Machine Learning Based Localization and Classification with Atomic Magnetometers

Abstract

We demonstrate identification of position, material, orientation and shape of objects imaged by an 85 Rb atomic magnetometer performing electromagnetic induction imaging supported by machine learning. Machine learning maximizes the information extracted from the images created by the magnetometer, demonstrating the use of hidden data. Localization 2.6 times better than the spatial resolution of the imaging system and successful classification up to 97% are obtained. This circumvents the need of solving the inverse problem, and demonstrates the extension of machine learning to diffusive systems such as low-frequency electrodynamics in media. Automated collection of taskrelevant information from quantum-based electromagnetic imaging will have a relevant impact from biomedicine to security. This is a preprint version of the article appeared in Physical Review Letters: C. Deans, L. D. Griffin, L. Marmugi, F. Renzoni, Phys. Rev. Lett. 120, 033204 (2018).DOI: 10.1103/PhysRevLett.120.033204Electromagnetic induction imaging (EMI), or magnetic induction tomography (MIT), with atomic magnetometers (AMs) was recently demonstrated for mapping the electric conductivity of objects and imaging of metallic samples [1][2][3][4].EMI and its classical counterpart with conventional magnetic field sensors [5] rely on the detection of the AC magnetic field generated by eddy currents excited in media. This poses severe problems for image reconstruction, particularly in cluttered contexts or with lowconductivity specimens. The inherently diffusive and non-linear nature of low-frequency electrodynamics in media makes conventional ray-optics analysis impossible. Consequently, back-projection approaches [6] are of limited use. Furthermore, the solution of the inverse problem for low-frequency electromagnetics is ill-posed, undetermined, and computationally challenging [7]. Ultimately, these limitations reduce the attainable information from EMI and its spatial resolution.In this Letter, we propose and demonstrate machine learning (ML) [8] as a method for enhancing the EMI capabilities and circumventing the problem of image reconstruction and interpretation. ML has thus far been applied in a wealth of fields [9][10][11][12][13][14][15][16][17]. ML-aided security screening in the X band [18,19] and biomedical imaging have been widely demonstrated [20,21], as well as image reconstruction through scattering media in the optical band [22,23]. All these applications to well-established imaging technologies are underpinned by linear systems, with ray-like propagation. * l.marmugi@ucl.ac.ukHere we present proof-of-concept demonstrations of EMI by an AM with metallic and non-metallic samples. Their localization, and material, orientation and shape classification from low-resolution images is aided by ML. AM-EMI supported by ML maximizes the information obtained from the images and provides relevant data for specific tasks, without requiring the inverse problem. This improves or enables identification of critical features, such as structural defec...

Sub-picotesla widely tunable atomic magnetometer operating at room-temperature in unshielded environments

“…The system can be easily extended to included two more loops maintaining B x = B y = 0. We found that this approach does not further improve the sensitivity though it is inherently more suited to long term data acquisition and field applications 30 .…”

Sub-Sm–1 electromagnetic induction imaging with an unshielded atomic magnetometer

“…If required, the spatial resolution could be enhanced via machine learning localization and classification algorithms. 31 The minimum imaged conductivity is improved by more than 50 times with respect to previous results 20 and by a factor 4 with respect to the recently reported detection of low-conductivity solutions in a shielded environment. 21 However, when considering the sensitivity of electromagnetic induction imaging instrumentation, the effective volume supporting eddy currents and thus contributing to the generation of the secondary field is also relevant.…”

Planar array magnetic induction tomography further improvement

“…While anisotropic TV can regularize differently along the depth direction, maybe a stronger prior may be needed to recover the loss of sensitivity with depth. In future work, we may consider machine learning, which has been recently proposed for learning inverse problems and seems to be a good candidate to model and correct for depth in planar MIT (Deans et al, 2018) (Tables V and VI ).…”