Camera Mechanism Tolerance Allocation for Performance and Cost

Hello. Last lesson examined how a camera corrects motion during an exposure. This final lesson of the module asks a quieter engineering question: when the camera is nominally still and correctly focused, how far can its parts deviate from their intended positions or timings before the photograph suffers?

You will turn optical and exposure requirements into tolerance budgets for focus, recording-surface alignment, and shutter timing. Then you will compare worst-case and statistical predictions, and see how adjustment during assembly can relax some manufacturing requirements without making errors disappear. Allow about 40–45 minutes, including the selected readings.

Begin with the photograph, not the part drawing

A tolerance such as “sensor position within $5\mu \mathrm{m}$” is meaningful only in relation to a performance requirement. The same displacement might be inconsequential for one aperture and output size but unacceptable for another. Likewise, a shutter-time error must be judged against a permitted exposure error, not against an arbitrarily precise clock.

A useful specification begins with measurable image outcomes:

• maximum permitted defocus blur, or minimum through-focus MTF at specified field positions;
• permitted exposure error and variation across the frame;
• permitted displacement or rotation of the recorded frame relative to its intended position.

Next, identify the datums—the surfaces or positions from which those outcomes are controlled. For focus, an important chain runs from the seated lens mount, through the positioned optics, to the actual light-sensitive surface. A dimension on the outside of the camera matters only insofar as it changes that chain.

The distinction between object-side depth of field and image-side depth of focus is particularly useful here. Read the indicated portions of Depth of Field and Bokeh, a technical article from ZEISS. Treat its geometric circle-of-confusion calculation as a starting criterion; its through-focus MTF discussion explains why passing that criterion does not mean image quality is constant throughout the permitted range.

For a nearly paraxial image-space cone, let $N_{\mathrm{im}}$ be the distance from the exit pupil to best focus divided by the exit-pupil diameter. If the recording surface is displaced axially from best focus by $\Delta z$, its geometric blur-circle diameter is approximately

Thus a chosen limit $c_{\max }$ gives an approximate one-sided axial allowance

At ordinary photographic distances, $N_{\mathrm{im}}$ is often close enough to the marked f-number for an initial estimate. At close focus or with unusual pupil geometry, use the actual image-space cone instead. The blur limit is a design choice tied to the intended output or an MTF requirement; it is not an intrinsic property of the sensor.

Allocate an axial focus and alignment budget

Consider an illustrative digital camera operated near f/2.8. Suppose its image-quality specification permits no more than an $8\mu \mathrm{m}$ geometric defocus circle—about two pitches if the pixel pitch happens to be $4\mu \mathrm{m}$. The approximate one-sided axial limit is then

The question for mechanical design is not “Which single part gets $22.4\mu \mathrm{m}$?” Several deviations can move the sensitive surface relative to the lens’s best-focus surface. With signs assigned consistently, a local linear model is

Here $e_0$ is the axial offset common to the frame; ${\alpha }_x$ and ${\alpha }_y$ describe relative tilt between the recording surface and best-focus surface; and $b(x,y)$ captures departures such as sensor bow or film curl. Each term can itself depend on several manufactured dimensions. Nominal lens field curvature belongs in the optical design; the tolerance question is how manufacturing and assembly change the intended relationship between that image surface and the recorder.

For an example worst-case allocation at a specified frame edge, take the following one-sided bounds. These are illustrative budgets, not measured capabilities of a particular camera.

If every contribution can take the sign that worsens this particular point, the total is $20\mu \mathrm{m}$. That leaves $2.4\mu \mathrm{m}$ between the allocated extremes and the $22.4\mu \mathrm{m}$ limit. At the extreme, the estimated defocus circle is $20/2.8$, or about $7.1\mu \mathrm{m}$. The slim margin is a reason to test the assumptions, not to declare the design finished.

Tilt deserves a separate calculation because it affects the frame nonuniformly. If the relevant edge lies $r=12\mathrm{mm}$ from the centre, an edge allowance of $4\mu \mathrm{m}$ corresponds to a small relative tilt of approximately

A centred sensor can therefore be acceptably focused in the middle yet soft toward one side. Translating the sensor axially can remove a common offset; it cannot, by itself, remove tilt or bow.

This is why an adjustable assembly can be more economical than making every component exceptionally precise. In the conference paper Image Acquisition Module with All Plastic Optics, study how the designers assessed performance over a space of manufacturing deviations and provided separate CCD position and tilt adjustment.

In the paper, sensor position and tilt are handled by adjustment rather than charged in full to the unadjusted optical stack. Adjustment does not make their tolerances vanish. It substitutes requirements for adjustment range, resolution, measurement accuracy, locking stability, and assembly time. If the initial error exceeds the adjustment travel, or the setting moves with temperature, the finished camera can still fail.

Film introduces a different recording surface. The gate establishes its intended plane, while the pressure plate helps hold film against that reference. Thickness variation, curl, dirt, and transport forces can make the emulsion depart locally from the plane even when the gate itself is accurately made. A larger acceptable blur circle may permit a looser axial budget than the digital example, but film flatness must be included explicitly; a precise lens mount cannot correct a locally buckled frame.

Alignment also has lateral and in-plane rotational components. These may shift a frame or change what lies within its edges without causing axial defocus. They need their own registration limits. A requirement of one-pixel alignment might be justified for a particular calibrated imaging task, but ordinary photography need not hold the entire sensor mechanically to one pixel of the housing datum if framing can be calibrated or cropped. State the intended function before assigning the tolerance.

Budget shutter error in exposure units

Mechanical precision governs exposure as well as focus. The focal-plane shutter pictured here makes the relevant parts tangible: curtains and their drive mechanism must deliver the intended opening over the recording area.

For a simplified constant-speed, travelling-slit shutter, let $w$ be the slit width measured along its travel and $v$ its speed. Each point on the recording surface is exposed for

Small changes give

Suppose the curtain travels at $2.4m/s$ and the desired local exposure is $1\mathrm{ms}$. The nominal slit width is $2.4\mathrm{mm}$. A requirement of $\pm 0.05$ exposure value (EV) corresponds to a time ratio between $2^{-0.05}$ and $2^{0.05}$, roughly a $3.4\%$ one-sided allowance near nominal. One possible illustrative allocation is slit-width variation bounded by $2\%$, or $0.048\mathrm{mm}$, and speed variation bounded by $1\%$, or $0.024m/s$. Combining their adverse signs gives a little over $3\%$ local time error, inside that illustrative limit.

That calculation describes exposure at a point, not uniformity across the image. Curtains that accelerate differently, fail to remain parallel, or change separation during travel can produce a different $T$ at the top and bottom of the frame. The specification must test both the average exposure and spatial variation. At longer settings, when the frame is fully uncovered for part of the exposure, a single constant-width slit model is no longer the appropriate complete description.

A leaf shutter calls for another model. Its opening area changes during opening and closing, so the effective exposure depends on the time integral of the pupil’s transmitted fraction, not just the interval between two mechanical events. Tolerances in blade motion can consequently affect effective time and throughput differently at different settings. In either design, mechanism friction, wear, supply conditions, and temperature belong in the operating-condition budget—not only part dimensions measured at assembly.

Choose a prediction method—and a way to pay for it

The sums above are worst-case stacks: each contributor is placed at its allowed extreme in the most damaging direction. They are appropriate when a guaranteed bound is required and the individual bounds are real. They can also be expensive, because the combination of all adverse extremes may be rare.

For a specified image point, if signed errors $e_i$ are independent and have measured standard deviations ${\sigma }_i$, a linear model instead gives

That is a prediction of spread, not a guaranteed maximum. Correlated errors require covariance terms. Errors sharing a mount, assembly fixture, or calibration procedure may be correlated; neither independence nor a normal distribution follows from writing a bilateral tolerance on a drawing. Tilts also give opposite signs at opposite edges, so a distribution calculated for one point is not automatically the distribution of the worst point anywhere in the frame.

Your experience with numerical models offers a direct extension when the linear approximation is inadequate. Sample plausible manufactured dimensions, assembly adjustments, temperature changes, and shutter dynamics; calculate focus and exposure metrics for each simulated camera; then inspect yield and the failure modes. Such a Monte Carlo analysis is only as credible as its inputs. Measured process distributions, correlations, drift, and realistic bounds matter more than the number of samples.

Finally, allocate tolerances by sensitivity and manufacturing cost, not equally among parts. In a local model $e={\sum }_i{s}_i\delta p_i$, the magnitude of $s_i$ tells you how strongly a deviation in parameter $p_i$ affects the chosen image metric. Tighten a sensitive dimension when the process can do so economically; otherwise consider changing the datum scheme, reducing the sensitivity through optical or mechanical redesign, or introducing an adjustment. A tolerance tighter than a process can reliably produce may cost more in inspection, yield loss, and rework than an additional adjustment step. Conversely, adjustment adds its own equipment, labour, and long-term stability costs.

The conference-paper example embodies this trade: a lens can have excellent nominal MTF yet be impractical to manufacture until its sensitivity and assembly strategy are reconsidered. A successful tolerance allocation is a specification for a reproducible camera, not merely for an ideal specimen.

Conclusion

Start with permitted photographic errors, convert them to mechanical or timing limits, and trace each limit through the relevant datums and mechanisms. Axial offset, tilt, and surface flatness consume the focus budget in different ways. Shutter width and speed affect local exposure, while motion across the frame can introduce nonuniformity. Worst-case stacks establish bounds; statistical models estimate yield only when their distributional assumptions are justified. Adjustment can buy manufacturing freedom, but brings requirements of its own.

This completes the module on exposure, timing, and focus control. The next module moves to electronic image capture, beginning with how absorbed photons become charge in an image-sensor pixel.