Regional Feature Analysis for Automated Welding Defect Classification: Statistical Decomposition Approaches

Automated welding defect detection is a critical challenge in manufacturing, where traditional image analysis fails to capture the spatially-dependent nature of defects. This study addresses this limitation by proposing and systematically evaluating a toolkit of five regional feature extraction methodologies. The proposed methods capture defect signatures across distinct regions: Cumulative Projection Profiling (CPP) for high-dimensional spatial analysis (6184 features); Multi-Scale Spatial Grid Feature Extraction (MSSGFE) for hierarchical statistics (336 features); Regional Edge Direction Analysis (REDA) for edge orientations (64 features); FFT Grid Feature Extraction (FFTGFE) for frequency-domain patterns (64 features); and LBP Directional Strip Analysis (LBPDSA) for textural characteristics (32 features).A standardized preprocessing pipeline and a consistent deep learning framework were used to ensure a rigorous comparative evaluation. Experimental results show a clear performance trade-off between computational efficiency and data footprint. The CPP, MSSGFE, and FFTGFE methods demonstrated suitability for high-throughput systems, achieving classification accuracies above 99% at processing speeds exceeding 95 FPS. In contrast, the REDA and LBPDSA methods produced extremely compact feature vectors (\((<)\)260 bytes), making them ideal for deployment on resource-constrained embedded devices. The findings establish a practical framework for implementing automated inspection systems, enabling engineers to select an optimal methodology based on specific hardware constraints and application requirements in Industry 4.0 environments.