Surface Defect Detection in Manufacturing using Logistic Regression and Hybridized Feature Engineering

Surface defect classification is a common bottleneck in many areas of manufacturing because manual visual inspection tends to be time-consuming and can have inconsistency between human inspectors who may not be as consistent as desired. In this study, we evaluate a lightweight, interpretable logistic regression pipeline developed on the NEU Surface Defect Database. We used feature engineering to classify six categories of defect. A comparison is given across several feature configurations composed of Histogram of Oriented Gradients (HOG), Principal Component Analysis (PCA), and edge detection. Among all the configurations that were tested by using 1,440 training and 360 test images (80/20 split), combining HOG and PCA was found to offer the best performance, achieving ∼75% overall accuracy and a balanced F1 score (macro-averaged F1 = 0.75) across all classification categories while only running on CPU hardware. In order to demonstrate the possibility of deploying this solution in practice, we have created an application using the Streamlit framework to support image uploading (for analysis), as well as displaying the predicted label, class probabilities and an output inspection tool to view the data that was subject to feature engineering. The results indicate that classical computer vision features can be effectively combined with linear models to provide a viable alternative to deep learning architectures in constrained manufacturing environments.