PURPOSE: To shorten the discriminating time and to improve the discriminating capacity by selecting a neural network which can input all flaw signal waveform range and has an input layer in which the input of a non-flaw signal waveform becomes minimum.

CONSTITUTION: When 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 reflected light is changed in its luminance according to the uneven surface of the plate, and the luminance change is detected as an electric signal. A neural network 2 which calculates the flaw range of the surface of the plate according to a waveform selector 15, inputs the entire flaw signal waveform from the calculated flow signal waveform range and has an input layer in which the input of the non-flaw signal waveform becomes minimum is selected by the detection signal of the sensor 13. The entire flaw signal waveform is input to a neural network arithmetic processor 16, which conducts the neural network calculation based on the input signal waveform and executes the neural network calculation according to the entire flaw signal waveform necessary to discriminate the flaw occupying the main region of the flaw signal waveform range.