Deep learning-based steel surface defect segmentation using multi-scale feature learning and attention mechanisms
Year: 2026
Type: dissertation
Abstract: Surface defect segmentation is important for quality control in manufacturing. However, defects with different sizes, low contrast, and backgrounds that look similar to defects can make segmentation less accurate. This thesis proposes ID-MSNet, a UNet3+-based model for steel surface defect segmentation. This model combines a Multi-Scale Feature Learning Module to improve feature extraction, an Improved Down-Sampling method to preserve spatial information, Squeeze-and-Excitation blocks, and a Convolutional Block Attention Module to support defect localization. ID-MSNet was evaluated on the SD-saliency-900 dataset and the Severstal Steel dataset based on metrics such as Precision, Recall, F1 score, Dice score, and mIoU. In SD-saliency-900, the F1 score of the model is 87.93%, and the mIoU is 86.17%. In Severstal Steel, the Dice scores ranged from 0.82 to 0.87 across different training and evaluation runs. The thesis also reviews recent work on Transformer models, foundation models, explainable AI, and real-time use.
Author: Saberironaghi, Mohammadreza
Institution: Ontario Tech University
Language: English
Subfield: Industrial and Manufacturing Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7213390189
