| 1 |
/* ============================================================ |
| 2 |
* |
| 3 |
* This file is a part of digiKam project |
| 4 |
* https://www.digikam.org |
| 5 |
* |
| 6 |
* Date : 2024-11-10 |
| 7 |
* Description : Integrated, multithread face detection / recognition |
| 8 |
* |
| 9 |
* SPDX-FileCopyrightText: 2024-2026 by Gilles Caulier <caulier dot gilles at gmail dot com> |
| 10 |
* SPDX-FileCopyrightText: 2024-2025 by Michael Miller <michael underscore miller at msn dot com> |
| 11 |
* |
| 12 |
* SPDX-License-Identifier: GPL-2.0-or-later |
| 13 |
* |
| 14 |
* ============================================================ */ |
| 15 |
|
| 16 |
#include "facepipelinebase.h" |
| 17 |
|
| 18 |
// Qt includes |
| 19 |
|
| 20 |
#include <QList> |
| 21 |
|
| 22 |
// KDE includes |
| 23 |
|
| 24 |
#include <klocalizedstring.h> |
| 25 |
|
| 26 |
// Local includes |
| 27 |
|
| 28 |
#include "digikam_debug.h" |
| 29 |
#include "digikam_opencv.h" |
| 30 |
#include "sharedqueue.h" |
| 31 |
#include "facepipelinepackagebase.h" |
| 32 |
#include "thumbnailloadthread.h" |
| 33 |
#include "dnnsfaceextractor.h" |
| 34 |
#include "dnnmodelmanager.h" |
| 35 |
#include "identityprovider.h" |
| 36 |
|
| 37 |
namespace Digikam |
| 38 |
{ |
| 39 |
|
| 40 |
FacePipelineBase::FacePipelineBase(const FaceScanSettings& _settings) |
| 41 |
: MLPipelineFoundation(), |
| 42 |
settings (_settings) |
| 43 |
{ |
| 44 |
} |
| 45 |
|
| 46 |
FacePipelineBase::~FacePipelineBase() |
| 47 |
{ |
| 48 |
} |
| 49 |
|
| 50 |
double FacePipelineBase::detectNoise1(const cv::Mat& cvGrayImage) const |
| 51 |
{ |
| 52 |
// Use a Gaussian filter to detect noise |
| 53 |
|
| 54 |
// apply Gaussian blur to the image |
| 55 |
|
| 56 |
cv::Mat blurred; |
| 57 |
cv::GaussianBlur(cvGrayImage, blurred, cv::Size(5, 5), 0); |
| 58 |
|
| 59 |
// calculate the difference between the original and blurred image |
| 60 |
|
| 61 |
cv::Mat noise = cvGrayImage - blurred; |
| 62 |
|
| 63 |
// calculate the standard deviation of the noise |
| 64 |
|
| 65 |
cv::Mat mean, stddev; |
| 66 |
cv::meanStdDev(noise, mean, stddev); |
| 67 |
|
| 68 |
double noiseLevel = stddev.at<double>(0, 0); |
| 69 |
|
| 70 |
return noiseLevel; |
| 71 |
} |
| 72 |
|
| 73 |
double FacePipelineBase::detectNoise2(const cv::Mat& cvGrayImage) const |
| 74 |
{ |
| 75 |
int H = cvGrayImage.rows; |
| 76 |
int W = cvGrayImage.cols; |
| 77 |
cv::Mat M = (cv::Mat_<double>(3, 3) << 1, -2, 1, |
| 78 |
-2, 4, -2, |
| 79 |
1, -2, 1); |
| 80 |
|
| 81 |
cv::Mat convolved; |
| 82 |
cv::filter2D(cvGrayImage, convolved, CV_64F, M); |
| 83 |
|
| 84 |
double sigma = cv::sum(cv::abs(convolved))[0]; |
| 85 |
sigma = sigma * std::sqrt(0.5 * M_PI) / (6 * (W - 2) * (H - 2)); |
| 86 |
|
| 87 |
return sigma; |
| 88 |
} |
| 89 |
|
| 90 |
double FacePipelineBase::detectBlur(const cv::Mat& cvGrayImage) const |
| 91 |
{ |
| 92 |
// Use a Fast Fourier Transform to detect blurriness |
| 93 |
|
| 94 |
int cx = cvGrayImage.cols / 2; |
| 95 |
int cy = cvGrayImage.rows / 2; |
| 96 |
|
| 97 |
// Convert the image to a flat float |
| 98 |
|
| 99 |
cv::Mat fImage; |
| 100 |
cvGrayImage.convertTo(fImage, CV_32F); |
| 101 |
|
| 102 |
// FFT |
| 103 |
|
| 104 |
cv::Mat fourierTransform; |
| 105 |
cv::dft(fImage, fourierTransform, cv::DFT_SCALE | cv::DFT_COMPLEX_OUTPUT); |
| 106 |
|
| 107 |
// center low frequencies in the middle |
| 108 |
// by shuffling the quadrants. |
| 109 |
|
| 110 |
cv::Mat q0(fourierTransform, cv::Rect(0, 0, cx, cy)); // Top-Left - Create a ROI per quadrant |
| 111 |
cv::Mat q1(fourierTransform, cv::Rect(cx, 0, cx, cy)); // Top-Right |
| 112 |
cv::Mat q2(fourierTransform, cv::Rect(0, cy, cx, cy)); // Bottom-Left |
| 113 |
cv::Mat q3(fourierTransform, cv::Rect(cx, cy, cx, cy)); // Bottom-Right |
| 114 |
|
| 115 |
// swap quadrants (Top-Left with Bottom-Right) |
| 116 |
|
| 117 |
cv::Mat tmp; |
| 118 |
q0.copyTo(tmp); |
| 119 |
q3.copyTo(q0); |
| 120 |
tmp.copyTo(q3); |
| 121 |
|
| 122 |
// swap quadrant (Top-Right with Bottom-Left) |
| 123 |
|
| 124 |
q1.copyTo(tmp); |
| 125 |
q2.copyTo(q1); |
| 126 |
tmp.copyTo(q2); |
| 127 |
|
| 128 |
// Block the low frequencies |
| 129 |
// #define BLOCK could also be a argument on the command line of course |
| 130 |
|
| 131 |
fourierTransform(cv::Rect(cx - BLOCK, cy - BLOCK, 2 * BLOCK, 2 * BLOCK)).setTo(0); |
| 132 |
|
| 133 |
//shuffle the quadrants to their original position |
| 134 |
|
| 135 |
cv::Mat orgFFT; |
| 136 |
fourierTransform.copyTo(orgFFT); |
| 137 |
cv::Mat p0(orgFFT, cv::Rect(0, 0, cx, cy)); // Top-Left - Create a ROI per quadrant |
| 138 |
cv::Mat p1(orgFFT, cv::Rect(cx, 0, cx, cy)); // Top-Right |
| 139 |
cv::Mat p2(orgFFT, cv::Rect(0, cy, cx, cy)); // Bottom-Left |
| 140 |
cv::Mat p3(orgFFT, cv::Rect(cx, cy, cx, cy)); // Bottom-Right |
| 141 |
|
| 142 |
// swap quadrant (Top-Left with Bottom-Right) |
| 143 |
|
| 144 |
p0.copyTo(tmp); |
| 145 |
p3.copyTo(p0); |
| 146 |
tmp.copyTo(p3); |
| 147 |
|
| 148 |
// swap quadrant (Top-Right with Bottom-Left) |
| 149 |
|
| 150 |
p1.copyTo(tmp); |
| 151 |
p2.copyTo(p1); |
| 152 |
tmp.copyTo(p2); |
| 153 |
|
| 154 |
// IFFT |
| 155 |
|
| 156 |
cv::Mat invFFT; |
| 157 |
cv::Mat logFFT; |
| 158 |
double minVal = 0.0; |
| 159 |
double maxVal = 0.0; |
| 160 |
|
| 161 |
cv::dft(orgFFT, invFFT, cv::DFT_INVERSE | cv::DFT_REAL_OUTPUT); |
| 162 |
|
| 163 |
invFFT = cv::abs(invFFT); |
| 164 |
cv::minMaxLoc(invFFT, &minVal, &maxVal); |
| 165 |
|
| 166 |
// check for impossible values |
| 167 |
|
| 168 |
if (maxVal <= 0.0) |
| 169 |
{ |
| 170 |
return 1; |
| 171 |
} |
| 172 |
|
| 173 |
cv::log(invFFT, logFFT); |
| 174 |
logFFT *= 20; |
| 175 |
|
| 176 |
cv::Scalar result = cv::mean(logFFT); |
| 177 |
|
| 178 |
return result.val[0]; |
| 179 |
} |
| 180 |
|
| 181 |
bool FacePipelineBase::useForTraining(const QSize& faceSize, const cv::Mat& cvImage) |
| 182 |
{ |
| 183 |
if (!detectorModel) |
| 184 |
{ |
| 185 |
detectorModel = DNNModelManager::instance()->getModel(QStringLiteral("yunet"), |
| 186 |
DNNModelUsage::DNNUsageFaceDetection); |
| 187 |
} |
| 188 |
|
| 189 |
// thumbnail must be at least minThumbnailSize of the size the detector expects |
| 190 |
|
| 191 |
if ( |
| 192 |
((detectorModel->info.imageSize * minThumbnailSize) > faceSize.width()) || |
| 193 |
((detectorModel->info.imageSize * minThumbnailSize) > faceSize.height()) |
| 194 |
) |
| 195 |
{ |
| 196 |
qCDebug(DIGIKAM_FACESENGINE_LOG) << "FacePipelineBase::useForTraining Image is too small for face detection" |
| 197 |
<< faceSize.width() << "x" << faceSize.height() |
| 198 |
<< "required:" << detectorModel->info.imageSize * minThumbnailSize; |
| 199 |
return false; |
| 200 |
} |
| 201 |
|
| 202 |
// convert to grayscale for use in noise and blur detection |
| 203 |
|
| 204 |
cv::Mat cvGrayImage; |
| 205 |
cv::cvtColor(cvImage, cvGrayImage, cv::COLOR_RGB2GRAY); |
| 206 |
|
| 207 |
// use a Gaussian filter to check for noisy images |
| 208 |
|
| 209 |
double noise1 = detectNoise1(cvGrayImage); |
| 210 |
double noise2 = detectNoise2(cvGrayImage); |
| 211 |
|
| 212 |
// check if image is too noisy |
| 213 |
|
| 214 |
if ((noise1 > noiseThreshold1) || (noise2 > noiseThreshold2)) |
| 215 |
{ |
| 216 |
return false; |
| 217 |
} |
| 218 |
|
| 219 |
// use a FFT filter to check for blurred images |
| 220 |
|
| 221 |
double blur = detectBlur(cvGrayImage); |
| 222 |
|
| 223 |
// check if image is too blurry |
| 224 |
|
| 225 |
if (blur < blurThreshold) |
| 226 |
{ |
| 227 |
return false; |
| 228 |
} |
| 229 |
|
| 230 |
return true; |
| 231 |
} |
| 232 |
|
| 233 |
bool FacePipelineBase::commonFaceThumbnailLoader(const QString& pipelineName, |
| 234 |
QThread::Priority stagePriority, |
| 235 |
MLPipelineStage thisStage, |
| 236 |
MLPipelineStage nextStage) |
| 237 |
{ |
| 238 |
MLPIPELINE_STAGE_START(stagePriority, thisStage, nextStage); |
| 239 |
FacePipelinePackageBase* package = nullptr; |
| 240 |
|
| 241 |
/* ========================================================================================= |
| 242 |
* Pipeline stage specific initialization code |
| 243 |
* |
| 244 |
* Use the block from here to MLPIPELINE_LOOP_START to initialize the stage. |
| 245 |
* The code in this block is run once per stage initialization. The number of instances |
| 246 |
* is at least 1. More instances are created by addMoreWorkers if needed. |
| 247 |
*/ |
| 248 |
|
| 249 |
ThumbnailLoadThread* const thumbnailLoadThread = new ThumbnailLoadThread; |
| 250 |
/* |
| 251 |
ThumbnailLoadThread* thumbnailLoadThread = ThumbnailLoadThread::defaultThread(); |
| 252 |
*/ |
| 253 |
thumbnailLoadThread->setPixmapRequested(false); |
| 254 |
thumbnailLoadThread->setThumbnailSize(ThumbnailLoadThread::maximumThumbnailSize()); |
| 255 |
thumbnailLoadThread->setPriority(QThread::NormalPriority); |
| 256 |
|
| 257 |
ThumbnailImageCatcher* const catcher = new ThumbnailImageCatcher(thumbnailLoadThread); |
| 258 |
|
| 259 |
catcher->setActive(true); |
| 260 |
|
| 261 |
MLPIPELINE_LOOP_START(thisStage, thisQueue); |
| 262 |
package = static_cast<FacePipelinePackageBase*>(mlpackage); |
| 263 |
|
| 264 |
/* ========================================================================================= |
| 265 |
* Start pipeline stage specific loop |
| 266 |
* |
| 267 |
* All code from here to MLPIPELINE_LOOP_END is in a try/catch block and loop. |
| 268 |
* This loop is run once per image. |
| 269 |
*/ |
| 270 |
|
| 271 |
{ |
| 272 |
catcher->thread()->find(ItemInfo::thumbnailIdentifier(package->face.imageId()), |
| 273 |
package->face.region().toRect()); |
| 274 |
catcher->enqueue(); |
| 275 |
QList<QImage> images = catcher->waitForThumbnails(); |
| 276 |
|
| 277 |
if (images.size() && !images[0].isNull()) |
| 278 |
{ |
| 279 |
// store the thumbnail image in the package |
| 280 |
|
| 281 |
package->thumbnail = images[0]; |
| 282 |
|
| 283 |
// create a thumbnail icon |
| 284 |
|
| 285 |
package->thumbnailIcon = QIcon(DImg(package->thumbnail).smoothScale(48, |
| 286 |
48, |
| 287 |
Qt::KeepAspectRatio).convertToPixmap()); |
| 288 |
|
| 289 |
// send the package to the next stage |
| 290 |
|
| 291 |
enqueue(nextQueue, package); |
| 292 |
|
| 293 |
package = nullptr; |
| 294 |
} |
| 295 |
else |
| 296 |
{ |
| 297 |
// send a notification that the file was skipped |
| 298 |
|
| 299 |
notify(MLPipelineNotification::notifySkipped, |
| 300 |
package->info.name(), |
| 301 |
package->info.filePath(), |
| 302 |
QString(), |
| 303 |
0, |
| 304 |
QIcon::fromTheme(QLatin1String("image-missing"))); |
| 305 |
|
| 306 |
// delete the package since it is not needed |
| 307 |
|
| 308 |
delete package; |
| 309 |
|
| 310 |
package = nullptr; |
| 311 |
} |
| 312 |
} |
| 313 |
|
| 314 |
/* ========================================================================================= |
| 315 |
* End pipeline stage specific loop |
| 316 |
*/ |
| 317 |
|
| 318 |
MLPIPELINE_LOOP_END(thisStage, QString(pipelineName + QLatin1String("::commonFaceThumbnailLoader")).toLocal8Bit().data()); |
| 319 |
|
| 320 |
/* ========================================================================================= |
| 321 |
* Pipeline stage specific cleanup |
| 322 |
* |
| 323 |
* Use the block from here to MLPIPELINE_STAGE_END to clean up any resources used by the stage. |
| 324 |
*/ |
| 325 |
|
| 326 |
catcher->setActive(false); |
| 327 |
catcher->thread()->stopAllTasks(); |
| 328 |
catcher->cancel(); |
| 329 |
|
| 330 |
delete catcher->thread(); |
| 331 |
delete catcher; |
| 332 |
|
| 333 |
MLPIPELINE_STAGE_END(thisStage, nextStage); |
| 334 |
} |
| 335 |
|
| 336 |
bool FacePipelineBase::commonFaceThumbnailExtractor(const QString& pipelineName, |
| 337 |
QThread::Priority stagePriority, |
| 338 |
MLPipelineStage thisStage, |
| 339 |
MLPipelineStage nextStage, |
| 340 |
bool trainingQualityCheck) |
| 341 |
{ |
| 342 |
MLPIPELINE_STAGE_START(stagePriority, thisStage, nextStage); |
| 343 |
FacePipelinePackageBase* package = nullptr; |
| 344 |
|
| 345 |
/* ========================================================================================= |
| 346 |
* Pipeline stage specific initialization code |
| 347 |
* |
| 348 |
* Use the block from here to MLPIPELINE_LOOP_START to initialize the stage. |
| 349 |
* The code in this block is run once per stage initialization. The number of instances |
| 350 |
* is at least 1. More instances are created by addMoreWorkers if needed. |
| 351 |
*/ |
| 352 |
|
| 353 |
DNNSFaceExtractor extractor; |
| 354 |
FaceUtils utils; |
| 355 |
|
| 356 |
MLPIPELINE_LOOP_START(thisStage, thisQueue); |
| 357 |
package = static_cast<FacePipelinePackageBase*>(mlpackage); |
| 358 |
|
| 359 |
/* ========================================================================================= |
| 360 |
* Start pipeline stage specific loop |
| 361 |
* |
| 362 |
* All code from here to MLPIPELINE_LOOP_END is in a try/catch block and loop. |
| 363 |
* This loop is run once per image. |
| 364 |
*/ |
| 365 |
|
| 366 |
{ |
| 367 |
QImage inputImage(package->thumbnail.copy()); |
| 368 |
|
| 369 |
// preprocess image to be in the correct format |
| 370 |
|
| 371 |
if (inputImage.format() != QImage::Format_BGR888) |
| 372 |
{ |
| 373 |
inputImage = inputImage.convertToFormat(QImage::Format_BGR888); |
| 374 |
} |
| 375 |
|
| 376 |
// create a cv::Mat image from the QImage |
| 377 |
|
| 378 |
cv::Mat cvImage = cv::Mat( |
| 379 |
inputImage.height(), |
| 380 |
inputImage.width(), |
| 381 |
CV_8UC3, |
| 382 |
inputImage.scanLine(0), |
| 383 |
inputImage.bytesPerLine() |
| 384 |
); |
| 385 |
|
| 386 |
if (!cvImage.empty()) |
| 387 |
{ |
| 388 |
cv::UMat cvUImage = cvImage.getUMat(cv::ACCESS_READ); |
| 389 |
|
| 390 |
// Extract the face features |
| 391 |
|
| 392 |
QPair<cv::Mat, cv::Mat> faceEmbedding = extractor.getFaceEmbedding(cvUImage); |
| 393 |
package->features = faceEmbedding.second; |
| 394 |
|
| 395 |
// Check for a valid feature set |
| 396 |
|
| 397 |
if (!package->features.empty() && trainingQualityCheck) |
| 398 |
{ |
| 399 |
/** |
| 400 |
* Compute the size of the face region in pixels as shown in the original image |
| 401 |
* This is used to check if the image is suitable for training. |
| 402 |
*/ |
| 403 |
|
| 404 |
// Prevent divide by zero - bug 513760. |
| 405 |
|
| 406 |
float width = static_cast<float>(package->info.dimensions().width()); |
| 407 |
float height = static_cast<float>(package->info.dimensions().height()); |
| 408 |
float thumbWidth = static_cast<float>(package->thumbnail.width()); |
| 409 |
float thumbHeight = static_cast<float>(package->thumbnail.height()); |
| 410 |
|
| 411 |
if ((width <= 0.0) || (height <= 0.0) || (thumbWidth <= 0.0) || (thumbHeight <= 0.0)) |
| 412 |
{ |
| 413 |
qCWarning(DIGIKAM_FACESENGINE_LOG) << "Invalid dimensions or thumbnail size, skipping training check."; |
| 414 |
package->useForTraining = false; |
| 415 |
} |
| 416 |
else |
| 417 |
{ |
| 418 |
// Percentage of the region compared to the image |
| 419 |
|
| 420 |
float xThumbnailPercent = static_cast<float>(package->face.region().toRect().width()) / width; |
| 421 |
float yThumbnailPercent = static_cast<float>(package->face.region().toRect().height()) / height; |
| 422 |
|
| 423 |
// Prevent infinite or not a number values - bug 513760. |
| 424 |
|
| 425 |
float xFacePos = faceEmbedding.first.at<float>(0, 2); |
| 426 |
float yFacePos = faceEmbedding.first.at<float>(0, 3); |
| 427 |
|
| 428 |
if ( |
| 429 |
std::isnan(xFacePos) || std::isnan(yFacePos) || |
| 430 |
std::isinf(xFacePos) || std::isinf(yFacePos) |
| 431 |
) |
| 432 |
{ |
| 433 |
qCWarning(DIGIKAM_FACESENGINE_LOG) << "Invalid face embedding values, skipping training check."; |
| 434 |
package->useForTraining = false; |
| 435 |
} |
| 436 |
else |
| 437 |
{ |
| 438 |
// Percentage of the face compared to the thumbnail |
| 439 |
|
| 440 |
float xFaceAreaPercent = xFacePos / thumbWidth; |
| 441 |
float yFaceAreaPercent = yFacePos / thumbHeight; |
| 442 |
|
| 443 |
// Compute the size of the face area in pixels |
| 444 |
|
| 445 |
QSize faceSize(qRound(package->info.dimensions().width() * xThumbnailPercent * xFaceAreaPercent), |
| 446 |
qRound(package->info.dimensions().height() * yThumbnailPercent * yFaceAreaPercent)); |
| 447 |
|
| 448 |
// Check if the image is suitable for training |
| 449 |
|
| 450 |
package->useForTraining = useForTraining(faceSize, cvImage); |
| 451 |
} |
| 452 |
} |
| 453 |
} |
| 454 |
|
| 455 |
for (const auto tagId : package->face.rejectedFaceTagList()) |
| 456 |
{ |
| 457 |
// add the Identity ID for the rejected face tag to the exclusion list |
| 458 |
|
| 459 |
QMultiMap<QString, QString> attributes = FaceTags::identityAttributes(tagId); |
| 460 |
Identity identity = IdentityProvider::instance()->findIdentity(attributes); |
| 461 |
|
| 462 |
if (!identity.isNull()) |
| 463 |
{ |
| 464 |
package->exclusionIdentityIds << identity.id(); |
| 465 |
} |
| 466 |
} |
| 467 |
|
| 468 |
// send the package to the next stage |
| 469 |
|
| 470 |
enqueue(nextQueue, package); |
| 471 |
|
| 472 |
package = nullptr; |
| 473 |
} |
| 474 |
else |
| 475 |
{ |
| 476 |
// send a notification that the file was skipped |
| 477 |
|
| 478 |
notify(MLPipelineNotification::notifySkipped, |
| 479 |
package->info.name(), |
| 480 |
package->info.filePath(), |
| 481 |
QString(), |
| 482 |
0, |
| 483 |
package->thumbnailIcon); |
| 484 |
|
| 485 |
// delete the package since it is not needed |
| 486 |
|
| 487 |
delete package; |
| 488 |
|
| 489 |
package = nullptr; |
| 490 |
} |
| 491 |
} |
| 492 |
|
| 493 |
/* ========================================================================================= |
| 494 |
* End pipeline stage specific loop |
| 495 |
*/ |
| 496 |
|
| 497 |
MLPIPELINE_LOOP_END(thisStage, QString(pipelineName + QLatin1String("::commonFaceThumbnailExtractor")).toLocal8Bit().data()); |
| 498 |
|
| 499 |
/* ========================================================================================= |
| 500 |
* Pipeline stage specific cleanup |
| 501 |
* |
| 502 |
* Use the block from here to MLPIPELINE_STAGE_END to clean up any resources used by the stage. |
| 503 |
*/ |
| 504 |
|
| 505 |
MLPIPELINE_STAGE_END(thisStage, nextStage); |
| 506 |
} |
| 507 |
|
| 508 |
bool FacePipelineBase::enqueue(MLPipelineQueue* thisQueue, MLPipelinePackageFoundation* package) |
| 509 |
{ |
| 510 |
if (nullptr != package) |
| 511 |
{ |
| 512 |
// calculate the package size. Only big items need to be checked |
| 513 |
|
| 514 |
package->size = static_cast<FacePipelinePackageBase*>(package)->image.size().width() * |
| 515 |
static_cast<FacePipelinePackageBase*>(package)->image.size().height() * 4; |
| 516 |
} |
| 517 |
|
| 518 |
return MLPipelineFoundation::enqueue(thisQueue, package); |
| 519 |
} |
| 520 |
|
| 521 |
} // namespace Digikam |
| 522 |
|
| 523 |
#include "moc_facepipelinebase.cpp" |