Research on size calibration algorithm of dark-field flaw on optics surface based on convolutional neural network

Optics in high-power solid-state laser devices will be damaged and adhere to contaminates on the surface during manufacturing and operation, which reduces the service life. To facilitate the positioning and recognition of the flaw （damage and contaminates） in the process of automatic repair, the dark-field flaw size calibration algorithm was studied. In this paper, a flaw size calibration model based on convolutional neural network was established. The size, exposure time, and loss function of damage and contaminate were optimized respectively. The training parameter of the damage regression model is determined to be size of less than 500μm, 30ms of exposure time and SmoothL1 loss function, and the parameter of contaminate is 0-100μm, 20ms of exposure and SmoothL1 loss function. After data balancing, the calibration precision was improved. To improve the calibration consistency of the full-size segment, a comprehensive size calibration strategy is proposed. The CNN regression algorithm is used for small size, and the pixel-level calibration algorithm is used for large size. Compared with the single regression calibration algorithm, the relative error and absolute error of the calibration results of small damage (<340μm) are reduced by 41.4% and 44.13% respectively. The relative error and absolute error of small contaminate (<85μm) are reduced by 40.26% and 30.63%.