Deep Learning Methods Based on ResNet50 Architecture for Wave Aberration Identification

Abstract In this paper, we consider deep learning methods based on the ResNet50 architecture for automatic identification of both single wave aberrations of optical systems and superposition of two types of aberration. The use of residual connections provides a high level of accuracy in classification of aberration types (defocusing, 1st and 2nd order astigmatism, coma, trefoil, quatrefoil and spherical) and their magnitudes from intensity distributions with several diffraction orders obtained using multichannel DOEs. Numerical experiments demonstrate the effectiveness of the proposed approach in automated recognition, which makes it promising for real-time use in adaptive optics and ophthalmology. Based on network training over 20 epochs, it was found that on the validation sample, the average absolute error of aberration recognition does not exceed 0.0056.