A context-driven approach to image-based crack detection

We present a novel context-driven approach to image-based crack detection for automated inspection of aircraft surface and subsurface defects. In contrast to existing image-based crack detection methods, which rely mostly on low-level image processing and data-driven methods, our method explicitly incorporates multiple high-level context into low-level processing. We present two classes of context: geometric/structural context and physical context. We formulate mathematically a sparse decomposition problem to incorporate the context and apply robust principal component analysis to decompose typical repetitive rivet regions into a normal component and a sparse component. Cracks are detected in the sparse component. By applying the proposed context-driven approach to coated and uncoated test specimens, we achieve significant reduction in false detections compared to the approach without exploiting context.