Toward Automated Quality Control: An End-to-End Image Processing and Machine Learning Pipeline for Bullet Case Mouth Defect Detection

The structural integrity of metallic bullet casings is critical not only for military applications—where a single defective casing can compromise ammunition performance and jeopardize personnel safety—but also for industrial manufacturing, where undetected flaws translate into costly rework and production delays. In this study, we present a comprehensive, automated pipeline that combines image processing, feature extraction and machine learning to detect defects in the mouth regions of bullet casings with high reliability and scalability. A dataset of 860 images (440 defective, 420 normal) was captured under varying lighting and camera angles. Each image undergoes adaptive thresholding and morphological opening/closing to isolate the rim contour, followed by Hough Transform–based circle detection and radius comparison to align and normalize the mouth region. Geometric features—including area, perimeter, circularity, aspect ratio, extent and solidity—are computed from the largest contour, while variance analysis of radial profiles captures subtle surface irregularities. Extracted features were fed into a suite of classifiers (Logistic Regression, Support Vector Machine, Decision Tree, Random Forest, XGBoost), alongside an Artificial Neural Network). The best-performing Artificial Neural Network model achieved 97.3% accuracy on held-out test data, substantially outperforming tree-based and linear baselines. In summary, effective results are obtained when classical image processing techniques are combined with modern learning algorithms for real-time quality control of bullet cartridges used in the military.