Effect of Image Preprocessing and Visualization Techniques on the Prediction Accuracy of Machine-Learning-Based Nut Defect Detection Algorithms

Defect detection in fastening nuts is critical for ensuring product reliability in manufacturing. This paper proposes optimal combinations of image preprocessing and visualization techniques to enhance the reliability and interpretability of machine-learning-based nut defect prediction algorithms. Polarization images were preprocessed using methods, such as region-of-interest extraction, flattening, Canny edge detection, and binarization. A VGG16-based convolutional neural network was trained for prediction. To analyze the decision-making of the model, various visualization techniques, including Grad-CAM, Grad-CAM++, and Score-CAM, were employed. The results showed that the prediction performance was sensitive to preprocessing, yielding an improvement of up to 3.7% in accuracy. Furthermore, visualization facilitated a comparative analysis of the focus areas of the model, enabling the identification of the optimal preprocessing–visualization combination. These findings demonstrate that the proposed approach can significantly contribute toward developing reliable and explainable AI-based defect prediction algorithms applicable directly to manufacturing sites.