Defect-Net: a new CNN model for steel surface defect classification

Image comprehension fundamentally lies in the extraction of relevant information and the subsequent recognition of domain-specific objects, analyzed in a predefined conceptual framework. Therefore, modeling of image content becomes paramount to exploit the inherent dependencies among patterns within a given domain. Constructing a Convolutional Neural Network model allows for the optimization of knowledge leverage in the nuanced interpretation of images within a specific domain. In this paper, we introduce Defect-Net, a new CNN-based architecture designed for the classification of steel surface defects. Our proposed classification system exhibits exceptional efficiency in discerning challenging defect types, showcasing high precision in classification. Through a series of experiments leveraging the CNN architecture, we enhance the efficiency of our defect classification system by refining the network optimizer. Validation of the proposed approach has been done on an industrial radio-graphs dataset NEU-DET. Accordingly, Defect-Net yields significant improvements of classification performances, boasting an impressive mean average precision (mAP) of ${9 7 . 7 4 \%}$. This accomplishment surpasses the performance of existing methods from the state of the art, underscoring the effectiveness and superiority of our developed CNN architecture in elevating the accuracy of surface defect detection.