhibits good performance on the unseen data in the present case the test accuracy and training accuracy differ by more than 10 percent indicating that model has learnt the significant patterns more effectively giving rise to good accuracy, the jagged pattern of the epoch vs validation accuracy shows overfitting of data this may be due to diversed class distribution in the original dataset. To avoid overfitting early stopping technique is introduced with 10 number of epochs as the patience value. Patience is defined as the number of consecutive epochs the model is allowed to run.