New Possibilistic Formalism Based on Confidence Levels for Metallic Surface Defect Classification

• Published in: IEEE access 2024, Vol.12, p.93010-93019
• Main Authors: Elleuch, Jihen Frikha, Mehdi, Mouna Zouari, Kammoun, Habib M., Bellaaj, Majd, Sellami, Dorra
• Format: Article
• Language: English
• Subjects: Ambiguity; Classification; Classification algorithms; Comparative studies; Confidence intervals; confidence levels; Data models; Deep learning; Defect detection; Feature extraction; Formalism; Image processing; Image quality; Industrial applications; Intervals; possibility transformation; probability intervals; Production systems; Quality assessment; Statistical analysis; Surface defect classification; Surface defects; Surface texture; textural wavelet quantization; Transforms; Uncertainty; Wavelet analysis
• ISSN: 2169-3536; 2169-3536
• DOI: 10.1109/ACCESS.2024.3414840

In several industrial applications, visual inspection impacts production and quality of products. In case of surface defect on metal structures, a decision should be made to discard the defective piece and prevent from possible failure in further steps of an industrial process. Accordingly, the decision-making should be automated using advanced image processing methods which must take into consideration the various surface defect types. Nevertheless, they can sometimes look similar causing a high level of ambiguity, even from the point of view of an expert. Possibility theory can offer a formalism and allows coping with such ambiguity, yielding more reliable decisions. Moreover, possibility transform based on probability intervals can tackle the problem of lack of information or sampling problem, and can describe in a more realistic way the features within a specific confidence region level based on Goodman concept. Then, a further transformation of interval probability to possibility using Masson et Denoeux formalism, allows a possibility mapping of the different classes at a given confidence level. In this paper, we propose a new formalism based on possibility transform, handling probability intervals, for addressing ambiguity and uncertainty affecting steel surface defects. The effectiveness of the proposed method is largely confirmed by carrying out comparative studies with other methods similar to deep learning or conventional methods. Accordingly, two validation databases NEU-DET and GC10-DET are deployed, demonstrating that the the possibility approach allows an improvement in the classification of surface defects by achieving accuracy rates sometimes approaching 100%, outperforming for the majority of the considered defect types, existing approaches.