HybridBinarizer port and builds

This commit is contained in:
Henry Schimke
2022-09-02 16:14:22 -05:00
parent 1b5ef1e51b
commit 833e7a5347
2 changed files with 310 additions and 237 deletions

View File

@@ -1,237 +0,0 @@
/*
* Copyright 2009 ZXing authors
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
package com.google.zxing.common;
import com.google.zxing.Binarizer;
import com.google.zxing.LuminanceSource;
import com.google.zxing.NotFoundException;
/**
* This class implements a local thresholding algorithm, which while slower than the
* GlobalHistogramBinarizer, is fairly efficient for what it does. It is designed for
* high frequency images of barcodes with black data on white backgrounds. For this application,
* it does a much better job than a global blackpoint with severe shadows and gradients.
* However it tends to produce artifacts on lower frequency images and is therefore not
* a good general purpose binarizer for uses outside ZXing.
*
* This class extends GlobalHistogramBinarizer, using the older histogram approach for 1D readers,
* and the newer local approach for 2D readers. 1D decoding using a per-row histogram is already
* inherently local, and only fails for horizontal gradients. We can revisit that problem later,
* but for now it was not a win to use local blocks for 1D.
*
* This Binarizer is the default for the unit tests and the recommended class for library users.
*
* @author dswitkin@google.com (Daniel Switkin)
*/
public final class HybridBinarizer extends GlobalHistogramBinarizer {
// This class uses 5x5 blocks to compute local luminance, where each block is 8x8 pixels.
// So this is the smallest dimension in each axis we can accept.
private static final int BLOCK_SIZE_POWER = 3;
private static final int BLOCK_SIZE = 1 << BLOCK_SIZE_POWER; // ...0100...00
private static final int BLOCK_SIZE_MASK = BLOCK_SIZE - 1; // ...0011...11
private static final int MINIMUM_DIMENSION = BLOCK_SIZE * 5;
private static final int MIN_DYNAMIC_RANGE = 24;
private BitMatrix matrix;
public HybridBinarizer(LuminanceSource source) {
super(source);
}
/**
* Calculates the final BitMatrix once for all requests. This could be called once from the
* constructor instead, but there are some advantages to doing it lazily, such as making
* profiling easier, and not doing heavy lifting when callers don't expect it.
*/
@Override
public BitMatrix getBlackMatrix() throws NotFoundException {
if (matrix != null) {
return matrix;
}
LuminanceSource source = getLuminanceSource();
int width = source.getWidth();
int height = source.getHeight();
if (width >= MINIMUM_DIMENSION && height >= MINIMUM_DIMENSION) {
byte[] luminances = source.getMatrix();
int subWidth = width >> BLOCK_SIZE_POWER;
if ((width & BLOCK_SIZE_MASK) != 0) {
subWidth++;
}
int subHeight = height >> BLOCK_SIZE_POWER;
if ((height & BLOCK_SIZE_MASK) != 0) {
subHeight++;
}
int[][] blackPoints = calculateBlackPoints(luminances, subWidth, subHeight, width, height);
BitMatrix newMatrix = new BitMatrix(width, height);
calculateThresholdForBlock(luminances, subWidth, subHeight, width, height, blackPoints, newMatrix);
matrix = newMatrix;
} else {
// If the image is too small, fall back to the global histogram approach.
matrix = super.getBlackMatrix();
}
return matrix;
}
@Override
public Binarizer createBinarizer(LuminanceSource source) {
return new HybridBinarizer(source);
}
/**
* For each block in the image, calculate the average black point using a 5x5 grid
* of the blocks around it. Also handles the corner cases (fractional blocks are computed based
* on the last pixels in the row/column which are also used in the previous block).
*/
private static void calculateThresholdForBlock(byte[] luminances,
int subWidth,
int subHeight,
int width,
int height,
int[][] blackPoints,
BitMatrix matrix) {
int maxYOffset = height - BLOCK_SIZE;
int maxXOffset = width - BLOCK_SIZE;
for (int y = 0; y < subHeight; y++) {
int yoffset = y << BLOCK_SIZE_POWER;
if (yoffset > maxYOffset) {
yoffset = maxYOffset;
}
int top = cap(y, subHeight - 3);
for (int x = 0; x < subWidth; x++) {
int xoffset = x << BLOCK_SIZE_POWER;
if (xoffset > maxXOffset) {
xoffset = maxXOffset;
}
int left = cap(x, subWidth - 3);
int sum = 0;
for (int z = -2; z <= 2; z++) {
int[] blackRow = blackPoints[top + z];
sum += blackRow[left - 2] + blackRow[left - 1] + blackRow[left] + blackRow[left + 1] + blackRow[left + 2];
}
int average = sum / 25;
thresholdBlock(luminances, xoffset, yoffset, average, width, matrix);
}
}
}
private static int cap(int value, int max) {
return value < 2 ? 2 : Math.min(value, max);
}
/**
* Applies a single threshold to a block of pixels.
*/
private static void thresholdBlock(byte[] luminances,
int xoffset,
int yoffset,
int threshold,
int stride,
BitMatrix matrix) {
for (int y = 0, offset = yoffset * stride + xoffset; y < BLOCK_SIZE; y++, offset += stride) {
for (int x = 0; x < BLOCK_SIZE; x++) {
// Comparison needs to be <= so that black == 0 pixels are black even if the threshold is 0.
if ((luminances[offset + x] & 0xFF) <= threshold) {
matrix.set(xoffset + x, yoffset + y);
}
}
}
}
/**
* Calculates a single black point for each block of pixels and saves it away.
* See the following thread for a discussion of this algorithm:
* http://groups.google.com/group/zxing/browse_thread/thread/d06efa2c35a7ddc0
*/
private static int[][] calculateBlackPoints(byte[] luminances,
int subWidth,
int subHeight,
int width,
int height) {
int maxYOffset = height - BLOCK_SIZE;
int maxXOffset = width - BLOCK_SIZE;
int[][] blackPoints = new int[subHeight][subWidth];
for (int y = 0; y < subHeight; y++) {
int yoffset = y << BLOCK_SIZE_POWER;
if (yoffset > maxYOffset) {
yoffset = maxYOffset;
}
for (int x = 0; x < subWidth; x++) {
int xoffset = x << BLOCK_SIZE_POWER;
if (xoffset > maxXOffset) {
xoffset = maxXOffset;
}
int sum = 0;
int min = 0xFF;
int max = 0;
for (int yy = 0, offset = yoffset * width + xoffset; yy < BLOCK_SIZE; yy++, offset += width) {
for (int xx = 0; xx < BLOCK_SIZE; xx++) {
int pixel = luminances[offset + xx] & 0xFF;
sum += pixel;
// still looking for good contrast
if (pixel < min) {
min = pixel;
}
if (pixel > max) {
max = pixel;
}
}
// short-circuit min/max tests once dynamic range is met
if (max - min > MIN_DYNAMIC_RANGE) {
// finish the rest of the rows quickly
for (yy++, offset += width; yy < BLOCK_SIZE; yy++, offset += width) {
for (int xx = 0; xx < BLOCK_SIZE; xx++) {
sum += luminances[offset + xx] & 0xFF;
}
}
}
}
// The default estimate is the average of the values in the block.
int average = sum >> (BLOCK_SIZE_POWER * 2);
if (max - min <= MIN_DYNAMIC_RANGE) {
// If variation within the block is low, assume this is a block with only light or only
// dark pixels. In that case we do not want to use the average, as it would divide this
// low contrast area into black and white pixels, essentially creating data out of noise.
//
// The default assumption is that the block is light/background. Since no estimate for
// the level of dark pixels exists locally, use half the min for the block.
average = min / 2;
if (y > 0 && x > 0) {
// Correct the "white background" assumption for blocks that have neighbors by comparing
// the pixels in this block to the previously calculated black points. This is based on
// the fact that dark barcode symbology is always surrounded by some amount of light
// background for which reasonable black point estimates were made. The bp estimated at
// the boundaries is used for the interior.
// The (min < bp) is arbitrary but works better than other heuristics that were tried.
int averageNeighborBlackPoint =
(blackPoints[y - 1][x] + (2 * blackPoints[y][x - 1]) + blackPoints[y - 1][x - 1]) / 4;
if (min < averageNeighborBlackPoint) {
average = averageNeighborBlackPoint;
}
}
}
blackPoints[y][x] = average;
}
}
return blackPoints;
}
}

View File

@@ -3755,3 +3755,313 @@ impl GlobalHistogramBinarizer {
Ok((bestValley as u32) << GlobalHistogramBinarizer::LUMINANCE_SHIFT)
}
}
/*
* Copyright 2009 ZXing authors
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
// package com.google.zxing.common;
// import com.google.zxing.Binarizer;
// import com.google.zxing.LuminanceSource;
// import com.google.zxing.NotFoundException;
/**
* This class implements a local thresholding algorithm, which while slower than the
* GlobalHistogramBinarizer, is fairly efficient for what it does. It is designed for
* high frequency images of barcodes with black data on white backgrounds. For this application,
* it does a much better job than a global blackpoint with severe shadows and gradients.
* However it tends to produce artifacts on lower frequency images and is therefore not
* a good general purpose binarizer for uses outside ZXing.
*
* This class extends GlobalHistogramBinarizer, using the older histogram approach for 1D readers,
* and the newer local approach for 2D readers. 1D decoding using a per-row histogram is already
* inherently local, and only fails for horizontal gradients. We can revisit that problem later,
* but for now it was not a win to use local blocks for 1D.
*
* This Binarizer is the default for the unit tests and the recommended class for library users.
*
* @author dswitkin@google.com (Daniel Switkin)
*/
pub struct HybridBinarizer {
//width: usize,
//height: usize,
//source: Box<dyn LuminanceSource>,
ghb: GlobalHistogramBinarizer,
// matrix :Option<BitMatrix>,
}
impl Binarizer for HybridBinarizer {
fn getLuminanceSource(&self) -> &Box<dyn LuminanceSource> {
self.ghb.getLuminanceSource()
}
fn getBlackRow(&self, y: usize, row: &mut BitArray) -> Result<BitArray, Exceptions> {
self.ghb.getBlackRow(y, row)
}
/**
* Calculates the final BitMatrix once for all requests. This could be called once from the
* constructor instead, but there are some advantages to doing it lazily, such as making
* profiling easier, and not doing heavy lifting when callers don't expect it.
*/
fn getBlackMatrix(&self) -> Result<BitMatrix, Exceptions> {
// if self.matrix.is_some() {
// return Ok(self.matrix.clone().unwrap())
// }
let matrix;
let source = self.getLuminanceSource();
let width = source.getWidth();
let height = source.getHeight();
if width >= HybridBinarizer::MINIMUM_DIMENSION
&& height >= HybridBinarizer::MINIMUM_DIMENSION
{
let luminances = source.getMatrix();
let mut subWidth = width >> HybridBinarizer::BLOCK_SIZE_POWER;
if (width & HybridBinarizer::BLOCK_SIZE_MASK) != 0 {
subWidth += 1;
}
let mut subHeight = height >> HybridBinarizer::BLOCK_SIZE_POWER;
if (height & HybridBinarizer::BLOCK_SIZE_MASK) != 0 {
subHeight += 1;
}
let blackPoints = Self::calculateBlackPoints(
&luminances,
subWidth as u32,
subHeight as u32,
width as u32,
height as u32,
);
let mut newMatrix = BitMatrix::new(width as u32, height as u32)?;
Self::calculateThresholdForBlock(
&luminances,
subWidth as u32,
subHeight as u32,
width as u32,
height as u32,
&blackPoints,
&mut newMatrix,
);
matrix = newMatrix;
} else {
// If the image is too small, fall back to the global histogram approach.
matrix = self.ghb.getBlackMatrix()?;
}
Ok(matrix)
}
fn createBinarizer(&self, source: Box<dyn LuminanceSource>) -> Box<dyn Binarizer> {
Box::new(HybridBinarizer::new(source))
}
fn getWidth(&self) -> usize {
self.ghb.getWidth()
}
fn getHeight(&self) -> usize {
self.ghb.getHeight()
}
}
impl HybridBinarizer {
// This class uses 5x5 blocks to compute local luminance, where each block is 8x8 pixels.
// So this is the smallest dimension in each axis we can accept.
const BLOCK_SIZE_POWER: usize = 3;
const BLOCK_SIZE: usize = 1 << HybridBinarizer::BLOCK_SIZE_POWER; // ...0100...00
const BLOCK_SIZE_MASK: usize = HybridBinarizer::BLOCK_SIZE - 1; // ...0011...11
const MINIMUM_DIMENSION: usize = HybridBinarizer::BLOCK_SIZE * 5;
const MIN_DYNAMIC_RANGE: usize = 24;
pub fn new(source: Box<dyn LuminanceSource>) -> Self {
Self {
ghb: GlobalHistogramBinarizer::new(source),
// matrix: None,
}
}
/**
* For each block in the image, calculate the average black point using a 5x5 grid
* of the blocks around it. Also handles the corner cases (fractional blocks are computed based
* on the last pixels in the row/column which are also used in the previous block).
*/
fn calculateThresholdForBlock(
luminances: &[u8],
subWidth: u32,
subHeight: u32,
width: u32,
height: u32,
blackPoints: &Vec<Vec<u32>>,
matrix: &mut BitMatrix,
) {
let maxYOffset = height - HybridBinarizer::BLOCK_SIZE as u32;
let maxXOffset = width - HybridBinarizer::BLOCK_SIZE as u32;
for y in 0..subHeight {
// for (int y = 0; y < subHeight; y++) {
let mut yoffset = y << HybridBinarizer::BLOCK_SIZE_POWER;
if yoffset > maxYOffset {
yoffset = maxYOffset;
}
let top = Self::cap(y, subHeight - 3);
for x in 0..subWidth {
// for (int x = 0; x < subWidth; x++) {
let mut xoffset = x << HybridBinarizer::BLOCK_SIZE_POWER;
if xoffset > maxXOffset {
xoffset = maxXOffset;
}
let left = Self::cap(x, subWidth - 3);
let mut sum = 0;
for z in -2i32..=2 {
// for (int z = -2; z <= 2; z++) {
let blackRow = &blackPoints[(top as i32 + z) as usize];
sum += blackRow[(left - 2) as usize]
+ blackRow[(left - 1) as usize]
+ blackRow[left as usize]
+ blackRow[(left + 1) as usize]
+ blackRow[(left + 2) as usize];
}
let average = sum / 25;
Self::thresholdBlock(luminances, xoffset, yoffset, average, width, matrix);
}
}
}
fn cap(value: u32, max: u32) -> u32 {
if value < 2 {
2
} else {
value.min(max)
}
}
/**
* Applies a single threshold to a block of pixels.
*/
fn thresholdBlock(
luminances: &[u8],
xoffset: u32,
yoffset: u32,
threshold: u32,
stride: u32,
matrix: &mut BitMatrix,
) {
let mut offset = yoffset * stride + xoffset;
for y in 0..HybridBinarizer::BLOCK_SIZE {
// for (int y = 0, offset = yoffset * stride + xoffset; y < HybridBinarizer::BLOCK_SIZE; y++, offset += stride) {
for x in 0..HybridBinarizer::BLOCK_SIZE {
// for (int x = 0; x < HybridBinarizer::BLOCK_SIZE; x++) {
// Comparison needs to be <= so that black == 0 pixels are black even if the threshold is 0.
if luminances[offset as usize + x] as u32 <= threshold {
matrix.set(xoffset + x as u32, yoffset + y as u32);
}
}
offset += stride;
}
}
/**
* Calculates a single black point for each block of pixels and saves it away.
* See the following thread for a discussion of this algorithm:
* http://groups.google.com/group/zxing/browse_thread/thread/d06efa2c35a7ddc0
*/
fn calculateBlackPoints(
luminances: &[u8],
subWidth: u32,
subHeight: u32,
width: u32,
height: u32,
) -> Vec<Vec<u32>> {
let maxYOffset = height as usize - HybridBinarizer::BLOCK_SIZE;
let maxXOffset = width as usize - HybridBinarizer::BLOCK_SIZE;
let mut blackPoints = vec![vec![0; subWidth as usize]; subHeight as usize];
for y in 0..subHeight {
// for (int y = 0; y < subHeight; y++) {
let mut yoffset = y << HybridBinarizer::BLOCK_SIZE_POWER;
if yoffset > maxYOffset as u32 {
yoffset = maxYOffset as u32;
}
for x in 0..subWidth {
// for (int x = 0; x < subWidth; x++) {
let mut xoffset = x << HybridBinarizer::BLOCK_SIZE_POWER;
if xoffset > maxXOffset as u32 {
xoffset = maxXOffset as u32;
}
let mut sum = 0;
let mut min = 0xFF;
let mut max = 0;
let mut offset = yoffset * width + xoffset;
for yy in 0..HybridBinarizer::BLOCK_SIZE {
// for (int yy = 0, offset = yoffset * width + xoffset; yy < HybridBinarizer::BLOCK_SIZE; yy++, offset += width) {
for xx in 0..HybridBinarizer::BLOCK_SIZE {
// for (int xx = 0; xx < HybridBinarizer::BLOCK_SIZE; xx++) {
let pixel = luminances[offset as usize + xx];
sum += pixel;
// still looking for good contrast
if pixel < min {
min = pixel;
}
if pixel > max {
max = pixel;
}
}
// short-circuit min/max tests once dynamic range is met
if (max - min) as usize > HybridBinarizer::MIN_DYNAMIC_RANGE {
// finish the rest of the rows quickly
offset += width;
for _yy_s in yy + 1..HybridBinarizer::BLOCK_SIZE {
// for (yy++, offset += width; yy < HybridBinarizer::BLOCK_SIZE; yy++, offset += width) {
for xx in 0..HybridBinarizer::BLOCK_SIZE {
// for (int xx = 0; xx < BLOCK_SIZE; xx++) {
sum += luminances[offset as usize + xx];
}
offset += width;
}
break;
}
offset += width;
}
// The default estimate is the average of the values in the block.
let mut average = sum >> (HybridBinarizer::BLOCK_SIZE_POWER * 2);
if (max - min) as usize <= HybridBinarizer::MIN_DYNAMIC_RANGE {
// If variation within the block is low, assume this is a block with only light or only
// dark pixels. In that case we do not want to use the average, as it would divide this
// low contrast area into black and white pixels, essentially creating data out of noise.
//
// The default assumption is that the block is light/background. Since no estimate for
// the level of dark pixels exists locally, use half the min for the block.
average = min / 2;
if y > 0 && x > 0 {
// Correct the "white background" assumption for blocks that have neighbors by comparing
// the pixels in this block to the previously calculated black points. This is based on
// the fact that dark barcode symbology is always surrounded by some amount of light
// background for which reasonable black point estimates were made. The bp estimated at
// the boundaries is used for the interior.
// The (min < bp) is arbitrary but works better than other heuristics that were tried.
let averageNeighborBlackPoint = (blackPoints[y as usize - 1][x as usize]
+ (2 * blackPoints[y as usize][x as usize - 1])
+ blackPoints[y as usize - 1][x as usize - 1])
/ 4;
if (min < averageNeighborBlackPoint) {
average = averageNeighborBlackPoint;
}
}
}
blackPoints[y as usize][x as usize] = average;
}
}
return blackPoints.into_iter().map(|x| x.iter().map(|y| *y as u32).collect()).collect();
}
}