Abstract: In detecting surface defects of a hot metal body, e.g. hot-rolled steel sheet, by using a first video signal derived from radiation spontaneously radiated from a target area of the hot metal surface and a second video signal derived from external light projected on and reflected from the same target area of the hot metal surface, the rate of false detection by the influence of noise signals attributed to metal oxide present on the hot metal surface is greatly reduced by converting the first and second video signals respectively into first and second binary signals, synchronously feeding the first and second binary signals to an AND circuit and detecting the locations of such defects based on the output of the AND circuit. Preferably near-infrared region of the radiation is used in producing the first video signal and visible light as the external light.