A defective part recognition device includes a microscope for obtaining a magnified image of a unit area for recognizing a defective part on the surface of a multi-layer film substrate; a spectral camera having an imaging surface where the magnified image is formed; and an information processing part for processing the spectrum information from the spectral camera. The information processing part includes a machine learning part for a clustering process on the spectrum information for each pixel, and a defect recognition part for recognizing a defective part from the result of the machine learning part. The machine learning part sets a cluster in the unit area and generates a histogram with a frequency, the number of pixels clustered into the cluster. The defect recognition part compares the frequency distribution of the generated histogram with that of a histogram free of defects and recognizes a defective part.