Getting Nikon Coolscan 5000 RGB+IR scans into NegPy, with aligned frame previews and dust removal

[rohanpandula]

I've been working on direct NegPy capture from a Nikon Super Coolscan 5000 ED.
SANE lists this model as untested in coolscan3, and the stock backend did not
expose a usable infrared scan on my LS-5000 ED with firmware 1.03.

My feeder started as an SA-21 and has the solderless SA-30 conversion. The
scanner reports 40 addressable positions, which is transport capacity. The
roll was sold as a 24-exposure roll, so I initially registered positions 1
through 24. I later matched those previews against an earlier sweep and found
image content at positions 25 through 27. Nominal roll length was not enough
to count the content-bearing positions, so automatic preview census is still
missing.

My local NegPy/SANE capture path works at 4000 dpi and 16-bit. NegPy first
captures RGB at 4x with infrared off, then captures RGBI at 1x through the same
SANE device open without moving the frame. It registers the 1x capture's IR
plane to the 4x RGB and writes an RGB TIFF plus an IR sidecar as one paired
commit.
Synchronous failures roll both files back; like any two-file scheme, this is
not power-loss atomic.

I also made a reference export with Nikon Scan 4.0.3 W in a Windows VM. That
file is only a Digital ICE reference. Nikon Scan is not part of the open
capture path.

The six-pane crop shows the open RGB capture, IR defect signal, NegPy repair,
Photoshop Content-Aware Fill from the same starting mask, Nikon Digital ICE
from a separate scan, and an amplified NegPy versus Photoshop difference. The
montage uses an earlier 1x RGBI capture. I captured the final 4x RGB plus 1x
RGBI validation master later from the same frame.

What works on this scanner

• 4000 dpi, 16-bit RGB at 4x followed by RGBI at 1x in one device reservation
• autofocus and auto-exposure on the first capture, held for the second
• RGB-only hardware multisampling at 2x and 16x at 1000 dpi, and 4x at 4000 dpi
• physical RGBI capture through NegPy over saned at 1000 and 4000 dpi
• paired RGB and IR TIFF output with DPI metadata
• preview-based frame positioning that retains the film borders without a
  broad black strip or stopped-transport tail
• NegPy C-41 inversion and IR-guided repair from the new 4x plus 1x master

I checked 24 aligned previews from positions 1 through 24. This was not a
complete exposure census: positions 25 through 27 still need current
registered previews. I then ran the full capture sequence on frames 03, 07,
14, and 16. Each frame used one SANE open and kept the same transport geometry
for both internal captures.

Frame 03's RGB capture was 3946 x 5004 before registration. The common-overlap
crop produced a 3946 x 5003 RGB and IR pair. The later registered pairs were
3945 x 5022, 3945 x 5039, and 3945 x 5049 after the same overlap step. These
dimensions are width x height. The other 20 positions in the registered set
from 1 through 24 have previews only. Positions 25 through 27 have historical
1x scans but no current registration, so this is not a full-resolution
validation or a complete preview census of the entire roll.

The first full scans of frames 07, 14, and 16 exposed another failure mode.
They reused frame 03's geometry and looked close at a glance, but the detector
found stopped-transport suffixes of 78, 254, and 304 rows. Running the same
frames with their own preview registration reduced all three suffixes to zero
rows.

The public proof image uses only frames 07 and 16. The attached JSON records
all six fixed-geometry and corrected captures for frames 07, 14, and 16. It
contains source hashes, dimensions, geometry, RGB and IR alignment values,
the detector contract, suffix starts and lengths, and acceptance verdicts.
I am keeping the raw reports, source TIFFs, and 24-position contact sheet
private and will not upload them.

alignment-evidence.json

On this feeder, the tested frames needed their own preview-derived positions.
A fixed leader offset can drift as the film advances, and a window that extends
beyond usable travel can finish with repeated rows instead of an obvious read
error.

Upstream plan

The local NegPy work is too large for one useful review. I am splitting it into
focused PRs:

1. 1x inline RGBI detection, fourth-channel recovery, and RGB/IR splitting
2. autofocus, sample count, checked frame selection, progress, paired TIFF
   writing, and DPI metadata
3. RGB Nx plus registered RGBI 1x in one scanner reservation
4. the preview-registration algorithm and a compact fixture
5. roll planning, orchestration, and registered transport geometry

The first NegPy PR is open as a draft. The SANE parameter-block fix has merged
into master, and the LS-5000 RGBI and multisample work remains a draft MR:
NegPy RGBI draft PR, SANE
parameter-block fix (merged),
and SANE LS-5000 RGBI and multisample draft MR.

The LS-5000 MR is rebased on the master commit that includes !936 and
contains only its own RGBI and multisample changes. It remains a draft because
its first maintainer question is whether four interleaved samples should
continue to be reported under SANE_FRAME_RGB, following the existing
pieusb precedent, or whether IR should be returned as a separate grayscale
frame. I opened the draft to collect maintainer feedback on that design
question.

Dust repair comparison

NegPy and Photoshop started with the same earlier 4000 dpi, 16-bit 1x RGBI
capture and the same IR mask. At 1:1, I found the finished crops hard to tell
apart, although Photoshop looked slightly more natural on a few larger
repairs.

The Nikon pane is a separate scan of the same frame with Nikon Scan 4.0.3 W,
MultiSample 4x, and Digital ICE Normal. Feature matching confirmed the scene,
but the capture, crop, color, and repair all differ, so I did not calculate a
Nikon pixel-error score.

A seeded synthetic test used a known clean image with 30 dust spots and two
hair scratches. Both methods substantially improved all 32 defects:

Workflow	MAE inside defects, 8-bit scale	Collateral MAE outside
Unhealed	53.41	0
Photoshop Content-Aware Fill	3.17	0.0019
NegPy IR heal	3.45	0.0000

Photoshop had the lower error on this synthetic image. I have not run a blind
perceptual test.

These are workflow timings, not a pure algorithm benchmark. On an Apple M4
Mac mini, NegPy's in-process repair took 0.9 seconds for a 5730 x 3600, 16-bit
frame. The Photoshop script took 10.7 seconds with Photoshop open and 14.7
seconds from a cold start, including file I/O and app automation.

One warning about multisampling and IR

A single RGBI scan with multisampling is unsafe on this scanner. When I tested
that combination earlier, the scanner became unresponsive until I
power-cycled it. The SANE draft rejects it in either option order and again
before scan start.

The working path uses two captures in one SANE open: RGB Nx with IR off, then
RGBI 1x without moving the frame. The second capture's RGB channels register
its IR map to the multisampled RGB. The four 4000 dpi runs passed the capture,
alignment, TIFF, smear, and scanner-response checks.

Offline clipping telemetry found small, fragmented red-channel saturation in
frames 03 and 07. Green, blue, and IR did not clip at the high end, and no
channel clipped at the low end. I am treating the red values as a review
warning until I have a same-frame lower-exposure comparison, not as a silent
pass or a reason to discard the masters.

Nikon's TIFF reports MultiSample 4x and Digital ICE, but it does not show
whether Nikon Scan sampled IR more than once or used multiple internal passes.
Here, practical parity means aligned 4000 dpi, 16-bit capture, autofocus,
auto-exposure, multisampled RGB, and a registered IR map. NegPy can then use
its own inversion, color controls, and deterministic IR-guided repair.

Hardware IR works with compatible color film, but not reliably with
IR-opaque media such as traditional silver black-and-white film or
Kodachrome.

I would be interested in results from an LS-50, another LS-5000 firmware
revision, or anyone familiar with SANE's preferred fourth-plane

[rohanpandula]

Measurements behind the opening post:

Practical-parity capture

The frame 03 baseline used 4000 dpi, 16-bit RGB 4x with IR off followed by
RGBI 1x with IR on. Both captures shared one SANE open, with no frame,
subframe, or scan-window rewrite between them. The final master was 3946 x
5003 pixels (width x height). Registration measured dx 0.000 px, dy -0.396 px,
and ECC 0.999981. The TIFF pair passed its reload checks, and the scanner
answered the follow-up probe.

Three later positions used the same sequence with preview-derived geometry:

Frame	subframe / br_y	Final RGB and IR, width x height	Smear suffix	ECC	Responsive
07	5.59 mm / 5023	3945 x 5022	0 rows	0.999976	Yes
14	4.38 mm / 5039	3945 x 5039	0 rows	0.999990	Yes
16	4.00 mm / 5049	3945 x 5049	0 rows	0.999994	Yes

The fixed-geometry negative controls used 6.35 mm and br_y=5003 on all
three frames. The current detector rejects them with suffixes beginning at
rows 4925, 4749, and 4699, for lengths of 78, 254, and 304 rows. The corrected
captures all return a clean zero-row result. Exact TIFF and private-report
hashes are in the attached alignment-evidence.json.

The detector evaluates the central 90 percent of sampled columns. A candidate
tail needs at least 64 matching row transitions at 4000 dpi, adjacent-row RMS
below 150, correlation above 0.9999, enough texture, and an abrupt boundary
from the preceding rows. Partial or low-texture signatures are indeterminate
and fail closed.

NegPy and Photoshop

For the dust montage, both repairs used the same crop and starting IR < 0.50
mask from one 4000 dpi, 16-bit 1x RGBI capture:

Measurement	Value
Exposure geometry	4910 x 3688
Initial masked pixels	40,805
Initial mask coverage	0.2253%
8-connected components	1,909

NegPy used its local IR-guided repair. In Photoshop, I put the mask in a
DustIR channel, expanded it two pixels, feathered it one pixel, and ran local
Content-Aware Fill. I did not use Generative Fill or Firefly.

The amplified difference also includes small TIFF and color-profile
round-trip changes. The largest differences still cluster around repairs.

Nikon registration

The Nikon Scan TIFF is a separate capture. SIFT matching and full affine
RANSAC found 99 inliers from 167 candidates. At full resolution, the median
residual was 0.71 pixel and the 95th percentile was 1.82 pixels. This confirms
the scene match, not pixel-level comparability.

The Nikon TIFF reports:

Setting	Value
Software	Nikon Scan 4.0.3 W
Image	5959 x 3946, 16-bit RGB, 4000 dpi
Film type	NEGATIVE (COLOR)
Multi-sample	4x
Digital ICE	Normal
Profile	Nikon Adobe RGB 4.0.0.3000

Links

• NegPy focused RGBI draft PR: Handle inline RGBI frames from SANE coolscan3 #471
• SANE parameter-block fix (merged): https://gitlab.com/sane-project/backends/-/merge_requests/936
• SANE LS-5000 RGBI and multisample draft MR: https://gitlab.com/sane-project/backends/-/merge_requests/937
• Existing SANE LS-5000 issue: https://gitlab.com/sane-project/backends/-/issues/8

The backend currently uses a 38.106 mm frame-pitch constant. I have a local
38.0 mm experiment, but it is not in the submission list. It needs a same-roll
A/B transport test before it can support an alignment claim.

I am also leaving out the experimental per-pass READ patch. Later tests did
not support its original transfer-size explanation, and RGBI still stopped

[marcinz606]

nice,I see on screenshots it plays well with inpainting improvements I pushed today

[marcinz606]

I'm just bidding on that...

[rohanpandula]

you are going to LOVE it! Ive spent so much money and time on camera scanning and this thing blows all of that out of the water. Make sure you look into the solder-less SA21 to SA30 mod. Took me like 5 min to do and now it can do full rolls automatically!

[marcinz606]

I won it! Hopefully I'll receive it before the weekend

[thetalkingdrum]

Congratulations! Great scanner.

Getting full roll on the V is a bit more involved than on a 5000 or 4000. Besides the SA21 mod you need to flash the firmware with kosmas mod (https://github.com/kosma/coolscan-mods), and cut a whole in the back.

You might want to test Nikon scan as well just to experience just how good ICE is.

[marcinz606]

I probably won't resort to cutting hole in the back :D at least not initially. I always cut my rolls to 6 exposure strips anyway and keep them in 6x7 archival sheets

[thetalkingdrum]

Yeah, I haven't done it either. Not for the faint-hearted! :)