A device for detection of surface defects on a terminal surface of an optical fiber. The device includes a digital microscope configured to capture an image of a terminal surface; and a mechanism for analyzing the image configured to detect surface defects present on the terminal surface, the analysis mechanism integrating a “U-Net”-type neural network having had its training phase carried out via an enhancer. The enhancer is configured to create training images, intended to train the neural network, based on reference images. The training images are obtained from the reference images by only applying flips, rotations and/or luminosity, contrast, or shade variations to the reference images.