Akademik Veri Yönetim Sistemi
MATERIALS TODAY COMMUNICATIONS, cilt.44, ss.2352-4928, 2025 (SCI-Expanded)
Corrosion is a problem that causes structural deterioration in steel materials as a result of chemical and electro-chemical processes, seriously adversely affecting the mechanical properties and long-term performance of the material. In order to preserve the service life and performance of steel materials, regular monitoring and evaluation of corrosion is of vital importance. The aim of this study is to investigate the corrosion detection performance of deep learning models and to develop a real-time corrosion detection system. In this scope, three different deep learning models, namely U-Net, Detectron2 and YOLOv8, are organised with different network architectures. In order to train and test the models, 812 images were taken from different steel surfaces, these images were tripled with various augmentation techniques and a dataset consisting of 2436 images in total was generated. In order to evaluate the performance of the models, the dataset was split into 80% training, 10% validation and 10% test set. Then, five different U-Net submodels, seven different Detectron2 submodels and five different YOLOv8 submodels were trained and tests were performed on these trained models. In the analyses performed on the test set, the highest mean Intersection over Union (mIoU) value (0.8815) between the U-Net, Detectron2 and YOLOv8 models was obtained with the ResNet50 U-Net model. The X_101_32x8d_FPN_3x model of Detectron2 gave the highest bounding box and segmentation mean Average Precision (mAP) results, while the YOLOv8x model provided the most successful results with 0.849 precision and 0.690 mAP50. In terms of inference time, YOLOv8n-seg was the fastest model. When the accuracy-speed relationship was evaluated, the YOLOv8l model was determined as the most suitable option for real-time segmentation and the ‘Corrosion-Seg’ software performed successful tests with an average value of 22 frame per second (fps).