High accuracy target point classification through dark-field detection of large-aperture optics surface based on machine vision

Surface defects on optics seriously affect their service life in high-power laser system. To address this issue, repair technologies for large-aperture optics are extensively employed. However, the operation of high-power laser system generates a large number of optics requiring repair. Currently, microscopic inspection is widely used for target points classification due to its high accuracy, but it has limitations such as small field of view and low efficiency. Utilizing the OTSU algorithm and Convolutional Neural Networks (CNN), a dark-field classification model for optical surface target points has been developed. Upon conducting experiments with several network models, the ResNet model, a probability threshold of 0.998 were determined. In conclusion, the classification accuracy for defects was 99.18% and for contaminants was 90.34%. The outcomes show that target points on the surfaces of optics can be accurately classified using this method.