Machine vision inspection of drill edge chipping

Chipped edges and scratches on drills, end mills and taps sit at the awkward end of automatic inspection: the surfaces are mirrored, the marks are tens of micrometers across, and good parts already carry grinding texture. None of that is a blocker today. This guide sizes a defect station for a standard cutting tool program: the pixel budget for a 50 µm chip, a lens and working distance that deliver it, the lighting geometry that keeps flute glare from hiding defects, and where dimensional gauging belongs. Standard parts make teaching practical, which is half the battle.

Can machine vision find chipping on a drill edge?

Yes. A camera resolving about 10 µm per pixel over a 10 mm field of view puts a 30 µm chip across three pixels and a 50 µm chip across five, and sensors that deliver this are ordinary catalog parts today.

The arithmetic is plain: 1024 pixels across 10 mm is 9.8 µm per pixel, the origin of the familiar rule of thumb that a megapixel camera over a 10 mm field sees 10 µm detail. A modern 5 MP sensor over the same field runs at 4 µm per pixel, so the same 30 µm chip spans seven pixels with margin for threshold tuning. Detection stopped being the hard part; teaching and lighting are where these stations succeed or fail.

The inspection program can also live on the device. Smart cameras and vision sensors now run defect tools, PLC communication and pass/fail I/O without a separate PC, which matters when the station shares a bench with the grinder. For the measurement side of tool making, subpixel edge measurement of cutting tools covers the micrometer-level methods.

Sensing routes for the defect station range from on-device judgment to open lens choice:

The LS-SMC-0017 takes any C-mount lens in front of an 8.8 MP, 1", global-shutter sensor. The MV-SC3050XC-16M-ATN-H is a 5 MP color vision sensor with a global shutter. The LS-SMC-0063 carries its own 16 mm lens and polarized white light, with a 1.6 MP sensor and 60 fps capture.

How many pixels must the smallest chip span?

Three pixels minimum across the smallest chip you must catch, five if software has to classify or measure it, so the budget for a 50 µm chip is 50 ÷ 3 ≈ 17 µm per pixel.

Below three pixels a chip merges into edge noise and texture. Subpixel methods locate an edge you can already see; they do not find a defect the sensor never sampled. The same arithmetic drives camera resolution for defect detection, and it is the first number to fix in a defect project.

Halving the object pixel halves the field of view, which multiplies the views you need per tool. Fix the reject threshold from your own tool standard before any hardware: this guide carries 50 µm through every example, and the table shows what 30 µm costs.

Run your own defect size and sensor width through the calculator:

What lens and working distance hit that pixel size?

A 35 mm lens on a 2/3" sensor at 100 mm working distance delivers exactly that: 0.55x magnification, a 16 x 12 mm field of view, and 10 µm per pixel from a 5.5 µm sensor pitch.

Worked example, defect station:

1. Sensor: a common 2/3" 2 MP device, 1600 x 1200 pixels at 5.5 µm pitch, so the active area is 1600 × 5.5 µm = 8.8 mm wide and 6.6 mm tall.
2. Budget: 10 µm per pixel, so m = 5.5 ÷ 10 = 0.55.
3. Working distance 100 mm, so f = 100 × 0.55 ÷ 1.55 = 35.5 mm — fit a 35 mm lens.
4. Field of view: 8.8 ÷ 0.55 = 16 mm wide, 6.6 ÷ 0.55 = 12 mm tall.
5. Check: 1600 pixels × 10 µm = 16 mm, and a 50 µm chip spans 50 ÷ 10 = 5 pixels.

The LS-FFL-0679 is this class of glass: 35 mm focal length, 2/3" format, 6 MP grade, C-mount.

Try your own working distance in the lens focal length calculator, or filter machine vision lenses by focal length in the lens finder.

Aperture is the quiet limit. The flute helix drops the edge several millimeters along its length, so you want depth of field — but the Airy disc grows as 2.44 × wavelength × f-number. At f/8 and 550 nm that is 2.44 × 0.55 µm × 8 = 10.7 µm, matched to a 10 µm pixel. Work between f/5.6 and f/8; past f/11 you trade away the contrast that reveals chips.

When the fixture rotates the tool between views, a fixed-magnification lens removes focus drift and perspective error in one step. The LS-TCL-0574 is a fixed-magnification C-mount lens: 0.4x at a 227 mm working distance.

At 0.4x, an 8.8 mm sensor width covers 8.8 ÷ 0.4 = 22 mm of flute and the object pixel grows to 5.5 ÷ 0.4 = 13.8 µm. A 50 µm chip still spans 3.6 pixels: enough to detect, tight to classify.

How do you light a mirrored flute without glare?

Leave coaxial light off the curved flutes and use a diffuse dome or low-angle dark-field bars there; a chip scatters the light the polish reflects away, so defect contrast appears exactly where the background goes uniform or dark.

A ground flute is a curved mirror. Aim a ring or coaxial beam at it and you get one traveling band of glare that follows part rotation across every image. A dome wraps the surface in light from all angles so no single reflection dominates; dark-field bars graze the surface at roughly 10-20 degrees so the polish images near-black and every broken edge lights up as scatter. The same reflection law decides bright field versus dark field lighting in every metal job, and lighting for metal surface defect inspection maps the wider menu.

Size the diffuse source at about twice your field of view, so the flute edges receive light, not shadow. On a mirror finish, cross-polarization is the second lever: polarizer on the light, analyzer on the lens rotated 90 degrees, and the specular streak vanishes while chips keep a little depolarized scatter. It costs roughly two to three stops, so budget strobe energy. Lighting shiny metal pins without glare works the identical problem on pins.

The flat clearance faces and any engraving are where coaxial light earns its slot: the beam travels along the optical axis, the flat polish returns it, and chips or scratches read as dark interruptions. Coaxial lights sized to the face do that job:

How do you separate real chips from grinding marks?

Change the lighting direction between two exposures and keep only what survives both: grinding lay flips its contrast when the light rotates 90 degrees, while a chip scatters in every direction and stays visible from all sides.

Carbide flutes carry intentional texture — wheel marks, honing striations, coating roughness — and it all runs in one direction. Fire a bar from the left, then from the top, and subtract: oriented texture changes sign between the two images, a chip does not. In software, a directional-derivative filter over the edge region is the same trick standing still. Chaining several lights through one camera is standard practice — sequential machine vision lighting for multi-view inspection lays out the triggering.

Teaching closes the gap. Because you inspect standard products, capture twenty to thirty accepted parts across lots, and set thresholds inside that texture envelope rather than against a perfect master. For gross damage, broken tool detection lighting finds a missing tip at far lower magnification.

Is defect inspection a separate unit from dimensional measurement?

Yes — plan two stations. A 0.1 mm dimensional tolerance needs only 25-35 µm per pixel with the whole tool in frame, while chipping inspection wants 10 µm per pixel on the edges, and the two jobs want opposite lighting.

The measurement side is silhouette territory. Backlight the tool and measure diameter, length and point geometry from the shadow: at 30 µm per pixel a subpixel edge repeats to about a tenth of a pixel, 3 µm, a 30-fold margin inside a 100 µm tolerance. The backlight and histogram edge inspection guide walks through that set-up, and machine vision dimensional measurement for small parts covers the metrology decisions.

The LS-FFL-0567 suits the wide station: 16 mm focal length, 1.1" format, 20 MP, C-mount, distortion < 0.3 %.

Even at 35 µm per pixel the tolerance spans three pixels, and a 20 MP frame over a 70 mm tool runs finer than that, so the wide station holds margin without trying.

Combine the two only when the reject size is large: at 30 µm per pixel a 100 µm chip still spans three pixels, so one wide station can carry both jobs. Below 100 µm reject size, split the stations and let each optic do one thing well.

How fast can one station clear a 10,000-piece lot?

Plan around two seconds per drill. Four strobed exposures of about 1 ms freeze the part, capture and judgment together stay under 200 ms, and indexing plus handling owns the rest — roughly 1,800 pieces per hour through one station.

The arithmetic: 10,000 pieces × 2 s = 20,000 s ≈ 5.6 hours, and a 300-piece pilot lot clears in ten minutes. Strobe the lights from a controller in sync with the camera trigger: the pulse overpowers shop lighting without heating the part, and short exposures erase any residual index motion. If takt time demands more, add a second head — the mechanical index, not the camera, is the bottleneck.

What should you check before ordering parts?

Work through this list before anything is ordered: the reject threshold sets the pixel budget, the pixel budget sets sensor and lens, and the lighting geometry decides whether the texture ever lets you use that resolution.

• Reject threshold, in µm. Every number below divides from this one.
• Pixel budget. Threshold ÷ 3 gives the object pixel; sensor pixels × object pixel gives the affordable field of view.
• Lens coverage. The image circle must span the sensor diagonal, and the lens resolution grade must meet the sensor's megapixel class.
• Aperture window. Stay between f/5.6 and f/8 so the Airy disc stays near the pixel size while flute depth remains in focus.
• Lighting per surface. Dome or dark field on curved flutes, coaxial on flat faces, crossed polarizers on mirror finishes.
• Two lighting directions. A second exposure from 90 degrees is the cheapest defense against grinding lay.
• Teach set. Twenty to thirty accepted parts from different lots define the allowed texture envelope.
• Indexing fixture. Four repeatable views need repeatable rotation, or thresholds never hold.

Start the defect station in the system configurator, compare lights at the lighting step, or send the pixel budget you calculated to the applications team for a parts list.