c1b8c4986e
Profiling the resize operation using Lanczos kernels showed that most of the time was spent in the Sin function, which is called twice per evaluation of the Lanczos kernel. Generating a lookup table and doing a linear interpolation between table entries speeds up the resize by a factor of 4.
258 lines
6.5 KiB
Go
258 lines
6.5 KiB
Go
/*
|
|
Copyright (c) 2012, Jan Schlicht <jan.schlicht@gmail.com>
|
|
|
|
Permission to use, copy, modify, and/or distribute this software for any purpose
|
|
with or without fee is hereby granted, provided that the above copyright notice
|
|
and this permission notice appear in all copies.
|
|
|
|
THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES WITH
|
|
REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND
|
|
FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY SPECIAL, DIRECT,
|
|
INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS
|
|
OF USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR OTHER
|
|
TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR PERFORMANCE OF
|
|
THIS SOFTWARE.
|
|
*/
|
|
|
|
package resize
|
|
|
|
import (
|
|
"image"
|
|
"image/color"
|
|
"math"
|
|
)
|
|
|
|
// restrict an input float32 to the range of uint16 values
|
|
func clampToUint16(x float32) (y uint16) {
|
|
y = uint16(x)
|
|
if x < 0 {
|
|
y = 0
|
|
} else if x > float32(0xfffe) {
|
|
// "else if x > float32(0xffff)" will cause overflows!
|
|
y = 0xffff
|
|
}
|
|
|
|
return
|
|
}
|
|
|
|
// describe a resampling filter
|
|
type filterModel struct {
|
|
// resampling is done by convolution with a (scaled) kernel
|
|
kernel func(float32) float32
|
|
|
|
// instead of blurring an image before downscaling to avoid aliasing,
|
|
// the filter is scaled by a factor which leads to a similar effect
|
|
factor [2]float32
|
|
|
|
// for optimized access to image points
|
|
converter
|
|
|
|
// temporaries used by Interpolate
|
|
tempRow, tempCol []colorArray
|
|
}
|
|
|
|
func (f *filterModel) convolution1d(x float32, p []colorArray, factor float32) colorArray {
|
|
var k float32
|
|
var sum float32 = 0
|
|
c := colorArray{0.0, 0.0, 0.0, 0.0}
|
|
|
|
for j := range p {
|
|
k = f.kernel((x - float32(j)) / factor)
|
|
sum += k
|
|
for i := range c {
|
|
c[i] += p[j][i] * k
|
|
}
|
|
}
|
|
|
|
// normalize values
|
|
for i := range c {
|
|
c[i] = c[i] / sum
|
|
}
|
|
|
|
return c
|
|
}
|
|
|
|
func (f *filterModel) Interpolate(x, y float32) color.RGBA64 {
|
|
xf, yf := int(x)-len(f.tempRow)/2+1, int(y)-len(f.tempCol)/2+1
|
|
x -= float32(xf)
|
|
y -= float32(yf)
|
|
|
|
for i := range f.tempCol {
|
|
for j := range f.tempRow {
|
|
f.tempRow[j] = f.at(xf+j, yf+i)
|
|
}
|
|
|
|
f.tempCol[i] = f.convolution1d(x, f.tempRow, f.factor[0])
|
|
}
|
|
|
|
c := f.convolution1d(y, f.tempCol, f.factor[1])
|
|
return color.RGBA64{
|
|
clampToUint16(c[0]),
|
|
clampToUint16(c[1]),
|
|
clampToUint16(c[2]),
|
|
clampToUint16(c[3]),
|
|
}
|
|
}
|
|
|
|
// createFilter tries to find an optimized converter for the given input image
|
|
// and initializes all filterModel members to their defaults
|
|
func createFilter(img image.Image, factor [2]float32, size int, kernel func(float32) float32) (f Filter) {
|
|
sizeX := size * (int(math.Ceil(float64(factor[0]))))
|
|
sizeY := size * (int(math.Ceil(float64(factor[1]))))
|
|
|
|
switch img.(type) {
|
|
default:
|
|
f = &filterModel{
|
|
kernel, factor,
|
|
&genericConverter{img},
|
|
make([]colorArray, sizeX), make([]colorArray, sizeY),
|
|
}
|
|
case *image.RGBA:
|
|
f = &filterModel{
|
|
kernel, factor,
|
|
&rgbaConverter{img.(*image.RGBA)},
|
|
make([]colorArray, sizeX), make([]colorArray, sizeY),
|
|
}
|
|
case *image.RGBA64:
|
|
f = &filterModel{
|
|
kernel, factor,
|
|
&rgba64Converter{img.(*image.RGBA64)},
|
|
make([]colorArray, sizeX), make([]colorArray, sizeY),
|
|
}
|
|
case *image.Gray:
|
|
f = &filterModel{
|
|
kernel, factor,
|
|
&grayConverter{img.(*image.Gray)},
|
|
make([]colorArray, sizeX), make([]colorArray, sizeY),
|
|
}
|
|
case *image.Gray16:
|
|
f = &filterModel{
|
|
kernel, factor,
|
|
&gray16Converter{img.(*image.Gray16)},
|
|
make([]colorArray, sizeX), make([]colorArray, sizeY),
|
|
}
|
|
case *image.YCbCr:
|
|
f = &filterModel{
|
|
kernel, factor,
|
|
&ycbcrConverter{img.(*image.YCbCr)},
|
|
make([]colorArray, sizeX), make([]colorArray, sizeY),
|
|
}
|
|
}
|
|
|
|
return
|
|
}
|
|
|
|
// Return a filter kernel that performs nearly identically to the provided
|
|
// kernel, but generates and uses a precomputed table rather than executing
|
|
// the kernel for each evaluation. The table is generated with tableSize
|
|
// values that cover the kernal domain from -maxX to +maxX. The input kernel
|
|
// is assumed to be symmetrical around 0, so the table only includes values
|
|
// from 0 to maxX.
|
|
func tableKernel(kernel func(float32) float32, tableSize int,
|
|
maxX float32) func(float32) float32 {
|
|
|
|
// precompute an array of filter coefficients
|
|
weights := make([]float32, tableSize+1)
|
|
for i := range weights {
|
|
weights[i] = kernel(maxX * float32(i) / float32(tableSize))
|
|
}
|
|
weights[tableSize] = 0.0
|
|
|
|
return func(x float32) float32 {
|
|
if x < 0.0 {
|
|
x = -x
|
|
}
|
|
indf := x / maxX * float32(tableSize)
|
|
ind := int(indf)
|
|
if ind >= tableSize {
|
|
return 0.0
|
|
}
|
|
return weights[ind] + (weights[ind+1]-weights[ind])*(indf-float32(ind))
|
|
}
|
|
}
|
|
|
|
// Nearest-neighbor interpolation
|
|
func NearestNeighbor(img image.Image, factor [2]float32) Filter {
|
|
return createFilter(img, factor, 2, func(x float32) (y float32) {
|
|
if x >= -0.5 && x < 0.5 {
|
|
y = 1
|
|
} else {
|
|
y = 0
|
|
}
|
|
|
|
return
|
|
})
|
|
}
|
|
|
|
// Bilinear interpolation
|
|
func Bilinear(img image.Image, factor [2]float32) Filter {
|
|
return createFilter(img, factor, 2, func(x float32) (y float32) {
|
|
absX := float32(math.Abs(float64(x)))
|
|
if absX <= 1 {
|
|
y = 1 - absX
|
|
} else {
|
|
y = 0
|
|
}
|
|
|
|
return
|
|
})
|
|
}
|
|
|
|
func splineKernel(B, C float32) func(float32) float32 {
|
|
factorA := 2.0 - 1.5*B - C
|
|
factorB := -3.0 + 2.0*B + C
|
|
factorC := 1.0 - 1.0/3.0*B
|
|
factorD := -B/6.0 - C
|
|
factorE := B + 5.0*C
|
|
factorF := -2.0*B - 8.0*C
|
|
factorG := 4.0/3.0*B + 4.0*C
|
|
return func(x float32) (y float32) {
|
|
absX := float32(math.Abs(float64(x)))
|
|
if absX <= 1 {
|
|
y = absX*absX*(factorA*absX+factorB) + factorC
|
|
} else if absX <= 2 {
|
|
y = absX*(absX*(absX*factorD+factorE)+factorF) + factorG
|
|
} else {
|
|
y = 0
|
|
}
|
|
|
|
return
|
|
}
|
|
}
|
|
|
|
// Bicubic interpolation (with cubic hermite spline)
|
|
func Bicubic(img image.Image, factor [2]float32) Filter {
|
|
return createFilter(img, factor, 4, splineKernel(0, 0.5))
|
|
}
|
|
|
|
// Mitchell-Netravali interpolation
|
|
func MitchellNetravali(img image.Image, factor [2]float32) Filter {
|
|
return createFilter(img, factor, 4, splineKernel(1.0/3.0, 1.0/3.0))
|
|
}
|
|
|
|
func lanczosKernel(a uint) func(float32) float32 {
|
|
return func(x float32) (y float32) {
|
|
if x > -float32(a) && x < float32(a) {
|
|
y = float32(Sinc(float64(x))) * float32(Sinc(float64(x/float32(a))))
|
|
} else {
|
|
y = 0
|
|
}
|
|
|
|
return
|
|
}
|
|
}
|
|
|
|
const lanczosTableSize = 300
|
|
|
|
// Lanczos interpolation (a=2)
|
|
func Lanczos2(img image.Image, factor [2]float32) Filter {
|
|
return createFilter(img, factor, 4,
|
|
tableKernel(lanczosKernel(2), lanczosTableSize, 2.0))
|
|
}
|
|
|
|
// Lanczos interpolation (a=3)
|
|
func Lanczos3(img image.Image, factor [2]float32) Filter {
|
|
return createFilter(img, factor, 6,
|
|
tableKernel(lanczosKernel(3), lanczosTableSize, 3.0))
|
|
}
|