"Single Image Wavefront Sensing Using AI Based Phase Retrieval"
Wavefront (WF) sensing is essential in many fields such as adaptive optics (AO), qualitative ophthalmic measurement or optical testing and alignment, to name few. For instance, in the context of telescopes, especially reflectors, alignment of the optical surfaces (mirrors and field lenses) is critical for reaching the diffraction limit.
Both AO for Earth atmospheric turbulence (seeing) compensation as well as real time (live) telescope optical alignment use actual or artificial stars and typically dedicated hardware for pupil, or near pupil, space WF sensing (i.e. Shack-Hartmann, Pyramid WF sensors, curvature sensing).
Image space WF sensing offers a very appealing alternative with a minimum or no extra hardware leading to a common optical path with the image plane which removes the need of a multi-path calibration and related WF reconstruction residual errors. Phase diversity has been widely used for image space WF sensing. This approach relies on, at least, two defocused images taken at two different defocus positions near the focal plane and iterative phase estimation algorithms at run time.
This lecture presents an image space WF seeing approach using AI for phase retrieval from a single defocused image, even under seeing limited conditions and noisy environments. It has 3 very desirable features:
- Requires only a single image.
- Fast processing (no iteration) at run time.
- Whole field WF sensing.
An artificial neural network (ANN) is trained with only synthetic data, simulated aberrated defocused star images, computed using the scalar diffraction theory. The ANN is a function approximation tool mapping single defocus images to their WFs, usually expressed in the form of Zernike annular polynomial coefficients. As a result, at run time, there is no iteration nor optimization and convergence concerns for phase retrieval calculation anymore, all the heavy work has been done during the training of the ANN. When presented with a new single defocused image of a star (artificial or natural) the ANN outputs the related WF data, (i.e. the Zernike annular polynomial coefficients). The feed forward structure of the chosen ANN leads to very fast phase retrieval calculations, consistent with video rate WFS sensing, a must for AO.
By nature, this approach can access WFs in the whole field, at once, when presented with a defocused star field image. Beside using an actual star, a point source far away (aka. a plane wave), various, even extended, types of sources, like spots, can be used with this method as long as their diffraction patterns can be calculated. For instance, the AO laser artificial guide stars created in the upper Earth’s mesospheric sodium layer results to spherical wavefronts which can be easily accounted for. Although our current application of this AI based WF sensing technology, is in the field of telescopes and astronomy it has much broader applications.
Gaston was born in Lausanne Switzerland along the Geneva’s Lake. He received his bachelor’s degree in electronics and mechanical engineering from the Lausanne Institute of Technology (ETML) in 1979 and his master’s degree in electrical engineering and computer science from the Swiss University for Applied Sciences (HEIG-VD) at Yverdon-les-Bains in 1982. Gaston followed a postgraduate course in biological and artificial neural network at the Swiss Polytechnical School of Lausannne (EPFL) from which is graduated in 1993 after accomplishing a project in the context of speech recognition. This was for him a great source of inspiration and passion for using artificial intelligence (AI), including in embedded applications and products. In 2006 he received a PhD’s degree in computer science in the field of machine learning from the French National Conservatory of Arts and Crafts (CNAM) at Paris.
Gaston has been deeply involved in optics, opto-electronics, and analog/digital sensor designs, as well as document sensing and pattern recognition using hyper spectral data, digital image processing, advanced statistics, and AI, leading to very successful products and filing many patents. He also published papers in scientific journals and conferences in statistics and machine learning.
In 2000 Gaston eventually moved to the US becoming the senior research director for MEI (Mars Electronics International) and later for CPI (part of Crane C.O.) from which he retired in 2018. Following his passion for astronomy he co-founded in 2007 Innovations Foresight for making innovative technologies and related products, such as the on-axis guider (ONAG) device for real time (live) focusing and full frame auto-guiding or the wavefront sensing using AI technology, for which all have been granted patents. Gaston has given many lectures on astronomical topics stretching from optimal guiding and focusing under seeing limited conditions to telescope optical alignment and AO using wavefront sensing. Since 2018, as the president of Innovations Foresight, Gaston is focused to deliver new cutting-edge innovative solutions in the fields of astronomy and optics. Meanwhile Gaston is still actively involved in many projects he has proposed over time with students and professors from various institutions (EPFL, HEIG-VD, ETML…).