Fig. 1. Schematic representation of the proposed U-Net model for defect detection. The boxes correspond to multi-channel feature maps, connected by different operations (denoted by arrows). The length and height of each box represent the number of filters ($N=32$) and the $x$–$y$ size, respectively.

Fig. 2. The overall architecture to train the model for detection of optical defects in real time.

Fig. 3. Examples of a potential damage site classified as: (a) real defect; (b) hardware reflection; (c) reflection from the exit surface (marked in the box); (d) light spot.

Fig. 4. Schematic diagram of the methodology in obtaining the online and offline images of the final optics.

Fig. 5. An example of the prepared training dataset: (a) the cropped region from the online image; (b) the matched region of (a) in the offline image; (c) the 0–1 mask created by (b), with 1 for real defect and 0 for background.

Fig. 6. Intensity distribution of the training samples (in log scale).

Fig. 7. The curves of training and validation loss with respect to the number of iterations. We used a learning rate of $10^{-3}$ for the first 500 iterations and changed the learning rate to $10^{-4}$ for later iterations.

Fig. 8. Predictions of real defects by the trained model on the test images. (a) The online image of an inspected optic. (b) 0–1 mask created by the offline images of the same inspected optic. (c) Predicted mask by the trained U-Net model. Bottom panels show a zoom-in on a highly contaminated region.