I. Mazur et alii, Frattura ed Integrità Strutturale, 37 (2016) 287-296; DOI: 10.3221/IGF-ESIS.37.38 290 group 3 – rolled metal of dynamo and alloyed steel grade; group 4 – rolled metal S355J2G3, St5sp, 09G2 and A36; group 5 – rolled metal with corrugation, and a threshold valuations of contrast were defined for each group and classification spheres were done . A presence of drops of water on the upper side of rolled metal was one more problem for the detection of surface defects on the upper side of rolled metal and also a presence of water splashes and vapor between the flat bar and upper video. Not less than 50-100 images of water will be classified as «defect» at the presence of 4-5 thousand images of water on one flat bar even at the high rate of classification (98-99 %). The design of water cooling of the roll-table rollers mounted in the zone of measurement shelters was altered for the elimination of water and another hydraulic blow-off assigned for the water drops moving off from the surface of flat bars with the width of more than 1200 mm., was installed in addition to the regular hydraulic blow-off after the fifth 5 roller and air blow off for the rolled metal with the width of 900-1200 mm. was installed after the seventh roller. It became possible to go to the next stage – a training of the system for the automatic classification of defects in the result of the setting of ASIS for the detection of the surface defects of rolled metal and the elimination of negative influence of water splashes for the detection. And though the use of data base of knowledge of standard images of defects obtained on different metallurgical plants was possible for this the stage of setting of ASIS for the automatic classification was performed in the manual mode. This is more laborious method of setting connected with the stoppages of units but effective because precision of classification of unknown defects is increased due to visual inspection. The examination of rolls containing the defects of unknown type was realized on the units of cutting and lines of preparation of hot-rolled rolls. An identification of defects, and their documenting and digital photographing was carried out by the experts of quality. A complex of metallographic examinations was carried out when it was necessary. A primary training of ASIS for the automatic classification of defects was carried out from the December 2003 up to the May of 2004. About 120 rolls with defects were examined altogether at that period of time and more than 1000 image of defects were entered into the data base of knowledge. It became possible to carry out the training of the system without carrying out of laborious visual examination of rolls or resort to it in the exceptional cases after the accumulation of a definite experience of work with the images of defects obtained from the system. A catalogue of digital images of defects developed by LSTU assisted to it (Fig. 2). Accumulated data base of knowledge ASIS and digital images of defects detected in the result of visual examination and identification of defects of surface of hot-rolled metal became a basis of that catalogue. The guarantee tests were carried out (June 2004) jointly with the participation of companies SYTCO AG and SIEMENS- VAI when the volume of trained images of defects became sufficient. The rolls with the surface defects detected by the system were selected for it. The selected rolls were examined according to the method analogous to the method used for the training of ASIS to the automatic classification. The deffects were registered manually on the units of preparation of hot-rolled rolls. A coordinate of defect from the trailing end of roll, the dimensions of the defect were fixing and its class and tenseness were defined. Then, a rate of detection d P and a rate of classification c P according to the results of manual and automated detection of deffects of surface of hot-rolled metal was calculated: %100 e d d N N P , %100 d c c N N P , where: d N – the number of deffects of the given class detected simultaneously by the system in automatic mode and by the expert at the visual examination on all the rolls the selected for the test; e N – the total number of deffects of the given class detected by the expert at the visual examination on all rolls the selected for the test; c N – the number of deffects of the given class detected simultaneously detected by the system in automatic mode and by the expert at the visual examination on all the rolls the selected for the test and classified correctly by the system (as the given class) at that.