MECH5720 Chap.7 Lensless Imaging and ADMM
Lensless Imaging and ADMM
Lensless imaging replaces a focusing lens with a coded element and an explicit operator y = Ax + n. Reconstruction must apply the forward operator and its adjoint accurately, then balance sensor agreement with a prior that stabilises the inverse problem. ADMM introduces a split variable so the data term and regulariser can be updated separately.
Data, primal and dual residuals reveal different failures and should be interpreted alongside validation targets rather than a final image alone. Each scene point produces a spread pattern on the detector instead of a focused point. The operator can be represented by a matrix in a small problem, but realistic systems rely on convolution, separable structure or fast transforms to apply A without storing every coefficient.
The adjoint maps a measurement residual back into scene coordinates. It is a backprojection, not an inverse, and should pass an inner-product test that checks flips, crops and scaling. Calibration determines the actual operator. Mask-to-sensor distance, scene depth, wavelength, detector crop, background and exposure can all alter the response.
Applying A to a known target should predict its measurement before an optimiser is trusted. Structured residuals can reveal shifted calibration, a wrong depth model or a missing background term more directly than a visually plausible reconstruction. A regularised objective combines data agreement with assumptions about the scene.
Sparsity can favour a small number of transform coefficients, total variation can favour piecewise smooth structure, and non-negativity can impose physical intensity. Each assumption can erase real features that violate it. Validation therefore needs targets with narrow lines, smooth regions and contrasts resembling the intended task. ADMM splits the estimate x from an auxiliary variable z constrained to agree with it.
The x-update handles the measurement operator, the z-update applies a proximal form of the prior, and the scaled dual update accumulates disagreement. Lambda determines prior strength in the original objective; rho influences how the numerical split is negotiated. Their roles should not be conflated. Stopping requires more than a fixed iteration number.
The data residual measures sensor agreement, the primal residual measures x minus z, and the dual residual reflects change in the split. A decreasing measurement error can still accompany noise fitting or model mismatch. Track a validation score, relative iterate change and residuals together, while saving intermediate images. Parameter stability is part of the evidence.
Sweep lambda, rho and stopping criteria under fixed normalisation, then report whether recovered features persist. If a small change creates or removes a claimed structure, mark it uncertain. An auditable result includes the operator calibration, objective, initialisation, parameter scale, iteration history and the exact measurement that the estimate predicts.
What this chapter covers
- 01
Lensless forward operators
- 02
Adjoint backprojection
- 03
Regularised objectives
- 04
ADMM variable splitting
- 05
Primal and dual residuals
- 06
Stopping and parameter selection
Worked example · free
Perform one scalar ADMM cycle
- +1Differentiate the x objective to obtain (x-3) + (x-1) = 0.
- +1Solve 2x - 4 = 0, giving x = 2.
- +1Soft-threshold 2 by 0.5 to obtain z = 1.5.
- +1Update u to 0 + 2 - 1.5 = 0.5.
- +1Interpret the primal residual x-z = 0.5 as remaining disagreement.
Key terms
- Lensless camera
- A camera that uses coded propagation and computation without a focusing lens.
- Adjoint operator
- A backprojection mapping measurement-space values to scene coordinates.
- Regulariser
- A penalty that favours selected structure in an inverse estimate.
- Proximal step
- An update that balances closeness to an input with a regularisation penalty.
- Primal residual
- The disagreement between variables constrained to be equal in ADMM.
Lensless Imaging and ADMM FAQ
Is the adjoint the inverse?
Not in general. The adjoint backprojects measurement information into scene coordinates, but it does not undo blur, mixing or information loss unless the operator has special orthonormal structure.
What do lambda and rho control?
Lambda sets the strength of the prior relative to data agreement in the reconstruction objective. Rho controls the ADMM penalty that coordinates split variables and can change convergence behaviour without redefining the target objective.
When should iteration stop?
Monitor primal and dual residuals, relative iterate change and a validation metric on comparable data. Stop when convergence criteria are met and validation no longer improves, rather than relying on visual sharpness or an arbitrary iteration count alone.
Exam move
Write the objective before memorising updates. Label which ADMM step serves the data term, which applies the prior and which accumulates disagreement. Trace one scalar example by hand, then connect structured residuals to alignment, crop, depth or calibration faults in the forward operator. Verify an adjoint with the inner-product identity before analysing convergence.
Plot data, primal and dual residuals on the same iteration axis and explain what each can reveal. Compare a weak and strong prior on targets containing both smooth regions and narrow lines. Preserve intermediate estimates so overfitting is visible. Finish with a reproducibility list covering operator version, normalisation, lambda, rho, initialisation, iteration count and stopping condition.
Predict how a wrong-depth operator would differ from an excessively strong sparsity prior in the residual.