The most common camera-sizing mistake starts at the catalog: "we need at least 12 megapixels". The reliable method starts at the opposite end - at the smallest defect your process must catch - and works backwards through four steps. Done properly, it takes ten minutes with a calculator and eliminates both underbuying (missed defects) and overbuying (megapixels nobody reads).
Step 1: Define the critical defect, in millimeters
Not "scratches" - the smallest scratch your customer contract or process spec requires you to catch. Call it D. If your real requirement is 0.15 mm but you specify 0.05 mm "to be safe", you have just quadrupled your camera class for a safety nobody asked for.
Step 2: Apply the three-pixel rule
Reliable machine detection needs the defect to span about 3 pixels (contrast tasks sometimes survive 2; measurement wants 5+). So your ground sampling distance - millimeters per pixel - must be D/3. For a 0.15 mm defect: 0.05 mm/pixel or finer.
Step 3: Divide your field of view
Pixels needed across = FOV / (D/3). Worked examples:
Conveyor part, 100 mm FOV, 0.15 mm defects: 100 / 0.05 = 2000 pixels across. A 5 MP camera (2448 across - the MV-CS050-10GM) covers it with margin; a 2 MP would not.
Full PCB panel, 250 mm FOV, 0.1 mm defects: 250 / 0.033 = 7500 pixels across. Now you are in 12 MP (4024 across, still short) versus 20 MP (5472, close - tighten the FOV to 180 mm and it fits) territory. This is the calculation that justifies the MV-CS200-10GM - and it also explains why panels are inspected in sections at 5 MP when throughput allows.
2D code reading, 20 mm FOV, module size 0.3 mm: 20 / 0.1 = 200 pixels across - any camera made this century resolves it; the real constraint becomes decoding speed and lighting, not resolution.
Step 4: Check the chain behind the sensor
The calculated resolution is only delivered if the chain holds up:
Lens: must resolve the pixel size you chose (approximately 1/(2 x pixel size) in lp/mm) across the full image circle, corners included.
Interface: pixels x frame rate must fit the wire. GigE carries about 118 MB/s; a 20 MP frame at 8-bit is roughly 20 MB, so 5.9 fps is the ceiling - as the MV-CS200-10GM's datasheet says, not because the sensor is slow but because the link is full.
Shutter and light: moving parts at this exposure need global shutter or strobing; the light must fill the chosen exposure to two-thirds saturation.
Common corrections
Aspect ratio mismatch: a wide, short field (400 x 60 mm) wastes a square sensor; a 5 MP 2448-wide camera covers 400 mm at 0.163 mm/px - recheck the defect math at the narrow dimension, or rotate the camera 90 degrees.
Overspec by habit: teams inherit "12 MP because the last project used 12 MP". Run the four steps fresh each time; the savings from one honest downgrade usually funds the lighting upgrade that actually improves detection.
Underspec by compression: JPEG or bandwidth-throttled pipelines can destroy the last pixel of effective resolution - compute the math on delivered data, not acquired data.
One table to rule the shortlist
Write down: FOV, smallest defect, computed pixels-across, required frame rate. Then read our camera pages' resolution and frame rate rows against it - the MV-CS050-10GM (2448 px, 24 fps), MV-CU120-10GM (4024 px, 9.7 fps), MV-CS200-10GM (5472 px, 5.9 fps) form a ladder that covers most discrete-part applications. The right camera is the cheapest rung whose numbers clear your four lines - and now you can prove it to purchasing with arithmetic instead of adjectives.