The system is applied to the non destructive testing of aluminium ingots by digital radioscopy. Internal defects are automatically detected and their type identified, while the ingot is moving at 10 m/min. the image is acquired in real time by a linear X-ray detector (THOMSON TE 1487). the image analysis consists of three steps: defect detection, feature extraction and recognition. Seven defect types are to be considered (cracks of different orientations, porosities, cavities, linear shrinkages, microshrinkages).

The detection method is mainly based on mathematical morphology. First, an adaptive thresholding method by the morphological top hat transform is applied, allowing a reliable detection of the defects but also the noise. Then, a procedure using the median filter (and/or a morphological erosion of the objects) permits the noise reduction, but in return affects the size of the defects. finally, a morphological reconstruction allows to recover the original defects, without the noise, by dilating the objects in the eroded image, conditionally to the image obtained after the top hat transform. Depending on the parameters of the top hat, one can obtain either an over-segmented image (almost all the defects are detected but with a bigger size) or a sub- segmented image (defects are detected with their real size, but the smallest defects are missed). The best compromise is chosen depending on the recognition procedure. For feature extraction, a study of the ingot production and of the defects forming allows to model the defect types, and to gather as much a priori information as possible. The selected features for defect modelling are shape, orientation, width, location and contrast.

The knowledge gathered from the radiographic expertise was used to build the recognition methodology. A sub-segmented image proved adequate because the various defect types are well separated, although some are missed. The recognition method is based on an identification tree, of which inputs are the features, and outputs are the defect types. Once the objects are separated in shape, orientation and width families, a clustering of aligned and neighbouring objects is performed, in such a way to gather the objects that would have been disconnected by segmentation process. Then, the features of the clustered objects are compared to those of the models, yielding the defect type identified.

As some a priori information is used in the recognition tree, this algorithm is specific to the aluminium ingot production. Nevertheless, a much wider range of defects was considered in the defect modelling, in particular the welding defects, and an interesting future application on welds is planned.