Estimate a log-linear formula for the extracted transfer function LUT.

This commit is contained in:
Nathan Vegdahl 2022-04-20 12:34:49 -07:00
parent 22db31fe5e
commit ab91aee328
5 changed files with 211 additions and 2 deletions

40
Cargo.lock generated
View File

@ -211,6 +211,7 @@ dependencies = [
"clap",
"colorbox",
"exr",
"simplers_optimization",
]
[[package]]
@ -228,6 +229,15 @@ dependencies = [
"getrandom",
]
[[package]]
name = "num-traits"
version = "0.2.14"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "9a64b1ec5cda2586e284722486d802acf1f7dbdc623e2bfc57e65ca1cd099290"
dependencies = [
"autocfg",
]
[[package]]
name = "num_cpus"
version = "1.13.1"
@ -238,6 +248,15 @@ dependencies = [
"libc",
]
[[package]]
name = "ordered-float"
version = "2.10.0"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "7940cf2ca942593318d07fcf2596cdca60a85c9e7fab408a5e21a4f9dcd40d87"
dependencies = [
"num-traits",
]
[[package]]
name = "os_str_bytes"
version = "6.0.0"
@ -267,6 +286,16 @@ dependencies = [
"syn",
]
[[package]]
name = "priority-queue"
version = "1.2.1"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "00ba480ac08d3cfc40dea10fd466fd2c14dee3ea6fc7873bc4079eda2727caf0"
dependencies = [
"autocfg",
"indexmap",
]
[[package]]
name = "proc-macro2"
version = "1.0.37"
@ -291,6 +320,17 @@ version = "1.1.0"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "d29ab0c6d3fc0ee92fe66e2d99f700eab17a8d57d1c1d3b748380fb20baa78cd"
[[package]]
name = "simplers_optimization"
version = "0.4.3"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "2cd97912bb2a16575a2c632c2a2f2bac8a527827ceaddd73e0ccc12d86adec43"
dependencies = [
"num-traits",
"ordered-float",
"priority-queue",
]
[[package]]
name = "smallvec"
version = "1.8.0"

View File

@ -7,3 +7,4 @@ edition = "2021"
exr = "1.4.1"
clap = { version = "3.1.8", default-features = false, features=["std"] }
colorbox = { git = "https://github.com/cessen/colorbox", branch = "master" }
simplers_optimization = "0.4.3"

107
src/linear_log.rs Normal file
View File

@ -0,0 +1,107 @@
/// A composite linear-log function.
///
/// `slope` is the slope of the linear segment.
/// `base` is the log base.
/// The offsets shift the linear and log parts of the curve along
/// the linear color axis.
pub fn linear_to_log(x: f64, line_offset: f64, slope: f64, log_offset: f64, base: f64) -> f64 {
// Transition point between log and linear.
let transition = 1.0 / (slope * base.ln());
let k = transition + log_offset;
let l = (transition - line_offset + log_offset) * slope - transition.log(base);
if x <= k {
(x - line_offset) * slope
} else {
(x - log_offset).log(base) + l
}
}
/// A composite linear-log function.
///
/// `slope` is the slope of the linear segment.
/// `base` is the log base.
/// The offsets shift the linear and log parts of the curve along
/// the linear color axis.
pub fn log_to_linear(x: f64, line_offset: f64, slope: f64, log_offset: f64, base: f64) -> f64 {
// Transition point between log and linear.
let transition = 1.0 / (slope * base.ln());
let k = (transition - line_offset + log_offset) * slope;
let l = (transition - line_offset + log_offset) * slope - transition.log(base);
if x <= k {
(x / slope) + line_offset
} else {
base.powf(x - l) + log_offset
}
}
//-------------------------------------------------------------
/// Generates Rust code for a linear-to-log transfer function with the
/// given parameters.
pub fn generate_linear_to_log(line_offset: f64, slope: f64, log_offset: f64, base: f64) -> String {
let transition = 1.0 / (slope * base.ln());
let k = transition + log_offset;
let l = (transition - line_offset + log_offset) * slope - transition.log(base);
format!(
r#"
pub fn linear_to_log(x: f32) -> f32 {{
const A: f32 = {};
const B: f32 = {};
const C: f32 = {};
const D: f32 = {};
const E: f32 = {};
const F: f32 = {};
if x <= A {{
(x - B) * C
}} else {{
(x - D).log2() * (1.0 / E) + F
}}
}}
"#,
k,
line_offset,
slope,
log_offset,
base.log2(),
l,
)
}
/// Generates Rust code for a log-to-linear transfer function with the
/// given parameters.
pub fn generate_log_to_linear(line_offset: f64, slope: f64, log_offset: f64, base: f64) -> String {
let transition = 1.0 / (slope * base.ln());
let k = (transition - line_offset + log_offset) * slope;
let l = (transition - line_offset + log_offset) * slope - transition.log(base);
format!(
r#"
pub fn log_to_linear(x: f32) -> f32 {{
const A: f32 = {};
const B: f32 = {};
const C: f32 = {};
const D: f32 = {};
const E: f32 = {};
const F: f32 = {};
if x <= A {{
(x * (1.0 / C)) + B
}} else {{
((x - F) * E).exp2() + D
}}
}}
"#,
k,
line_offset,
slope,
log_offset,
base.log2(),
l,
)
}

View File

@ -1,9 +1,11 @@
mod linear_log;
mod optimize_log;
mod test_image;
use std::{fs::File, io::BufWriter, path::Path};
use clap::{Arg, Command};
use colorbox::transfer_functions::srgb;
// use colorbox::transfer_functions::srgb;
use test_image::{GRADIENT_LEN, RES_X, RES_Y};
@ -46,7 +48,7 @@ fn main() {
} else {
let input_path = args.value_of("input").unwrap();
let base_image = test_image::build();
// let base_image = test_image::build();
let mut input_image = {
let (image, res_x, res_y) = read_rgb_exr(input_path).unwrap();
assert_eq!(
@ -87,6 +89,8 @@ fn main() {
gray_b.push(rgb[2]);
}
optimize_log::find_parameters(&gray_r);
// Write the LUT.
colorbox::formats::cube::write_1d(
BufWriter::new(File::create("test.cube").unwrap()),

57
src/optimize_log.rs Normal file
View File

@ -0,0 +1,57 @@
use crate::linear_log::log_to_linear as log_to_lin;
pub fn find_parameters(lut: &[f32]) {
let lin_norm = 1.0 / (lut.len() - 1) as f64;
// Compute the stuff that we can without estimation.
let offset = lut[0] as f64;
let slope = lin_norm / (lut[1] as f64 - lut[0] as f64);
// Select a range of points from the lookup table to fit to.
let idxs: Vec<_> = (0..lut.len()).step_by(lut.len() / 256).collect();
let coords: Vec<(f64, f64)> = idxs
.iter()
.map(|i| (*i as f64 * lin_norm, lut[*i] as f64))
.collect();
// Do the fitting.
let f = |v: &[f64]| {
let mut avg_sqr_err = 0.0f64;
for (x, y) in coords.iter().copied() {
let e = (log_to_lin(x, offset, slope, v[0], v[1]) - y).abs() / y.abs();
avg_sqr_err += e * e;
}
let last_y = lut[lut.len() - 1] as f64;
let e = (log_to_lin(1.0, offset, slope, v[0], v[1]) - last_y).abs() / last_y.abs();
avg_sqr_err += e * e;
avg_sqr_err
};
let input_interval = vec![(-0.2, 0.2), (1.1, 1000.0)];
let (_, params) = simplers_optimization::Optimizer::minimize(&f, &input_interval, 1000000);
// Calculate the error of our model.
let mut max_err = 0.0f64;
let mut avg_err = 0.0f64;
let mut avg_samples = 0usize;
for (i, y) in lut.iter().map(|y| *y as f64).enumerate() {
// We only record error for values that aren't crazy tiny, since
// their relative error isn't representative.
if y.abs() > 0.0001 {
let x = i as f64 * lin_norm;
let e = (log_to_lin(x, offset, slope, params[0], params[1]) - y).abs()
/ y.abs();
max_err = max_err.max(e);
avg_err += e;
avg_samples += 1;
}
}
avg_err /= avg_samples as f64;
println!("Max Err: {:.4}%\nAvg Err: {:.4}%", max_err * 100.0, avg_err * 100.0);
println!(
"{}{}",
crate::linear_log::generate_linear_to_log(offset, slope, params[0], params[1],),
crate::linear_log::generate_log_to_linear(offset, slope, params[0], params[1],),
);
}