According to the requirements and characteristics of the hard resin lens on-line detection, an on-line defect detection method based on machine vision is proposed. An on-line detection system with two image acquisition structures is designed to detect all of the defects with different sizes. The first image acquisition system is used to detect the characteristics of the defects such as the point-like impurities and bubbles, and to obtain the information, including the length, amount and location of the scratches, and the location of sensitive areas contain the scratches and feathery impurities, and to control the working status of the second image acquisition system. The second image acquisition system is used to detect the diameter of the scratches and feathery impurities, after locating the position of sensitive areas with the movement of the conveyor belt and the one-dimensional linear guiding of the second image acquisition structure. This proposed method is able to solve the technical problems such as high cost, low-detection-speed and mass-data, which meets the needs of online detection. Experiments show that detecting speed of the system is 1.5 s / piece. Compared with subaperture stitching, the least amount of data is reduced by about 88% and a decrease of approximately 96% at most.