Comparison of Traditional Tool Combination and Deep Learning Methods for Automatic Classification of Steel Surface Defects

In industry, the automatic recognition of surface defects of flat steel products still represents a real challenge. Indeed, in addition to constraints such as the image noise or blur, there is neither an agreed standard of these defects, which can be of different sizes, shapes, orientations, and locations, nor a standard defect identification method, whether based on traditional tools such as statistical-based, spatial/frequency filtering based or model-based or on newer techniques such as deep learning. Hence, the complexity of the algorithm that deals with this matter always depends on the specific needs of the application. This paper focuses on the comparison of deep learning approaches with a new combination of traditional tools, namely the Gabor wavelets (GW) and the Local Phase Quantization, respectively known for image abrupt change representation and for the invariance to image blur. Both approaches are assessed on the North University (NEU) and a Research Center (RC_DDB) datasets, which include surface defect images of hot rolled steel product. Interesting correct-prediction rates and processing times are achieved for both approaches, with slightly higher scores for the Deep Learning approaches in terms of accuracy, but better computation times for the proposed combination.