Model-Centric or Data-Centric Approach? A Case Study on the Classification of Surface Defects in Steel Hot Rolling Using Convolutional Neural Networks

Any industrial application that uses convolutional neural networks (CNNs) requires initial data and resources in order to train the models. However, the selection of models must be appropriate to the quality and quantity of the available data and computational resources. This study analyses the influence of data quantity and quality on the performance of CNN models of different complexity. Image preprocessing and image transformation data augmentation techniques are applied to generate different amounts of synthetic data with which to train the aforementioned models, shedding light on the following question: does the quality and quantity of the data or the depth of the model have more influence? Different experiments are performed using the Northeastern University (NEU) Steel Surface Defects Database, which contains surface defects found in hot-rolled steel. After analyzing the results, the authors conclude that data quality and quantity have a much greater influence than model choice. As resources and time are often limited in industry and the ultimate goal is to maximize profit by increasing efficiency, the authors encourage researchers to carefully consider the industrial application at hand and analyze the available data and resources before selecting CNN models.