Automatic detection of cracks and delaminations in thermal images

We present an algorithm that takes as input an image obtained by flying-spot active thermography, and outputs two detection masks: the first one for surface breaking defects for instance wear cracks – and underlying defects close to the surface; the second one for delaminations. In a first step, we perform a one-dimensional analysis of lines along the laser’s motion to detect 1D signatures of thermal blocking defects. To evaluate defects’ signal/noise ratio, we introduce a global image noise estimation that is robust to outliers (defects or acquisition artifacts). In order to discriminate further between real defects and image artifacts, we also introduce a symmetry measure. This one-dimensional approach may lead to partial detection of noisy objects: in a second step, we reconstruct defects in the orthogonal direction, with a controlled tolerance to noise. Finally, each connected component of the resulting mask is associated different criteria, among which area, width, signal/noise ratio, or symmetry. It is then possible to interactively select defects according to these measures, for instance eliminating false positives or keeping only large enough defects. The methodology we propose has been implemented and confronted to experts’ annotations, exhibiting good performance for different types of defects on samples of different materials.