Lensless (Diffuser Camera) Imaging

lensless Microscopy Lensless Computational Incoherent
View Benchmarks (1)

Lensless imaging replaces the objective lens with a thin optical element (phase diffuser or coded mask) placed directly near the sensor. Scene light produces a multiplexed caustic pattern encoding the entire scene. The forward model is y = H * x + n where H is determined by the mask's phase profile and mask-to-sensor distance. Each scene point contributes across many sensor pixels, yielding a multiplexing advantage. Reconstruction solves a large-scale inverse problem via ADMM or FISTA with total-variation or learned priors.

Forward Model

Psf Convolution Or Linear Operator

Noise Model

Poisson Gaussian

Default Solver

admm tv

Sensor

CMOS

Forward-Model Signal Chain

Each primitive represents a physical operation in the measurement process. Arrows show signal flow left to right.

P diffuser Diffuser Propagation D g, η₁ Bare Sensor
Spec Notation

P(diffuser) → D(g, η₁)

Benchmark Variants & Leaderboards

Lensless

Lensless (Diffuser Camera) Imaging

Full Benchmark Page →
Spec Notation

P(diffuser) → D(g, η₁)

Standard Leaderboard (Top 10)

# Method Score PSNR (dB) SSIM Trust Source
🥇 Uformer 0.768 33.5 0.920 ✓ Certified Wang et al., CVPR 2022
🥈 FlatNet 0.725 31.8 0.890 ✓ Certified Khan et al., IEEE TPAMI 2020
🥉 PnP-ADMM 0.603 27.5 0.790 ✓ Certified Monakhova et al., Opt. Express 2019
4 Wiener-ADMM 0.462 23.5 0.640 ✓ Certified Antipa et al., Optica 2018
Mismatch Parameters (3) click to expand
Name Symbol Description Nominal Perturbed
diffuser_psf ΔPSF Diffuser PSF calibration error (%) 0 5.0
sensor_distance Δd Diffuser-sensor distance error (mm) 0 0.2
wavelength Δλ Wavelength mismatch (nm) 0 5.0

Reconstruction Triad Diagnostics

The three diagnostic gates (G1, G2, G3) characterize how reconstruction quality degrades under different error sources. Each bar shows the relative attribution.

G1 — Forward Model Accuracy How well does the mathematical model match reality?

Model: psf convolution or linear operator — Mismatch modes: psf calibration error, mask sensor misalignment, depth dependent psf variation, stray light

G2 — Noise Characterization Is the noise model correctly specified?

Noise: poisson gaussian — Typical SNR: 10.0–30.0 dB

G3 — Calibration Quality Are instrument parameters accurately measured?

Requires: psf measurement, mask to sensor distance, flatfield, background subtraction

Modality Deep Dive

Principle

Lensless (diffuser-cam) imaging replaces the imaging lens with a thin diffuser or coded mask placed directly before the sensor. The sensor records a multiplexed pattern (caustic or speckle) that encodes the 3-D scene. Computational reconstruction inverts the known point-spread function of the diffuser to recover the image, enabling an extremely compact, lightweight camera suitable for miniaturized or in-vivo applications.

How to Build the System

Place a thin diffuser (ground glass, engineered phase mask, or Scotch tape) at a fixed, small distance (~1-5 mm) from a bare sensor (CMOS, e.g., Sony IMX sensor). Precisely characterize the diffuser PSF by scanning a point source across the field of view. Mount rigidly to prevent any relative motion between diffuser and sensor. For 3-D reconstruction, the depth-dependent PSF must be calibrated at multiple axial planes.

Common Reconstruction Algorithms

  • ADMM (alternating direction method of multipliers) with TV regularization
  • Wiener deconvolution (fast, single-step but lower quality)
  • Gradient descent with learned priors (DiffuserCam, neural network prior)
  • Tikhonov-regularized least squares
  • Unrolled optimization networks (physics-informed deep learning)

Common Mistakes

  • Inaccurate PSF calibration causing reconstruction artifacts
  • Insufficient sensor dynamic range for the caustic intensity peaks
  • Motion between diffuser and sensor during capture invalidating the PSF model
  • Regularization too strong, over-smoothing fine details in the reconstruction
  • Ignoring the depth-dependence of the PSF when imaging 3-D scenes

How to Avoid Mistakes

  • Calibrate PSF carefully with a point source at the exact sample distance
  • Use HDR acquisition or high-bit-depth sensors to capture full caustic range
  • Rigidly bond the diffuser to the sensor; verify alignment stability
  • Tune regularization weight (e.g., via L-curve or cross-validation)
  • Calibrate PSF at multiple depths for 3-D scenes; use depth-varying reconstruction

Forward-Model Mismatch Cases

  • The widefield fallback uses a Gaussian PSF, but lensless cameras use a coded aperture (phase mask, diffuser, or amplitude mask) that creates a highly structured, non-Gaussian PSF — the caustic pattern is fundamentally different from a Gaussian
  • The lensless PSF encodes the scene through a known, shift-variant pattern — the widefield shift-invariant Gaussian blur does not capture the scene-dependent structure of the lensless measurement and produces incorrect reconstruction input

How to Correct the Mismatch

  • Use the lensless operator with the calibrated PSF of the specific coded aperture (measured from a point source or computed from the mask design): y = H * x, where H is the non-Gaussian, possibly shift-variant PSF
  • Reconstruct using Wiener deconvolution, ADMM with TV prior, or learned methods (FlatNet, PhlatCam) that use the correct coded-aperture PSF for the specific mask in use

Experimental Setup

Instrument

DiffuserCam / FlatCam prototype

Sensor

Raspberry Pi HQ Camera (Sony IMX477, 4056x3040)

Pixel Pitch Um

1.55

Diffuser Type

optical diffuser (Luminit 0.5-deg) / coded mask

Diffuser To Sensor Mm

2.5

Field Of View Deg

40

Image Size

2592x1944

Signal Chain Diagram

Experimental setup diagram for Lensless (Diffuser Camera) Imaging

Key References

  • Antipa et al., 'DiffuserCam: lensless single-exposure 3D imaging', Optica 5, 1-9 (2018)
  • Asif et al., 'FlatCam: Thin, Lensless Cameras Using Coded Aperture', IEEE TCI 3, 384-397 (2017)

Canonical Datasets

  • DiffuserCam lensless mirflickr dataset (Monakhova et al.)
  • PhlatCam benchmark (Boominathan et al., IEEE TPAMI 2022)

Benchmark Pages