Java：ロスレスとコンテンツ対応へのフォールバックを試してください
（これまでで最高のロスレス結果！）
この質問を最初に見たとき、これはパズルやチャレンジではなく、プログラムとコードを必死に必要としている誰かだけだと思いました;）しかし、ビジョンの問題を解決するのは私の性質ですので、自分がこのチャレンジを試みるのを止めることはできませんでした！
次のアプローチとアルゴリズムの組み合わせを思いつきました。
擬似コードでは、次のようになります。
function crop(image, desired) {
int sizeChange = 1;
while(sizeChange != 0 and image.width > desired){
Look for a repeating and connected set of lines (top to bottom) with a minimum of x lines
Remove all the lines except for one
sizeChange = image.width - newImage.width
image = newImage;
}
if(image.width > desired){
while(image.width > 2 and image.width > desired){
Create a "pixel energy" map of the image
Find the path from the top of the image to the bottom which "costs" the least amount of "energy"
Remove the lowest cost path from the image
image = newImage;
}
}
}
int desiredWidth = ?
int desiredHeight = ?
Image image = input;
crop(image, desiredWidth);
rotate(image, 90);
crop(image, desiredWidth);
rotate(image, -90);
使用されたテクニック：
- 強度グレースケール
- 膨張
- 等しい列の検索と削除
- シームカービング
- ソーベルエッジ検出
- しきい値処理
プログラム
このプログラムはスクリーンショットをロスレスでトリミングできますが、100％ロスレスではないコンテンツ認識トリミングにフォールバックするオプションがあります。プログラムの引数を調整して、より良い結果を得ることができます。
注：プログラムはさまざまな方法で改善することができます（余暇はあまりありません！）
引数
File name = file
Desired width = number > 0
Desired height = number > 0
Min slice width = number > 1
Compare threshold = number > 0
Use content aware = boolean
Max content aware cycles = number >= 0
コード
import java.awt.Graphics2D;
import java.awt.image.BufferedImage;
import java.awt.image.ColorModel;
import java.io.File;
import java.io.IOException;
import javax.imageio.ImageIO;
import javax.swing.ImageIcon;
import javax.swing.JLabel;
import javax.swing.JOptionPane;
/**
* @author Rolf Smit
* Share and adapt as you like, but don't forget to credit the author!
*/
public class MagicWindowCropper {
public static void main(String[] args) {
if(args.length != 7){
throw new IllegalArgumentException("At least 7 arguments are required: (file, desiredWidth, desiredHeight, minSliceSize, sliceThreshold, forceRemove, maxForceRemove)!");
}
File file = new File(args[0]);
int minSliceSize = Integer.parseInt(args[3]); //4;
int desiredWidth = Integer.parseInt(args[1]); //400;
int desiredHeight = Integer.parseInt(args[2]); //400;
boolean forceRemove = Boolean.parseBoolean(args[5]); //true
int maxForceRemove = Integer.parseInt(args[6]); //40
MagicWindowCropper.MATCH_THRESHOLD = Integer.parseInt(args[4]); //3;
try {
BufferedImage result = ImageIO.read(file);
System.out.println("Horizontal cropping");
//Horizontal crop
result = doDuplicateColumnsMagic(result, minSliceSize, desiredWidth);
if (result.getWidth() != desiredWidth && forceRemove) {
result = doSeamCarvingMagic(result, maxForceRemove, desiredWidth);
}
result = getRotatedBufferedImage(result, false);
System.out.println("Vertical cropping");
//Vertical crop
result = doDuplicateColumnsMagic(result, minSliceSize, desiredHeight);
if (result.getWidth() != desiredHeight && forceRemove) {
result = doSeamCarvingMagic(result, maxForceRemove, desiredHeight);
}
result = getRotatedBufferedImage(result, true);
showBufferedImage("Result", result);
ImageIO.write(result, "png", getNewFileName(file));
} catch (IOException e) {
e.printStackTrace();
}
}
private static BufferedImage doSeamCarvingMagic(BufferedImage inputImage, int max, int desired) {
System.out.println("Seam Carving magic:");
int maxChange = Math.min(inputImage.getWidth() - desired, max);
BufferedImage last = inputImage;
int total = 0, change;
do {
int[][] energy = getPixelEnergyImage(last);
BufferedImage out = removeLowestSeam(energy, last);
change = last.getWidth() - out.getWidth();
total += change;
System.out.println("Carves removed: " + total);
last = out;
} while (change != 0 && total < maxChange);
return last;
}
private static BufferedImage doDuplicateColumnsMagic(BufferedImage inputImage, int minSliceWidth, int desired) {
System.out.println("Duplicate columns magic:");
int maxChange = inputImage.getWidth() - desired;
BufferedImage last = inputImage;
int total = 0, change;
do {
BufferedImage out = removeDuplicateColumn(last, minSliceWidth, desired);
change = last.getWidth() - out.getWidth();
total += change;
System.out.println("Columns removed: " + total);
last = out;
} while (change != 0 && total < maxChange);
return last;
}
/*
* Duplicate column methods
*/
private static BufferedImage removeDuplicateColumn(BufferedImage inputImage, int minSliceWidth, int desiredWidth) {
if (inputImage.getWidth() <= minSliceWidth) {
throw new IllegalStateException("The image width is smaller than the minSliceWidth! What on earth are you trying to do?!");
}
int[] stamp = null;
int sliceStart = -1, sliceEnd = -1;
for (int x = 0; x < inputImage.getWidth() - minSliceWidth + 1; x++) {
stamp = getHorizontalSliceStamp(inputImage, x, minSliceWidth);
if (stamp != null) {
sliceStart = x;
sliceEnd = x + minSliceWidth - 1;
break;
}
}
if (stamp == null) {
return inputImage;
}
BufferedImage out = deepCopyImage(inputImage);
for (int x = sliceEnd + 1; x < inputImage.getWidth(); x++) {
int[] row = getHorizontalSliceStamp(inputImage, x, 1);
if (equalsRows(stamp, row)) {
sliceEnd = x;
} else {
break;
}
}
//Remove policy
int canRemove = sliceEnd - (sliceStart + 1) + 1;
int mayRemove = inputImage.getWidth() - desiredWidth;
int dif = mayRemove - canRemove;
if (dif < 0) {
sliceEnd += dif;
}
int mustRemove = sliceEnd - (sliceStart + 1) + 1;
if (mustRemove <= 0) {
return out;
}
out = removeHorizontalRegion(out, sliceStart + 1, sliceEnd);
out = removeLeft(out, out.getWidth() - mustRemove);
return out;
}
private static BufferedImage removeHorizontalRegion(BufferedImage image, int startX, int endX) {
int width = endX - startX + 1;
if (endX + 1 > image.getWidth()) {
endX = image.getWidth() - 1;
}
if (endX < startX) {
throw new IllegalStateException("Invalid removal parameters! Wow this error message is genius!");
}
BufferedImage out = deepCopyImage(image);
for (int x = endX + 1; x < image.getWidth(); x++) {
for (int y = 0; y < image.getHeight(); y++) {
out.setRGB(x - width, y, image.getRGB(x, y));
out.setRGB(x, y, 0xFF000000);
}
}
return out;
}
private static int[] getHorizontalSliceStamp(BufferedImage inputImage, int startX, int sliceWidth) {
int[] initial = new int[inputImage.getHeight()];
for (int y = 0; y < inputImage.getHeight(); y++) {
initial[y] = inputImage.getRGB(startX, y);
}
if (sliceWidth == 1) {
return initial;
}
for (int s = 1; s < sliceWidth; s++) {
int[] row = new int[inputImage.getHeight()];
for (int y = 0; y < inputImage.getHeight(); y++) {
row[y] = inputImage.getRGB(startX + s, y);
}
if (!equalsRows(initial, row)) {
return null;
}
}
return initial;
}
private static int MATCH_THRESHOLD = 3;
private static boolean equalsRows(int[] left, int[] right) {
for (int i = 0; i < left.length; i++) {
int rl = (left[i]) & 0xFF;
int gl = (left[i] >> 8) & 0xFF;
int bl = (left[i] >> 16) & 0xFF;
int rr = (right[i]) & 0xFF;
int gr = (right[i] >> 8) & 0xFF;
int br = (right[i] >> 16) & 0xFF;
if (Math.abs(rl - rr) > MATCH_THRESHOLD
|| Math.abs(gl - gr) > MATCH_THRESHOLD
|| Math.abs(bl - br) > MATCH_THRESHOLD) {
return false;
}
}
return true;
}
/*
* Seam carving methods
*/
private static BufferedImage removeLowestSeam(int[][] input, BufferedImage image) {
int lowestValue = Integer.MAX_VALUE; //Integer overflow possible when image height grows!
int lowestValueX = -1;
// Here be dragons
for (int x = 1; x < input.length - 1; x++) {
int seamX = x;
int value = input[x][0];
for (int y = 1; y < input[x].length; y++) {
if (seamX < 1) {
int top = input[seamX][y];
int right = input[seamX + 1][y];
if (top <= right) {
value += top;
} else {
seamX++;
value += right;
}
} else if (seamX > input.length - 2) {
int top = input[seamX][y];
int left = input[seamX - 1][y];
if (top <= left) {
value += top;
} else {
seamX--;
value += left;
}
} else {
int left = input[seamX - 1][y];
int top = input[seamX][y];
int right = input[seamX + 1][y];
if (top <= left && top <= right) {
value += top;
} else if (left <= top && left <= right) {
seamX--;
value += left;
} else {
seamX++;
value += right;
}
}
}
if (value < lowestValue) {
lowestValue = value;
lowestValueX = x;
}
}
BufferedImage out = deepCopyImage(image);
int seamX = lowestValueX;
shiftRow(out, seamX, 0);
for (int y = 1; y < input[seamX].length; y++) {
if (seamX < 1) {
int top = input[seamX][y];
int right = input[seamX + 1][y];
if (top <= right) {
shiftRow(out, seamX, y);
} else {
seamX++;
shiftRow(out, seamX, y);
}
} else if (seamX > input.length - 2) {
int top = input[seamX][y];
int left = input[seamX - 1][y];
if (top <= left) {
shiftRow(out, seamX, y);
} else {
seamX--;
shiftRow(out, seamX, y);
}
} else {
int left = input[seamX - 1][y];
int top = input[seamX][y];
int right = input[seamX + 1][y];
if (top <= left && top <= right) {
shiftRow(out, seamX, y);
} else if (left <= top && left <= right) {
seamX--;
shiftRow(out, seamX, y);
} else {
seamX++;
shiftRow(out, seamX, y);
}
}
}
return removeLeft(out, out.getWidth() - 1);
}
private static void shiftRow(BufferedImage image, int startX, int y) {
for (int x = startX; x < image.getWidth() - 1; x++) {
image.setRGB(x, y, image.getRGB(x + 1, y));
}
}
private static int[][] getPixelEnergyImage(BufferedImage image) {
// Convert Image to gray scale using the luminosity method and add extra
// edges for the Sobel filter
int[][] grayScale = new int[image.getWidth() + 2][image.getHeight() + 2];
for (int x = 0; x < image.getWidth(); x++) {
for (int y = 0; y < image.getHeight(); y++) {
int rgb = image.getRGB(x, y);
int r = (rgb >> 16) & 0xFF;
int g = (rgb >> 8) & 0xFF;
int b = (rgb & 0xFF);
int luminosity = (int) (0.21 * r + 0.72 * g + 0.07 * b);
grayScale[x + 1][y + 1] = luminosity;
}
}
// Sobel edge detection
final double[] kernelHorizontalEdges = new double[] { 1, 2, 1, 0, 0, 0, -1, -2, -1 };
final double[] kernelVerticalEdges = new double[] { 1, 0, -1, 2, 0, -2, 1, 0, -1 };
int[][] energyImage = new int[image.getWidth()][image.getHeight()];
for (int x = 1; x < image.getWidth() + 1; x++) {
for (int y = 1; y < image.getHeight() + 1; y++) {
int k = 0;
double horizontal = 0;
for (int ky = -1; ky < 2; ky++) {
for (int kx = -1; kx < 2; kx++) {
horizontal += ((double) grayScale[x + kx][y + ky] * kernelHorizontalEdges[k]);
k++;
}
}
double vertical = 0;
k = 0;
for (int ky = -1; ky < 2; ky++) {
for (int kx = -1; kx < 2; kx++) {
vertical += ((double) grayScale[x + kx][y + ky] * kernelVerticalEdges[k]);
k++;
}
}
if (Math.sqrt(horizontal * horizontal + vertical * vertical) > 127) {
energyImage[x - 1][y - 1] = 255;
} else {
energyImage[x - 1][y - 1] = 0;
}
}
}
//Dilate the edge detected image a few times for better seaming results
//Current value is just 1...
for (int i = 0; i < 1; i++) {
dilateImage(energyImage);
}
return energyImage;
}
private static void dilateImage(int[][] image) {
for (int x = 0; x < image.length; x++) {
for (int y = 0; y < image[x].length; y++) {
if (image[x][y] == 255) {
if (x > 0 && image[x - 1][y] == 0) {
image[x - 1][y] = 2; //Note: 2 is just a placeholder value
}
if (y > 0 && image[x][y - 1] == 0) {
image[x][y - 1] = 2;
}
if (x + 1 < image.length && image[x + 1][y] == 0) {
image[x + 1][y] = 2;
}
if (y + 1 < image[x].length && image[x][y + 1] == 0) {
image[x][y + 1] = 2;
}
}
}
}
for (int x = 0; x < image.length; x++) {
for (int y = 0; y < image[x].length; y++) {
if (image[x][y] == 2) {
image[x][y] = 255;
}
}
}
}
/*
* Utilities
*/
private static void showBufferedImage(String windowTitle, BufferedImage image) {
JOptionPane.showMessageDialog(null, new JLabel(new ImageIcon(image)), windowTitle, JOptionPane.PLAIN_MESSAGE, null);
}
private static BufferedImage deepCopyImage(BufferedImage input) {
ColorModel cm = input.getColorModel();
return new BufferedImage(cm, input.copyData(null), cm.isAlphaPremultiplied(), null);
}
private static final BufferedImage getRotatedBufferedImage(BufferedImage img, boolean back) {
double oldW = img.getWidth(), oldH = img.getHeight();
double newW = img.getHeight(), newH = img.getWidth();
BufferedImage out = new BufferedImage((int) newW, (int) newH, img.getType());
Graphics2D g = out.createGraphics();
g.translate((newW - oldW) / 2.0, (newH - oldH) / 2.0);
g.rotate(Math.toRadians(back ? -90 : 90), oldW / 2.0, oldH / 2.0);
g.drawRenderedImage(img, null);
g.dispose();
return out;
}
private static BufferedImage removeLeft(BufferedImage image, int startX) {
int removeWidth = image.getWidth() - startX;
BufferedImage out = new BufferedImage(image.getWidth() - removeWidth,
image.getHeight(), image.getType());
for (int x = 0; x < startX; x++) {
for (int y = 0; y < out.getHeight(); y++) {
out.setRGB(x, y, image.getRGB(x, y));
}
}
return out;
}
private static File getNewFileName(File in) {
String name = in.getName();
int i = name.lastIndexOf(".");
if (i != -1) {
String ext = name.substring(i);
String n = name.substring(0, i);
return new File(in.getParentFile(), n + "-cropped" + ext);
} else {
return new File(in.getParentFile(), name + "-cropped");
}
}
}
結果
希望のサイズなしのXPスクリーンショットロスレス（最大ロスレス圧縮）
引数： "image.png" 1 1 5 10 false 0
結果： 836 x 323
XPスクリーンショットを800x600に
引数： "image.png" 800600 6 10 true 60
結果： 800 x 600
ロスレスアルゴリズムは、約155の水平線を削除しますが、一部のアーティファクトが見られるため、アルゴリズムはコンテンツ認識の削除にフォールバックします。
700x300へのWindows 10スクリーンショット
引数： "image.png" 700 300 6 10 true 60
結果： 700 x 300
ロスレスアルゴリズムは270本の水平線を削除しますが、アルゴリズムはコンテンツに対応した削除にフォールバックし、29本を削除します。垂直ロスレスアルゴリズムのみが使用されます。
400x200（テスト）に対応したWindows 10スクリーンショット
引数： "image.png" 400 200 5 10 true 600
結果： 400 x 200
これは、結果認識された画像がコンテンツ認識機能の厳しい使用後にどのように見えるかを確認するためのテストでした。結果はひどく破損していますが、認識できません。
