PERFORMANCE OF DIFFERENT CNN-BASED MODELS ON CLASSIFICATION OF STEEL SHEET SURFACE DEFECTS

Today, hot rolled steel strip has been widely used in the automobile industry and construction. However, the manufacturing process of this steel often causes various defects on the surface, thereby causing great damage to enterprises. The detection of steel surface defects becomes an indispensable part of iron and steel production facilities. Recently, deep learning has been broadly applied in the domain of image recognition. However, training a deep learning model from scratch creates many problems because the state-of-the-art CNN requires considerable training size and computational resources. For small data scale, transfer learning is an option. This paper focuses on cross-comparing the performance of different CNN-based models on the NEU database of 1800 images which consists of six different typical surface defects. In addition, the performance of optimization algorithms is also cross-compared by incorporating these algorithms in the model with the best performance or highest accuracy. The results show that the DenseNet121 network using the Adam optimizer performs the most effectively with an accuracy of 99.26% for testing set.