Improving the reliability of deep learning computational ghost imaging with prediction uncertainty based on neighborhood feature maps.

Defect inspection is required in various fields, and many researchers have attempted deep-learning algorithms for inspections. Deep-learning algorithms have advantages in terms of accuracy and measurement time; however, the reliability of deep-learning outputs is problematic in precision measurements. This study demonstrates that iterative estimation using neighboring feature maps can evaluate the uncertainty of the outputs and shows that unconfident error predictions have higher uncertainties. In ghost imaging using deep learning, the experimental results show that removing outputs with higher uncertainties improves the accuracy by approximately 15.7%.

Relationship between the kernel size of a convolutional layer and the optical point spread function in ghost imaging using deep learning for identifying defect locations.

We explore the contribution of convolutional neural networks to correcting for the effect of the point spread function (PSF) of the optics when applying ghost imaging (GI) combined with deep learning to identify defect positions in materials. GI can be accelerated by combining GI and deep learning. However, no method has been established for determining the relevant model parameters. A simple model with different kernel sizes was built. Its accuracy was evaluated for data containing the effects of different PSFs. Numerical analysis and empirical experiments demonstrate that the accuracy of defect identification improved by matching the kernel size with the PSF of the optics.

Noise-robust deep learning ghost imaging using a non-overlapping pattern for defect position mapping.

Defect detection requires highly sensitive and robust inspection methods. This study shows that non-overlapping illumination patterns can improve the noise robustness of deep learning ghost imaging (DLGI) without modifying the convolutional neural network (CNN). Ghost imaging (GI) can be accelerated by combining GI and deep learning. However, the robustness of DLGI decreases in exchange for higher speed. Using non-overlapping patterns can decrease the noise effects in the input data to the CNN. This study evaluates the DLGI robustness by using non-overlapping patterns generated based on binary notation. The results show that non-overlapping patterns improve the position accuracy by up to 51%, enabling the detection of defect positions with higher accuracy in noisy environments.