Document Type : Original/Review Paper
Authors
Computer engineering department, Yazd University
Abstract
Surface defect detection in steel billets is a key component of quality assurance in the steel industry. This paper presents an inspection framework based on deep neural networks for identifying surface defects in steel billets at the finishing section of the Iran Alloy Steel Factory. The proposed system tackles major industrial challenges in this category, including surface light reflections, limited labeled data, high defect variability, high billet movement speed, and the complexity of visual inspection. First, a controlled-illumination light box is designed, through which the billet passes centrally. Three Basler cameras are mounted on a custom circular disk at 120-degree intervals to capture surface images. Image preprocessing and data augmentation techniques are employed to reduce image noise and compensate for data scarcity. Subsequently, a Convolutional Neural Network (CNN)-based method is developed to analyze the preprocessed images and detect surface defects. The dataset consists of manually annotated images of steel billets collected under diverse operating conditions. Due to the limited dataset size, the early layers of the designed networks are frozen, and only the parameters of the final layers are fine-tuned. Experiments are conducted using both color and grayscale images. In total, eighteen models are trained and evaluated based on accuracy, sensitivity, and real-time applicability. Experimental results indicate that the VGGNet-based model achieves a detection accuracy of 95%, demonstrating its effectiveness for real-time industrial applications.
Keywords
Main Subjects
)