MECH5720 Chap.1 Computational Imaging Systems
Computational Imaging Systems
Computational imaging designs optics, sensing and algorithms as one measurement system. The central object is a forward model that predicts sensor data from a scene. Calibration tests whether that model matches the assembled camera, while reconstruction estimates the scene or task variable from noisy measurements.
The chapter develops a reusable chain from scene and encoding through detector, inverse method and validation, with special attention to identifiability, conditioning and bounded claims. Begin by defining the task. A camera intended to recognise an object, estimate depth or reproduce a visually faithful image may need different measurements even when all three observe the same scene.
The unknown vector should therefore represent what the task requires, and the measurement operator should expose the variables that distinguish plausible answers. A high-resolution detector cannot compensate for an encoding that maps two important scene states to the same data. Linear notation y = Ax + n makes the logic explicit.
Columns of A describe how individual scene elements affect the detector, while rows describe individual measurements. Rank tests noiseless uniqueness. Singular values show whether some combinations are weakly observed and will amplify noise or calibration error during inversion. These are separate questions: a full-rank operator may remain too ill conditioned for a reliable physical system.
Calibration turns intended geometry into a usable operator. A point target, known illumination pattern or reference pose can reveal response, alignment, gain and offset. The calibration should match depth, wavelength, field position and exposure where those variables change the response.
A residual with repeatable structure is not merely a bad score; it identifies a systematic way in which the predicted measurement disagrees with the camera. Reversing a crop, shifting a mask or using the wrong depth kernel can each leave a characteristic pattern. Reconstruction then needs an inference rule. Direct inversion may be suitable for a small, well-conditioned system.
Regularisation, constrained optimisation or task-specific estimation is required when information is weak or noisy. Any prior contributes assumptions, so recovered features should be checked against the measurement and a target whose answer is known. Parameter sweeps, repeated captures and complementary measurements test whether a conclusion survives reasonable changes. The final claim must match the validated range.
Evidence from one central calibration plane does not establish performance across every depth or field location. A credible engineering account states the model, records the acquisition, shows the control, quantifies uncertainty and names the situation in which the method stops working.
What this chapter covers
- 01
Task-led camera design
- 02
Forward measurement models
- 03
Calibration and validation
- 04
Identifiability and conditioning
- 05
Residual-based diagnosis
- 06
Information loss and recovery
Worked example · free
Separate uniqueness from stability
- +1Write the measurement matrix with rows [1, 1] and [1, 1.01].
- +1Its determinant is 0.01, so the matrix is invertible and the noiseless solution is unique.
- +1The nearly parallel rows provide almost repeated information, so small measurement errors cause large solution changes.
- +1Choose codes with more distinct response vectors or collect a complementary measurement.
- +1Validate the new design by comparing singular values and reconstruction error under realistic noise.
Key terms
- Forward model
- A map from scene variables to predicted measurements.
- Identifiability
- The ability to distinguish unknown scenes from the available measurements.
- Conditioning
- The sensitivity of an inferred solution to errors in data or model.
- Calibration
- Measurement of system parameters needed by the forward model.
- Residual
- The difference between observed data and model prediction.
Computational Imaging Systems FAQ
What makes imaging computational?
The optics and sensor produce an encoded measurement that is interpreted with an explicit model and algorithm. Computation is not merely post-processing; it is part of how the system captures and separates information for a defined task.
Why is calibration part of the model?
A design specifies intended geometry, while calibration measures the response of the assembled system. Misalignment, aberration, detector offsets and environmental conditions can change the operator, so reconstruction should use validated parameters rather than an ideal diagram alone.
Can regularisation recover missing information?
Regularisation can select a stable, plausible estimate when measurements are noisy or incomplete, but it does not create uniquely measured detail. A feature supported mainly by the prior should be labelled uncertain and tested with a complementary capture.
Exam move
Draw the scene-to-measurement chain for every system. Label known variables, unknown variables, calibration data and noise. For each inverse problem, ask separately whether the scene is identifiable and whether recovery is well conditioned. Finish by predicting a residual pattern for one model failure and naming the control capture that would test it.
Build one small numerical matrix and change two nearly independent rows until they almost coincide. Track the determinant, smallest singular direction and recovery change under a fixed perturbation. Next, take a calibration fault such as a shifted crop and sketch the structured residual it should create. Practise writing conclusions that name the tested depth, field and noise range.
End each session by explaining why a visually plausible estimate may still be unsupported by the measurement. Add one alternative system design and state which new measurement direction makes its inverse more stable. Verify the revised design with a simulated perturbation before proposing physical hardware changes.