Scratches on the surface of optical components have serious impacts on optical system such as imaging quality of lens and/or mirrors in optical imaging systems, light-collecting abilities of laser fusion and solar concentrator systems. The size of the scratches is a key issue for analyzing and assessing the impacts quantitatively. Most of the available testing methods for scratches depend on human visual inspection (HVI) with naked eyes by workers, which leads to low efficiency and accuracy. This paper presents an automatic detecting method for the scratches on optical surface with machine vision inspection (MVI) method. The microscopic dark-field scattering imaging system is used as the front end of the detection system. A dedicated algorithm is designed for non-closing scratch detection. The core merits of this algorithm lies in three folds: 1) automatic processing capabilities, which includes positioning, clustering, and precise estimation of the length of the scratches; 2) high efficiency, which is characterized by a short time interval, i.e., about 0.138 second per binary image with 2724 × 2724 pixels in our experiments; 3) high accuracy, where the error rate of the total length of the scratches detected is less than 5% when compared with the nominal visual measurement result obtained via HVI method. The proposed scratch detecting algorithm can be used for non-destructive testing (NDT) of the glass-like surfaces.