Abstract: There are some scratches of shallow depth (about 25 nm) or narrow width (collectively referred to as weak scratches) during defects detection on the super-smooth optical components. In the dark field imaging detection, the scattered light of the weak scratches produces such a low gray value, even masked in the background light, as a result, they are very difficult to identify by visual or conventional machine vision, and could be undetected. To solve this problem, based on the existing surface defects evaluation system (SDES), according to the scratches gray features, the dual-threshold classification method to extract scratches defects is proposed. In low threshold scratches processing, according to the frequency characteristics of the weak scratches, background contrast and spatial characteristics, the algorithm of frequency domain filtering with background difference is adopted. By spatial domain and frequency domain filtering algorithms, interference points and high luminance noise points are eliminated. In addition, background can be extracted since its geometric characteristics are different from weak scratches. Further, weak scratches with enhanced contrast are extracted, and finally, these scratches are used in high threshold information extraction of subsequent scratches feature together with the normal gray level scratches. Thus all the scratches information is obtained for calculating the total length of scratches and grading of defects maximum length. Experimental results show that the algorithm proposed avoids introducing irregular background noise by a low binarization threshold, and also improves the contrast of scratches and background. At present, the algorithm has been applied to inertial confinement fusion systems in large aperture optical surface scratches quantitative detection, and the length detection accuracy has been raised to about 80%.