Evaluation of the convolutional neural network’s performance in classifying steel strip’s surface defects

Abstract

Steel strip plays a vital role in many industrial fields. Its defects will impact the manifestation of the product and also reduce the features of the product, resulting in a huge economic loss. Deep learning algorithms, such as Convolutional Neural Network (CNN), have successfully been applied to image classification, while featuring a great level of abstraction and learning capabilities. These features are keys to detect and classify surface defects in a robust and reliable manner. The images used for training and testing the model are obtained from the NEU Surface Defect Database which contains six kinds of typical surface defects of steel strips that are rolled-in-scale, patches, crazing, pitted surface, inclusion and scratches. These images are pre-processing to enhance them and extract some useful information from them. After that, the CNN models are trained and tested with these images to evaluate their performance. The specific hyperparameters for the CNN model which are tuned are number of epochs, batch size, number of convolutional layers, input image size and kernel size. For each hyperparameter, the CNN model is trained and tested several times using different values of that hyperparameter. The training accuracy, testing accuracy and training time are recorded and analyzed. Lastly, the final CNN model with high performance is produced. The final CNN model based on the optimum hyperparameters is produced. It has a very high training accuracy of 95.12% and a fairly high testing accuracy of 85.43%. This paper focused on the application of CNN in the classification of the steel strip’s surface defects. The performance of the CNN models with different values of hyperparameters are also evaluated.