CS180 Proj1 · Colorizing the Prokudin-Gorskii Collection

Overview

Each input plate stacks B-G-R vertically. I split channels, then translate-align G/R to B by maximizing ZNCC on Sobel gradient magnitude with interior-only comparison (ratio-based border crop). For large .tif plates, a coarse-to-fine image pyramid yields accurate and fast alignment.

Original BGR plate — **Input** · BGR plate (top→bottom: B, G, R)

Colorized output — **Output** · Aligned & composed RGB

Methodology

Single-scale alignment (baseline, small JPGs)

Split: split_bgr_plate() slices the vertical plate into B, G, R.
Feature: optionally convert to Sobel gradient magnitude (feature_grad() / grad_mag()).
Interior-only metric: crop_interior() supports pixel or ratio borders.
Search: exhaustive translations within [-R, R] (default ±15 px).
Score: ZNCC (zncc()) on the cropped interior; pick the max.
Compose: stack aligned [R, G, B] to RGB and save.

Implementation notes: crop_interior() accepts int (pixels) or float (ratio); ZNCC is “higher is better”.

Coarse-to-fine pyramid (large TIFs)

Downsample: build pyramids from coarse → fine with factor scale = 0.5.
Coarsest init: run local search at coarsest level using Sobel + ZNCC.
Propagate: rescale offsets by ×2 at each finer level.
Refine: refine with a small-radius local search.
Ratio crop: use ratio borders (e.g., 0.08).

Key routines: build_pyramid(), align_local_search(), align_pyramid().

Results (with Offsets)

Colorized outputs with corresponding offsets (G,R relative to B).

cathedral.jpg
G:(+2,+5), R:(+3,+12)
single · NCC

monastery.jpg
G:(+2,-3), R:(+2,+3)
single · NCC

tobolsk.jpg
G:(+2,+3), R:(+3,+6)
single · NCC

church.tif
G:(+4,+25), R:(-4,+58)
pyramid · edges+ZNCC

emir.tif
G:(+23,+49), R:(+40,+107)
pyramid · edges+ZNCC

harvesters.tif
G:(+17,+60), R:(+14,+124)
pyramid · edges+ZNCC

icon.tif
G:(+17,+42), R:(+23,+90)
pyramid · edges+ZNCC

italil.tif
G:(+22,+38), R:(+36,+77)
pyramid · edges+ZNCC

lastochikino.tif
G:(-2,-3), R:(-8,+76)
pyramid · edges+ZNCC

lugano.tif
G:(-17,+41), R:(-29,+92)
pyramid · edges+ZNCC

melons.tif
G:(+10,+80), R:(+13,+177)
pyramid · edges+ZNCC

self_portrait.tif
G:(+29,+78), R:(+37,+176)
pyramid · edges+ZNCC

siren.tif
G:(-7,+49), R:(-24,+96)
pyramid · edges+ZNCC

three_generations.tif
G:(+12,+54), R:(+9,+111)
pyramid · edges+ZNCC

LoC Extras

Additional images from the Library of Congress collection. Only colorized outputs are shown. Offsets are (G,R) relative to B, format (dx, dy), right/down positive.

test1.tif
G:(+34,+56), R:(+60,+125)
pyramid · edges+ZNCC

test2.tif
G:(+4,+63), R:(-1,+132)
pyramid · edges+ZNCC

test3.tif
G:(-7,+15), R:(-17,+83)
pyramid · edges+ZNCC

test4.tif
G:(+15,+41), R:(+25,+91)
pyramid · edges+ZNCC

Tip: outputs are saved to ./out/ as testX_color.jpg. Offsets filled from console logs. Input plates are intentionally omitted.

Discussion

Failures, lessons, runtime

Challenging cases (e.g., strong cross-channel intensity change like emir); edges helped.
Runtime: < 1 min per image with pyramid + vectorization.
Border cropping and interior-only metrics improved stability.