The practical consequence is that engineers must think in terms of aggregate throughput rather than per-camera specifications alone. A line with eight cameras does not necessarily need eight equally powerful processing nodes; it needs enough total capacity, correctly routed, to meet the combined cycle time requirement across all stations during peak load.
How Do Calibration and Optics Affect the Final Measurement? No sub-pixel algorithm can compensate for uncorrected lens distortion or an uncalibrated pixel-to-millimeter conversion. Before any sub-pixel routine adds value, the system needs a calibration target, typically a dot grid or checkerboard pattern, imaged at the exact working distance and lighting conditions used in production, so the software can map pixel coordinates to real-world units and correct for barrel or pincushion distortion introduced by the lens. Telecentric lenses are frequently specified in high-precision metrology applications specifically because they minimize perspective error across depth of field, which otherwise would cause apparent part size to shift as the object’s position varies slightly within the inspection station.
Sub-pixel precision is not a property of the camera or the software alone; it emerges only when optics, illumination, sensor characteristics, and algorithm selection are matched to the specific measurement task at hand. Consider a worked example: an inspection station needs to measure the diameter of a stamped washer with a nominal specification of 10.00 mm ± 0.02 mm. The camera delivers a pixel size of 15 microns after calibration. A whole-pixel measurement system can only resolve to ±15 microns, which already consumes most of the allowable tolerance band before accounting for any other error source. Applying a sub-pixel edge algorithm capable of resolving to 0.1 pixel drops the theoretical resolution to roughly 1.5 microns, leaving comfortable margin for repeatability variation, thermal drift, and fixture tolerance. This is the arithmetic that justifies the additional processing overhead in tolerance-critical applications. machine vision systems
No. Standard CMOS and CCD sensors used in visible-light cameras are built on silicon photodiodes that have very low quantum efficiency above roughly 1000 nm, so adding an SWIR bandpass filter to a silicon-based sensor mainly blocks light without producing a usable image. A dedicated InGaAs sensor is required to achieve meaningful sensitivity across the 900-1700 nm range used in wafer transmission imaging.
The angular nature of entocentric imaging also means that lighting and shadow behavior change across the frame, further complicating edge detection algorithms used in automated inspection software. Engineers often try to compensate with software correction factors or calibration lookup tables, but these are approximations that degrade whenever the part’s height profile deviates from the calibration sample. This is precisely the gap that telecentric optical design was engineered to close. machine vision systems
Why Does Silicon Become Transparent Under SWIR Illumination? The physics behind this behavior relates directly to silicon’s bandgap energy, which sits at approximately 1.12 electron volts. Photons with energy below this threshold – corresponding to wavelengths longer than roughly 1100 nm – lack sufficient energy to excite electrons across the bandgap, so they pass through the material largely unabsorbed rather than being reflected or scattered at the surface. This is fundamentally different from how silicon interacts with visible light, where photons are absorbed almost immediately at or near the surface, which is why a silicon wafer looks like an opaque, mirror-like disc to the naked eye.
Under good contrast and lighting, sub-pixel algorithms typically resolve edges to within 0.05 to 0.1 of a pixel, compared to a full pixel of uncertainty in standard thresholding, effectively improving measurement resolution by a factor of ten or more in favorable conditions.
Why Do Some Vision Projects Pay Back in Months While Others Never Break Even? The difference almost always traces back to scope definition before hardware selection. Teams that succeed start by quantifying the cost of the problem they are solving – missed defects, rework labor, warranty claims, or manual inspection headcount – before choosing a camera or lens. A project aimed at replacing three inspectors working three shifts, for example, has a labor-cost baseline that can be measured in weeks, while a project justified only by “improved quality” has no baseline at all and struggles to demonstrate value to finance stakeholders.
What Illumination Setup Works Best for Wafer Transmission Imaging? Because the technique relies on transmitting light through the wafer rather than reflecting it off the surface, illumination geometry differs fundamentally from standard machine vision setups. Backlighting with a uniform SWIR LED or laser diode array positioned directly opposite the camera is the most common configuration, producing a transmission image where defects appear as shadows, distortions, or scattering artifacts against an otherwise even bright field. Illumination uniformity across the full wafer diameter is critical, since even modest intensity gradients can be misread as defects or, worse, mask real ones near the edges of the field. machine vision systems
