Automatic surface crack detection on textured materials

Abstract A novel image transformation is proposed for the automatic detection of crack-like faults on inspected surfaces with “busy” or patterned backgrounds (textured surfaces). The textured background is modeled by the proposed transformation, which consists of the Walsh transformation of a nonlinear function of the image intensity. The algorithm includes three stages: training the system to learn the underlying background textural pattern by using faultless images, obtaining the local difference from the underlying texture for the images to be tested, and postprocessing the difference map to isolate the fault pixels. By this method, crack-like faults in random and regularly textured backgrounds can be detected reliably and efficiently. The performance of the algorithm is demonstrated on artificially created faults on some Brodatz texture images as well as on some real images of materials with genuine faults on them.