[Paper Review] Unlabelled Far-field Deeply Subwavelength Superoscillatory Imaging (DSSI)

Tanchao Pu, Vassili Savinov+3 more|arXiv (Cornell University)|Aug 2, 2019

Digital Holography and Microscopy23 references4 citations

TL;DR

This paper introduces Deeply Subwavelength Superoscillatory Imaging (DSSI), a far-field imaging technique that reconstructs subwavelength object features using intensity patterns from scattered light under superoscillatory illumination. By training a convolutional neural network on simulated scattering events, DSSI achieves resolution beyond the diffraction limit—demonstrated at λ/200 for a dimer of subwavelength particles.

ABSTRACT

Recently it was reported that deeply subwavelength features of free space superoscillatory electromagnetic fields can be observed experimentally and used in optical metrology with nanoscale resolution [Science 364, 771 (2019)]. Here we introduce a new type of imaging, termed Deeply Subwavelength Superoscillatory Imaging (DSSI), that reveals the fine structure of a physical object through its far-field scattering pattern under superoscillatory illumination. The object is reconstructed from intensity profiles of scattered light recorded for different positions of the object in the superoscillatory field. The reconstruction is performed with a convolutional neural network trained on a large number of scattering events. We show that DSSI offers resolution far beyond the conventional 'diffraction limit'. In modelling experiments, a dimer comprising two subwavelength opaque particles is imaged with a resolution exceeding $λ/200$.

Motivation & Objective

• To overcome the diffraction limit in far-field optical imaging by exploiting superoscillatory fields.
• To enable high-resolution imaging of subwavelength objects without near-field detection or label-based methods.
• To develop a machine learning-based reconstruction framework that operates solely on far-field intensity measurements.
• To demonstrate subwavelength resolution in a realistic, experimentally feasible imaging setup using deep learning.

Proposed method

• Illuminates a subwavelength object with a superoscillatory field that exhibits highly localized intensity maxima below the diffraction limit.
• Records far-field intensity patterns of the scattered light as the object is translated across the superoscillatory field.
• Trains a convolutional neural network (CNN) on a large dataset of simulated scattering events to learn the mapping from intensity profiles to object structure.
• Uses the trained CNN to reconstruct the object from measured far-field intensity data without requiring phase information or prior knowledge of the object.
• Employs a data-driven approach to bypass the inverse scattering problem's ill-posedness by leveraging the representational capacity of deep learning.
• Validates the method through numerical simulations of a dimer composed of two subwavelength opaque particles.

Experimental results

Research questions

• RQ1 Can deep learning be used to reconstruct subwavelength object features from far-field intensity patterns under superoscillatory illumination?
• RQ2 To what extent can DSSI surpass the conventional diffraction limit in far-field imaging?
• RQ3 How does the resolution of DSSI compare to conventional optical imaging techniques in the absence of near-field detection?
• RQ4 Can the method reliably reconstruct complex subwavelength structures using only intensity measurements and a trained neural network?
• RQ5 What is the minimum resolvable feature size achievable with this approach in a realistic simulation setting?

Key findings

• DSSI achieves a spatial resolution exceeding λ/200 in numerical simulations, significantly surpassing the conventional diffraction limit.
• The method successfully reconstructs the fine structure of a dimer composed of two subwavelength opaque particles using only far-field intensity measurements.
• The convolutional neural network is able to generalize across diverse object configurations and accurately recover subwavelength features without requiring phase information.
• The approach operates entirely in the far field, avoiding the need for near-field probes or complex interferometric setups.
• The reconstruction quality remains robust even under moderate noise levels, indicating potential for experimental implementation.
• The results demonstrate that superoscillatory fields combined with deep learning enable a new paradigm for high-resolution optical imaging beyond classical limits.