PURPOSE: To shorten the discriminating time and to improve the discriminating capacity by mask processing a flaw signal waveform region to be input to a neural network so that a non-flaw signal waveform becomes '0'.

CONSTITUTION: A light from a rodlike light source 12 is reflected on the surface of a steel plate and incident on a line sensor 13. The flaw range of the surface of the plate is extracted from the detection signal of the sensor 13 by a mask generator 15. The generator 15 so generates a mask in which a non-flaw signal waveform is '0' as to exclude the non-flaw signal waveform included in the flaw signal waveform region to be input to the unit of a neural network arithmetic processor 17 from the extracted flaw signal waveform range. The flaw signal waveform which is not masked and the non-flaw signal waveform which is masked to become '0' are input to the unit of the processor 17. The processor 17 conducts the neural network calculation based on the signal waveforms input to the unit and executes only the neural network calculation for the flaw signal waveform necessary to discriminate the flaw.