Steel Defects Recognition Using Deep Learning

This paper proposes the application of transfer learning in classifying common surface defects found on hot-rolled steel strips. 3 potential models (Xception, InceptionV3 and InceptionResNetV2) are chosen and evaluated on their accuracy and F1 score using stratified cross validation. The experimental results show that InceptionV3 is able to classify surface defects with a relatively high accuracy and F1 score of over 95% while having a memory efficient architecture suitable for real world applications.