Abstract:
Defect detection in steel surfaces is critical for maintaining
product quality and ensuring safety in manufacturing processes. This paper
presents a comparative study of various transfer learning models applied to the
classification of steel surface defects, including rolled, crazing, pitting, scratches,
inclusion, and patches. We evaluate five prominent convolutional neural
networks: VGG16, VGG19, MobileNetV2, EfficientNetB0, and ResNet50, which
have been pre-trained on large-scale image datasets. The models were fine
tuned on a dataset specifically curated for steel surface defects, consisting of
diverse images to capture varia tions in appearance and context. Through
rigorous experimentation, we assess each model’s performance based on key
metrics such as accuracy, precision, recall, and F1 score, as well as their
computational efficiency in terms of execution time and memory usage. The
findings reveal signifi-cant differences in classification capabilities among the
models, highlight ing the strengths and weaknesses of each architecture in the
context of defect detection. Our study not only identifies the most effective
models for this specific application but also provides insights into the trade-offs
between accuracy and resource requirements, offering guidance for prac-
titioners in the field of industrial quality control.