/**
* Constructs a TemplateMatcher object. If a mask shape is specified, then only pixels that are
* entirely inside the mask are included in the template.
*
* @param image the image to match
* @param maskShape a shape to define inside pixels (may be null)
*/
public TemplateMatcher(BufferedImage image, Shape maskShape) {
original = new BufferedImage(image.getWidth(), image.getHeight(), BufferedImage.TYPE_INT_ARGB);
original.createGraphics().drawImage(image, 0, 0, null);
mask = maskShape;
getTemplate();
// set up the Gaussian curve fitter
dataset = new Dataset();
fitter = new DatasetCurveFitter(dataset);
fitter.setAutofit(true);
f = new UserFunction("gaussian"); // $NON-NLS-1$
f.setParameters(
new String[] {"a", "b", "c"}, // $NON-NLS-1$ //$NON-NLS-2$ //$NON-NLS-3$
new double[] {1, 0, 1});
f.setExpression("a*exp(-(x-b)^2/c)", new String[] {"x"}); // $NON-NLS-1$ //$NON-NLS-2$
}
/**
* Gets the template location at which the best match occurs in a rectangle. May return null.
*
* @param target the image to search
* @param searchRect the rectangle to search within the target image
* @return the optimized template location at which the best match, if any, is found
*/
public TPoint getMatchLocation(BufferedImage target, Rectangle searchRect) {
wTarget = target.getWidth();
hTarget = target.getHeight();
// determine insets needed to accommodate template
int left = wTemplate / 2, right = left;
if (wTemplate % 2 > 0) right++;
int top = hTemplate / 2, bottom = top;
if (hTemplate % 2 > 0) bottom++;
// trim search rectangle if necessary
searchRect.x = Math.max(left, Math.min(wTarget - right, searchRect.x));
searchRect.y = Math.max(top, Math.min(hTarget - bottom, searchRect.y));
searchRect.width = Math.min(wTarget - searchRect.x - right, searchRect.width);
searchRect.height = Math.min(hTarget - searchRect.y - bottom, searchRect.height);
if (searchRect.width <= 0 || searchRect.height <= 0) {
peakHeight = Double.NaN;
peakWidth = Double.NaN;
return null;
}
// set up test pixels to search (rectangle plus template)
int xMin = Math.max(0, searchRect.x - left);
int xMax = Math.min(wTarget, searchRect.x + searchRect.width + right);
int yMin = Math.max(0, searchRect.y - top);
int yMax = Math.min(hTarget, searchRect.y + searchRect.height + bottom);
wTest = xMax - xMin;
hTest = yMax - yMin;
if (target.getType() != BufferedImage.TYPE_INT_RGB) {
BufferedImage image = new BufferedImage(wTarget, hTarget, BufferedImage.TYPE_INT_RGB);
image.createGraphics().drawImage(target, 0, 0, null);
target = image;
}
targetPixels = new int[wTest * hTest];
target.getRaster().getDataElements(xMin, yMin, wTest, hTest, targetPixels);
// find the rectangle point with the minimum difference
double matchDiff = largeNumber; // larger than typical differences
int xMatch = 0, yMatch = 0;
double avgDiff = 0;
for (int x = 0; x <= searchRect.width; x++) {
for (int y = 0; y <= searchRect.height; y++) {
double diff = getDifferenceAtTestPoint(x, y);
avgDiff += diff;
if (diff < matchDiff) {
matchDiff = diff;
xMatch = x;
yMatch = y;
}
}
}
avgDiff /= (searchRect.width * searchRect.height);
peakHeight = avgDiff / matchDiff - 1;
peakWidth = Double.NaN;
double dx = 0, dy = 0;
// if match is not exact, fit a Gaussian and find peak
if (!Double.isInfinite(peakHeight)) {
// fill data arrays
xValues[1] = yValues[1] = peakHeight;
for (int i = -1; i < 2; i++) {
if (i == 0) continue;
double diff = getDifferenceAtTestPoint(xMatch + i, yMatch);
xValues[i + 1] = avgDiff / diff - 1;
diff = getDifferenceAtTestPoint(xMatch, yMatch + i);
yValues[i + 1] = avgDiff / diff - 1;
}
// estimate peakHeight = peak of gaussian
// estimate offset dx of gaussian
double pull = 1 / (xValues[1] - xValues[0]);
double push = 1 / (xValues[1] - xValues[2]);
if (Double.isNaN(pull)) pull = LARGE_NUMBER;
if (Double.isNaN(push)) push = LARGE_NUMBER;
dx = 0.6 * (push - pull) / (push + pull);
// estimate width wx of gaussian
double ratio = dx > 0 ? peakHeight / xValues[0] : peakHeight / xValues[2];
double wx = dx > 0 ? dx + 1 : dx - 1;
wx = wx * wx / Math.log(ratio);
// estimate offset dy of gaussian
pull = 1 / (yValues[1] - yValues[0]);
push = 1 / (yValues[1] - yValues[2]);
if (Double.isNaN(pull)) pull = LARGE_NUMBER;
if (Double.isNaN(push)) push = LARGE_NUMBER;
dy = 0.6 * (push - pull) / (push + pull);
// estimate width wy of gaussian
ratio = dy > 0 ? peakHeight / yValues[0] : peakHeight / yValues[2];
double wy = dy > 0 ? dy + 1 : dy - 1;
wy = wy * wy / Math.log(ratio);
// set x parameters and fit to x data
dataset.clear();
dataset.append(pixelOffsets, xValues);
double rmsDev = 1;
for (int k = 0; k < 3; k++) {
double c = k == 0 ? wx : k == 1 ? wx / 3 : wx * 3;
f.setParameterValue(0, peakHeight);
f.setParameterValue(1, dx);
f.setParameterValue(2, c);
rmsDev = fitter.fit(f);
if (rmsDev < 0.01) { // fitter succeeded (3-point fit should be exact)
dx = f.getParameterValue(1);
peakWidth = f.getParameterValue(2);
break;
}
}
if (!Double.isNaN(peakWidth)) {
// set y parameters and fit to y data
dataset.clear();
dataset.append(pixelOffsets, yValues);
for (int k = 0; k < 3; k++) {
double c = k == 0 ? wy : k == 1 ? wy / 3 : wy * 3;
f.setParameterValue(0, peakHeight);
f.setParameterValue(1, dx);
f.setParameterValue(2, c);
rmsDev = fitter.fit(f);
if (rmsDev < 0.01) { // fitter succeeded (3-point fit should be exact)
dy = f.getParameterValue(1);
peakWidth = (peakWidth + f.getParameterValue(2)) / 2;
break;
}
}
if (rmsDev > 0.01) peakWidth = Double.NaN;
}
}
double xImage = xMatch + searchRect.x - left - trimLeft + dx;
double yImage = yMatch + searchRect.y - top - trimTop + dy;
return new TPoint(xImage, yImage);
}
/**
* Gets the template location at which the best match occurs in a rectangle and along a line. May
* return null.
*
* @param target the image to search
* @param searchRect the rectangle to search within the target image
* @param x0 the x-component of a point on the line
* @param y0 the y-component of a point on the line
* @param slope the slope of the line
* @param spread the spread of the line (line width = 1+2*spread)
* @return the optimized template location of the best match, if any
*/
public TPoint getMatchLocation(
BufferedImage target, Rectangle searchRect, double x0, double y0, double theta, int spread) {
wTarget = target.getWidth();
hTarget = target.getHeight();
// determine insets needed to accommodate template
int left = wTemplate / 2, right = left;
if (wTemplate % 2 > 0) right++;
int top = hTemplate / 2, bottom = top;
if (hTemplate % 2 > 0) bottom++;
// trim search rectangle if necessary
searchRect.x = Math.max(left, Math.min(wTarget - right, searchRect.x));
searchRect.y = Math.max(top, Math.min(hTarget - bottom, searchRect.y));
searchRect.width = Math.min(wTarget - searchRect.x - right, searchRect.width);
searchRect.height = Math.min(hTarget - searchRect.y - bottom, searchRect.height);
if (searchRect.width <= 0 || searchRect.height <= 0) {
peakHeight = Double.NaN;
peakWidth = Double.NaN;
return null;
}
// set up test pixels to search (rectangle plus template)
int xMin = Math.max(0, searchRect.x - left);
int xMax = Math.min(wTarget, searchRect.x + searchRect.width + right);
int yMin = Math.max(0, searchRect.y - top);
int yMax = Math.min(hTarget, searchRect.y + searchRect.height + bottom);
wTest = xMax - xMin;
hTest = yMax - yMin;
if (target.getType() != BufferedImage.TYPE_INT_RGB) {
BufferedImage image = new BufferedImage(wTarget, hTarget, BufferedImage.TYPE_INT_RGB);
image.createGraphics().drawImage(target, 0, 0, null);
target = image;
}
targetPixels = new int[wTest * hTest];
target.getRaster().getDataElements(xMin, yMin, wTest, hTest, targetPixels);
// get the points to search along the line
ArrayList<Point2D> searchPts = getSearchPoints(searchRect, x0, y0, theta);
if (searchPts == null) {
peakHeight = Double.NaN;
peakWidth = Double.NaN;
return null;
}
// collect differences in a map as they are measured
HashMap<Point2D, Double> diffs = new HashMap<Point2D, Double>();
// find the point with the minimum difference from template
double matchDiff = largeNumber; // larger than typical differences
int xMatch = 0, yMatch = 0;
double avgDiff = 0;
Point2D matchPt = null;
for (Point2D pt : searchPts) {
int x = (int) pt.getX();
int y = (int) pt.getY();
double diff = getDifferenceAtTestPoint(x, y);
diffs.put(pt, diff);
avgDiff += diff;
if (diff < matchDiff) {
matchDiff = diff;
xMatch = x;
yMatch = y;
matchPt = pt;
}
}
avgDiff /= searchPts.size();
peakHeight = avgDiff / matchDiff - 1;
peakWidth = Double.NaN;
double dl = 0;
int matchIndex = searchPts.indexOf(matchPt);
// if match is not exact, fit a Gaussian and find peak
if (!Double.isInfinite(peakHeight) && matchIndex > 0 && matchIndex < searchPts.size() - 1) {
// fill data arrays
Point2D pt = searchPts.get(matchIndex - 1);
double diff = diffs.get(pt);
xValues[0] = -pt.distance(matchPt);
yValues[0] = avgDiff / diff - 1;
xValues[1] = 0;
yValues[1] = peakHeight;
pt = searchPts.get(matchIndex + 1);
diff = diffs.get(pt);
xValues[2] = pt.distance(matchPt);
yValues[2] = avgDiff / diff - 1;
// determine approximate offset (dl) and width (w) values
double pull = -xValues[0] / (yValues[1] - yValues[0]);
double push = xValues[2] / (yValues[1] - yValues[2]);
if (Double.isNaN(pull)) pull = LARGE_NUMBER;
if (Double.isNaN(push)) push = LARGE_NUMBER;
dl = 0.3 * (xValues[2] - xValues[0]) * (push - pull) / (push + pull);
double ratio = dl > 0 ? peakHeight / yValues[0] : peakHeight / yValues[2];
double w = dl > 0 ? dl - xValues[0] : dl - xValues[2];
w = w * w / Math.log(ratio);
// set parameters and fit to x data
dataset.clear();
dataset.append(xValues, yValues);
double rmsDev = 1;
for (int k = 0; k < 3; k++) {
double c = k == 0 ? w : k == 1 ? w / 3 : w * 3;
f.setParameterValue(0, peakHeight);
f.setParameterValue(1, dl);
f.setParameterValue(2, c);
rmsDev = fitter.fit(f);
if (rmsDev < 0.01) { // fitter succeeded (3-point fit should be exact)
dl = f.getParameterValue(1);
peakWidth = f.getParameterValue(2);
break;
}
}
}
double dx = dl * Math.cos(theta);
double dy = dl * Math.sin(theta);
double xImage = xMatch + searchRect.x - left - trimLeft + dx;
double yImage = yMatch + searchRect.y - top - trimTop + dy;
return new TPoint(xImage, yImage);
}
/**
* Rebuilds the template from an input image. The input image dimensions must match the original.
* The input and original are overlaid onto the existing template, if any. Pixels that fall
* outside the mask are ignored in the final template.
*
* @param image the input image
* @param alphaInput the opacity with which the input image is overlaid
* @param alphaOriginal the opacity with which the original image is overlaid
*/
public void rebuildTemplate(BufferedImage image, int alphaInput, int alphaOriginal) {
int w = image.getWidth();
int h = image.getHeight();
// return if image dimensions do not match original image
if (original.getWidth() != w || original.getHeight() != h) return;
// return if both alphas are zero
if (alphaInput == 0 && alphaOriginal == 0) return;
// draw image onto argb input
BufferedImage input = new BufferedImage(w, h, BufferedImage.TYPE_INT_ARGB);
input.createGraphics().drawImage(image, 0, 0, null);
// create working image if needed
if (working == null) {
working = new BufferedImage(w, h, BufferedImage.TYPE_INT_ARGB);
}
// reset template dimensions and create new template if needed
if (template == null || w != wTemplate || h != hTemplate) {
wTemplate = w;
hTemplate = h;
int len = w * h;
template = new BufferedImage(w, h, BufferedImage.TYPE_INT_ARGB);
pixels = new int[len];
templateR = new int[len];
templateG = new int[len];
templateB = new int[len];
isPixelTransparent = new boolean[len];
}
// set alpha of input and draw onto working
Graphics2D gWorking = working.createGraphics();
alphaInput = Math.max(0, Math.min(255, alphaInput));
if (alphaInput > 0) { // overlay only if not transparent
gWorking.setComposite(getComposite(alphaInput));
input.getRaster().getDataElements(0, 0, w, h, pixels);
gWorking.drawImage(input, 0, 0, null);
}
// set alpha of original and draw onto working
alphaOriginal = Math.max(0, Math.min(255, alphaOriginal));
if (alphaOriginal > 0) { // overlay only if not transparent
gWorking.setComposite(getComposite(alphaOriginal));
original.getRaster().getDataElements(0, 0, w, h, pixels);
gWorking.drawImage(original, 0, 0, null);
}
// read pixels from working raster
working.getRaster().getDataElements(0, 0, wTemplate, hTemplate, pixels);
if (mask != null) {
// set pixels outside mask to transparent
for (int i = 0; i < pixels.length; i++) {
boolean inside = true;
// pixel is inside only if all corners are inside
int x = i % wTemplate, y = i / wTemplate;
for (int j = 0; j < 2; j++) {
for (int k = 0; k < 2; k++) {
p.setLocation(x + j, y + k);
inside = inside && mask.contains(p);
}
}
if (!inside) pixels[i] = pixels[i] & (0 << 24); // set alpha to zero (transparent)
}
}
// write pixels to template raster
template.getRaster().setDataElements(0, 0, wTemplate, hTemplate, pixels);
// trim transparent edges from template
int trimRight = 0, trimBottom = 0;
trimLeft = trimTop = 0;
// left edge
boolean transparentEdge = true;
while (transparentEdge && trimLeft < wTemplate) {
for (int line = 0; line < hTemplate; line++) {
int i = line * wTemplate + trimLeft;
transparentEdge = transparentEdge && getAlpha(pixels[i]) == 0;
}
if (transparentEdge) trimLeft++;
}
// right edge
transparentEdge = true;
while (transparentEdge && (trimLeft + trimRight) < wTemplate) {
for (int line = 0; line < hTemplate; line++) {
int i = (line + 1) * wTemplate - 1 - trimRight;
transparentEdge = transparentEdge && getAlpha(pixels[i]) == 0;
}
if (transparentEdge) trimRight++;
}
// top edge
transparentEdge = true;
while (transparentEdge && trimTop < hTemplate) {
for (int col = 0; col < wTemplate; col++) {
int i = trimTop * wTemplate + col;
transparentEdge = transparentEdge && getAlpha(pixels[i]) == 0;
}
if (transparentEdge) trimTop++;
}
// bottom edge
transparentEdge = true;
while (transparentEdge && (trimTop + trimBottom) < hTemplate) {
for (int col = 0; col < wTemplate; col++) {
int i = (hTemplate - 1 - trimBottom) * wTemplate + col;
transparentEdge = transparentEdge && getAlpha(pixels[i]) == 0;
}
if (transparentEdge) trimBottom++;
}
// reduce size of template if needed
if (trimLeft + trimRight + trimTop + trimBottom > 0) {
wTemplate -= (trimLeft + trimRight);
hTemplate -= (trimTop + trimBottom);
pixels = new int[wTemplate * hTemplate];
templateR = new int[wTemplate * hTemplate];
templateG = new int[wTemplate * hTemplate];
templateB = new int[wTemplate * hTemplate];
isPixelTransparent = new boolean[wTemplate * hTemplate];
BufferedImage bi = new BufferedImage(wTemplate, hTemplate, BufferedImage.TYPE_INT_ARGB);
bi.createGraphics().drawImage(template, -trimLeft, -trimTop, null);
template = bi;
template.getRaster().getDataElements(0, 0, wTemplate, hTemplate, pixels);
}
// set up rgb and transparency arrays for faster matching
for (int i = 0; i < pixels.length; i++) {
int val = pixels[i];
templateR[i] = getRed(val); // red
templateG[i] = getGreen(val); // green
templateB[i] = getBlue(val); // blue
isPixelTransparent[i] = getAlpha(val) == 0; // alpha
}
}
