| 1 |
/* ============================================================ |
| 2 |
* |
| 3 |
* This file is a part of digiKam project |
| 4 |
* https://www.digikam.org |
| 5 |
* |
| 6 |
* Date : 2024-11-10 |
| 7 |
* Description : Performs face detection and 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 "facepipelinedetectrecognize.h" |
| 17 |
|
| 18 |
// C++ includes |
| 19 |
|
| 20 |
#include <climits> |
| 21 |
#include <cmath> |
| 22 |
|
| 23 |
// Qt includes |
| 24 |
|
| 25 |
#include <QSet> |
| 26 |
#include <QList> |
| 27 |
#include <QRectF> |
| 28 |
#include <QElapsedTimer> |
| 29 |
|
| 30 |
// KDE includes |
| 31 |
|
| 32 |
#include <klocalizedstring.h> |
| 33 |
|
| 34 |
// Local includes |
| 35 |
|
| 36 |
#include "digikam_debug.h" |
| 37 |
#include "digikam_opencv.h" |
| 38 |
#include "sharedqueue.h" |
| 39 |
#include "collectionmanager.h" |
| 40 |
#include "album.h" |
| 41 |
#include "coredb.h" |
| 42 |
#include "iteminfo.h" |
| 43 |
#include "previewloadthread.h" |
| 44 |
#include "faceutils.h" |
| 45 |
#include "facescansettings.h" |
| 46 |
#include "facepipelinepackagebase.h" |
| 47 |
#include "thumbnailloadthread.h" |
| 48 |
#include "faceclassifier.h" |
| 49 |
#include "identityprovider.h" |
| 50 |
#include "qtopencvimg.h" |
| 51 |
#include "ocvocldnnsetter.h" |
| 52 |
|
| 53 |
namespace Digikam |
| 54 |
{ |
| 55 |
|
| 56 |
FacePipelineDetectRecognize::FacePipelineDetectRecognize(const FaceScanSettings& _settings) |
| 57 |
: FacePipelineBase(_settings) |
| 58 |
{ |
| 59 |
} |
| 60 |
|
| 61 |
FacePipelineDetectRecognize::~FacePipelineDetectRecognize() |
| 62 |
{ |
| 63 |
if (faceDetector) |
| 64 |
{ |
| 65 |
delete faceDetector; |
| 66 |
} |
| 67 |
|
| 68 |
// extractor is singleton, so no need to delete it |
| 69 |
} |
| 70 |
|
| 71 |
bool FacePipelineDetectRecognize::start() |
| 72 |
{ |
| 73 |
// create the face detector and extractor |
| 74 |
|
| 75 |
try |
| 76 |
{ |
| 77 |
faceDetector = new DNNFaceDetectorYuNet; |
| 78 |
faceExtractor = static_cast<DNNModelSFace*>(DNNModelManager::instance()->getModel(QStringLiteral("sface"), |
| 79 |
DNNModelUsage::DNNUsageFaceRecognition)); |
| 80 |
faceExtractor->getNet(); |
| 81 |
} |
| 82 |
|
| 83 |
catch (const std::exception& e) |
| 84 |
{ |
| 85 |
std::cerr << e.what() << '\n'; |
| 86 |
|
| 87 |
if (faceDetector) |
| 88 |
{ |
| 89 |
delete faceDetector; |
| 90 |
faceDetector = nullptr; |
| 91 |
} |
| 92 |
|
| 93 |
// extractor is singleton, so no need to delete it |
| 94 |
} |
| 95 |
|
| 96 |
catch (...) |
| 97 |
{ |
| 98 |
if (faceDetector) |
| 99 |
{ |
| 100 |
delete faceDetector; |
| 101 |
faceDetector = nullptr; |
| 102 |
} |
| 103 |
} |
| 104 |
|
| 105 |
// check if the detector and extractor were created |
| 106 |
|
| 107 |
if ((nullptr == faceDetector) || (nullptr == faceExtractor) || !faceExtractor->modelLoaded) |
| 108 |
{ |
| 109 |
return false; |
| 110 |
} |
| 111 |
|
| 112 |
// set the face detection size and accuracy |
| 113 |
|
| 114 |
faceDetector->setFaceDetectionSize(settings.detectSize); |
| 115 |
faceDetector->uiConfidenceThreshold = settings.detectAccuracy; |
| 116 |
|
| 117 |
{ |
| 118 |
// use the mutex to synchronize the start of the threads |
| 119 |
|
| 120 |
QMutexLocker lock(&mutex); |
| 121 |
|
| 122 |
// add the worker threads for this pipeline |
| 123 |
|
| 124 |
addWorker(MLPipelineStage::Finder); |
| 125 |
addWorker(MLPipelineStage::Loader); |
| 126 |
addWorker(MLPipelineStage::Extractor); |
| 127 |
addWorker(MLPipelineStage::Classifier); |
| 128 |
addWorker(MLPipelineStage::Writer); |
| 129 |
} |
| 130 |
|
| 131 |
return FacePipelineBase::start(); |
| 132 |
} |
| 133 |
|
| 134 |
bool FacePipelineDetectRecognize::finder() |
| 135 |
{ |
| 136 |
ActionThreadBase::setCurrentThreadName(QString::fromLatin1("FaceDetReco%1").arg(QLatin1String(__FUNCTION__))); // To customize thread name |
| 137 |
|
| 138 |
MLPIPELINE_FINDER_START(settings.workerThreadPriority, MLPipelineStage::Loader); |
| 139 |
|
| 140 |
/* ========================================================================================= |
| 141 |
* Pipeline finder specific initialization code |
| 142 |
* |
| 143 |
* Use the block from here to MLPIPELINE_FINDER_END to find the IDs images to process. |
| 144 |
* The code in this block is run once per stage initialization. The number of instances |
| 145 |
* is always 1. |
| 146 |
*/ |
| 147 |
|
| 148 |
// get the IDs to process |
| 149 |
|
| 150 |
QSet<qlonglong> filter; |
| 151 |
int serialNumber = 0; |
| 152 |
|
| 153 |
for (const Album* const album : std::as_const(settings.albums)) |
| 154 |
{ |
| 155 |
if (cancelled) |
| 156 |
{ |
| 157 |
break; |
| 158 |
} |
| 159 |
|
| 160 |
if (!album->isTrashAlbum()) |
| 161 |
{ |
| 162 |
QList<qlonglong> imageIds; |
| 163 |
|
| 164 |
// get the image IDs for the album |
| 165 |
|
| 166 |
if (album->type() == Album::PHYSICAL) |
| 167 |
{ |
| 168 |
imageIds = CoreDbAccess().db()->getImageIds(album->id(), |
| 169 |
DatabaseItem::Status::Visible, |
| 170 |
(FaceScanSettings::AlreadyScannedHandling::Skip != |
| 171 |
settings.alreadyScannedHandling)); |
| 172 |
} |
| 173 |
else if (album->type() == Album::TAG) |
| 174 |
{ |
| 175 |
imageIds = CoreDbAccess().db()->getItemIDsInTag(album->id()); |
| 176 |
} |
| 177 |
|
| 178 |
// iterate over the image IDs and add unique IDs to the queue for processing |
| 179 |
|
| 180 |
for (qlonglong imageId : std::as_const(imageIds)) |
| 181 |
{ |
| 182 |
// filter out duplicate image IDs |
| 183 |
|
| 184 |
if (!filter.contains(imageId)) |
| 185 |
{ |
| 186 |
// quick check if we should add threads. |
| 187 |
|
| 188 |
if (!moreCpu) |
| 189 |
{ |
| 190 |
moreCpu = checkMoreWorkers(totalItemCount, imageIds.size(), settings.useFullCpu); |
| 191 |
} |
| 192 |
|
| 193 |
++totalItemCount; |
| 194 |
filter << imageId; |
| 195 |
|
| 196 |
// add the image ID to the queue for processing |
| 197 |
|
| 198 |
enqueue(nextQueue, new FacePipelinePackageBase(imageId, ++serialNumber)); |
| 199 |
} |
| 200 |
} |
| 201 |
} |
| 202 |
} |
| 203 |
|
| 204 |
for (const ItemInfo& info : std::as_const(settings.infos)) |
| 205 |
{ |
| 206 |
// filter out duplicate image IDs |
| 207 |
|
| 208 |
qlonglong imageId = info.id(); |
| 209 |
|
| 210 |
if (!filter.contains(imageId)) |
| 211 |
{ |
| 212 |
// quick check if we should add threads. |
| 213 |
|
| 214 |
if (!moreCpu) |
| 215 |
{ |
| 216 |
moreCpu = checkMoreWorkers(totalItemCount, settings.infos.size(), settings.useFullCpu); |
| 217 |
} |
| 218 |
|
| 219 |
++totalItemCount; |
| 220 |
filter << imageId; |
| 221 |
|
| 222 |
// add the image ID to the queue for processing |
| 223 |
|
| 224 |
enqueue(nextQueue, new FacePipelinePackageBase(imageId, ++serialNumber)); |
| 225 |
} |
| 226 |
} |
| 227 |
|
| 228 |
/* ========================================================================================= |
| 229 |
* Pipeline finder specific cleanup |
| 230 |
* |
| 231 |
* Use the block from here to MLPIPELINE_FINDER_END to clean up any resources used by the stage. |
| 232 |
*/ |
| 233 |
|
| 234 |
MLPIPELINE_FINDER_END(MLPipelineStage::Loader); |
| 235 |
} |
| 236 |
|
| 237 |
bool FacePipelineDetectRecognize::loader() |
| 238 |
{ |
| 239 |
ActionThreadBase::setCurrentThreadName(QString::fromLatin1("FaceDetReco%1").arg(QLatin1String(__FUNCTION__))); // To customize thread name |
| 240 |
|
| 241 |
MLPIPELINE_STAGE_START(QThread::LowPriority, MLPipelineStage::Loader, MLPipelineStage::Extractor); |
| 242 |
FacePipelinePackageBase* package = nullptr; |
| 243 |
|
| 244 |
/* ========================================================================================= |
| 245 |
* Pipeline stage specific initialization code |
| 246 |
* |
| 247 |
* Use the block from here to MLPIPELINE_LOOP_START to initialize the stage. |
| 248 |
* The code in this block is run once per stage initialization. The number of instances |
| 249 |
* is at least 1. More instances are created by addMoreWorkers if needed. |
| 250 |
*/ |
| 251 |
|
| 252 |
MLPIPELINE_LOOP_START(MLPipelineStage::Loader, thisQueue); |
| 253 |
package = static_cast<FacePipelinePackageBase*>(mlpackage); |
| 254 |
|
| 255 |
/* ========================================================================================= |
| 256 |
* Start pipeline stage specific loop |
| 257 |
* |
| 258 |
* All code from here to MLPIPELINE_LOOP_END is in a try/catch block and loop. |
| 259 |
* This loop is run once per image. |
| 260 |
*/ |
| 261 |
{ |
| 262 |
// check if the ID is for an image (not video or other file type) |
| 263 |
|
| 264 |
bool sendNotification = true; |
| 265 |
|
| 266 |
if (DatabaseItem::Category::Image == package->info.category()) |
| 267 |
{ |
| 268 |
PreviewSettings::RawLoading rawLoadingMode = PreviewSettings::RawPreviewAutomatic; |
| 269 |
|
| 270 |
if (QStringLiteral("RAW-ARW") == package->info.format()) |
| 271 |
{ |
| 272 |
// fix for #447767: ARW preview is different aspect than the image |
| 273 |
|
| 274 |
bool loadFullSize = |
| 275 |
( |
| 276 |
qMin(package->info.dimensions().width() / 2, package->info.dimensions().height() / 2) < |
| 277 |
qMin(faceDetector->nnInputSizeRequired().height / 2, faceDetector->nnInputSizeRequired().width / 2) |
| 278 |
); |
| 279 |
|
| 280 |
rawLoadingMode = (loadFullSize ? PreviewSettings::RawPreviewFromRawFullSize |
| 281 |
: PreviewSettings::RawPreviewFromRawHalfSize); |
| 282 |
} |
| 283 |
|
| 284 |
// load high quality image for detection. |
| 285 |
|
| 286 |
package->image = PreviewLoadThread::loadHighQualitySynchronously(package->info.filePath(), rawLoadingMode); |
| 287 |
|
| 288 |
// check for corrupted images that cannot be loaded |
| 289 |
|
| 290 |
if (!package->image.isNull()) |
| 291 |
{ |
| 292 |
// create a thumbnail for the notification |
| 293 |
|
| 294 |
package->thumbnailIcon = QIcon(package->image.smoothScale(48, 48, Qt::KeepAspectRatio).convertToPixmap()); |
| 295 |
|
| 296 |
// send to the next stage |
| 297 |
|
| 298 |
enqueue(nextQueue, package); |
| 299 |
|
| 300 |
package = nullptr; |
| 301 |
|
| 302 |
sendNotification = false; |
| 303 |
} |
| 304 |
} |
| 305 |
|
| 306 |
if (sendNotification) |
| 307 |
{ |
| 308 |
// send a notification that the file was skipped |
| 309 |
|
| 310 |
notify(MLPipelineNotification::notifySkipped, |
| 311 |
package->info.name(), |
| 312 |
package->info.relativePath(), |
| 313 |
QString(), |
| 314 |
0, |
| 315 |
package->thumbnailIcon); |
| 316 |
|
| 317 |
// delete the package since it is not needed |
| 318 |
|
| 319 |
delete package; |
| 320 |
|
| 321 |
package = nullptr; |
| 322 |
} |
| 323 |
} |
| 324 |
|
| 325 |
/* ========================================================================================= |
| 326 |
* End pipeline stage specific loop |
| 327 |
*/ |
| 328 |
|
| 329 |
MLPIPELINE_LOOP_END(MLPipelineStage::Loader, "FacePipelineDetectRecognize::loader"); |
| 330 |
|
| 331 |
/* ========================================================================================= |
| 332 |
* Pipeline stage specific cleanup |
| 333 |
* |
| 334 |
* Use the block from here to MLPIPELINE_STAGE_END to clean up any resources used by the stage. |
| 335 |
*/ |
| 336 |
|
| 337 |
MLPIPELINE_STAGE_END(MLPipelineStage::Loader, MLPipelineStage::Extractor); |
| 338 |
} |
| 339 |
|
| 340 |
bool FacePipelineDetectRecognize::extractor() |
| 341 |
{ |
| 342 |
ActionThreadBase::setCurrentThreadName(QString::fromLatin1("FaceDetReco%1").arg(QLatin1String(__FUNCTION__))); // To customize thread name |
| 343 |
|
| 344 |
MLPIPELINE_STAGE_START(QThread::LowPriority, MLPipelineStage::Extractor, MLPipelineStage::Classifier); |
| 345 |
FacePipelinePackageBase* package = nullptr; |
| 346 |
|
| 347 |
/* ========================================================================================= |
| 348 |
* Pipeline stage specific initialization code |
| 349 |
* |
| 350 |
* Use the block from here to MLPIPELINE_LOOP_START to initialize the stage. |
| 351 |
* The code in this block is run once per stage initialization. The number of instances |
| 352 |
* is at least 1. More instances are created by addMoreWorkers if needed. |
| 353 |
*/ |
| 354 |
|
| 355 |
FaceUtils utils; |
| 356 |
|
| 357 |
MLPIPELINE_LOOP_START(MLPipelineStage::Extractor, thisQueue); |
| 358 |
package = static_cast<FacePipelinePackageBase*>(mlpackage); |
| 359 |
|
| 360 |
/* ========================================================================================= |
| 361 |
* Start pipeline stage specific loop |
| 362 |
* |
| 363 |
* All code from here to MLPIPELINE_LOOP_END is in a try/catch block and loop. |
| 364 |
* This loop is run once per image. |
| 365 |
*/ |
| 366 |
|
| 367 |
{ |
| 368 |
// preprocess the image |
| 369 |
|
| 370 |
// copy the image to a cv::UMat |
| 371 |
|
| 372 |
cv::UMat cvUResizedImage; |
| 373 |
cv::UMat cvUImage = QtOpenCVImg::image2Mat( |
| 374 |
package->image, |
| 375 |
CV_8UC3, |
| 376 |
QtOpenCVImg::MatColorOrder::MCO_BGR |
| 377 |
) |
| 378 |
.getUMat(cv::ACCESS_RW); |
| 379 |
|
| 380 |
// resize the image if needed. Only resize if the image is larger than the input size of the detector |
| 381 |
|
| 382 |
cv::Size inputImageSize = faceDetector->nnInputSizeRequired(); |
| 383 |
float resizeFactor = 1.0F; |
| 384 |
|
| 385 |
if (std::max(cvUImage.cols, cvUImage.rows) > std::max(inputImageSize.width, inputImageSize.height)) |
| 386 |
{ |
| 387 |
// Image should be resized. YuNet image sizes are much more flexible than SSD and YOLO |
| 388 |
// so we just need to make sure no one bound exceeds the max. No padding needed. |
| 389 |
|
| 390 |
resizeFactor = std::min(static_cast<float>(inputImageSize.width) / static_cast<float>(cvUImage.cols), |
| 391 |
static_cast<float>(inputImageSize.height) / static_cast<float>(cvUImage.rows)); |
| 392 |
|
| 393 |
int newWidth = (int)(resizeFactor * cvUImage.cols); |
| 394 |
int newHeight = (int)(resizeFactor * cvUImage.rows); |
| 395 |
cv::resize(cvUImage, cvUResizedImage, cv::Size(newWidth, newHeight)); |
| 396 |
} |
| 397 |
else |
| 398 |
{ |
| 399 |
cvUResizedImage = cvUImage; |
| 400 |
} |
| 401 |
|
| 402 |
// get reciprocal factor for resizing the face to back the original image size |
| 403 |
|
| 404 |
float reciprocalFactor = 1.0F / resizeFactor; |
| 405 |
|
| 406 |
// lock OpenCV OpenCL DNN settings while using UMat |
| 407 |
|
| 408 |
OpenCVOpenCLDNNSetter openCLDNNSetter;; |
| 409 |
|
| 410 |
// detect any faces in the image |
| 411 |
|
| 412 |
cv::UMat udetectionResults = faceDetector->callModel(cvUResizedImage); |
| 413 |
|
| 414 |
// process detected faces |
| 415 |
|
| 416 |
if (udetectionResults.rows > 0) |
| 417 |
{ |
| 418 |
cv::Mat detectionResults = udetectionResults.getMat(cv::ACCESS_READ); |
| 419 |
|
| 420 |
// get list of all previously detected faces |
| 421 |
|
| 422 |
QList<FaceTagsIface> faces = utils.databaseFaces(package->info.id()); |
| 423 |
|
| 424 |
QList<QRectF> faceFRects; |
| 425 |
|
| 426 |
// Loop through the faces found. |
| 427 |
|
| 428 |
for (int i = 0 ; i < detectionResults.rows ; ++i) |
| 429 |
{ |
| 430 |
// Create the rect of the face. |
| 431 |
|
| 432 |
const float x = detectionResults.at<float>(i, 0); |
| 433 |
const float y = detectionResults.at<float>(i, 1); |
| 434 |
const float width = detectionResults.at<float>(i, 2); |
| 435 |
const float height = detectionResults.at<float>(i, 3); |
| 436 |
|
| 437 |
if (!std::isfinite(x) || |
| 438 |
!std::isfinite(y) || |
| 439 |
!std::isfinite(width) || |
| 440 |
!std::isfinite(height) || |
| 441 |
(width <= 0.0F) || |
| 442 |
(height <= 0.0F)) |
| 443 |
{ |
| 444 |
continue; |
| 445 |
} |
| 446 |
|
| 447 |
const int X = static_cast<int>(x); |
| 448 |
const int Y = static_cast<int>(y); |
| 449 |
const int iwidth = static_cast<int>(width); |
| 450 |
const int iheight = static_cast<int>(height); |
| 451 |
|
| 452 |
if ( |
| 453 |
(X == INT_MIN) || |
| 454 |
(Y == INT_MIN) || |
| 455 |
(iwidth == INT_MIN) || |
| 456 |
(iheight == INT_MIN) |
| 457 |
) |
| 458 |
{ |
| 459 |
qCWarning(DIGIKAM_FACESENGINE_LOG) << "Invalid detection rectangle" |
| 460 |
<< "(" << X<< ", " << Y |
| 461 |
<< "[" << iwidth << ", " << iheight << "]" |
| 462 |
<< "for image" |
| 463 |
<< package->info.filePath(); |
| 464 |
|
| 465 |
continue; |
| 466 |
} |
| 467 |
|
| 468 |
// Add the rect to result list. |
| 469 |
|
| 470 |
QRectF currentFaceFRect(qreal(X) / qreal(cvUResizedImage.cols), |
| 471 |
qreal(Y) / qreal(cvUResizedImage.rows), |
| 472 |
qreal(iwidth) / qreal(cvUResizedImage.cols), |
| 473 |
qreal(iheight)/ qreal(cvUResizedImage.rows)); |
| 474 |
|
| 475 |
faceFRects << currentFaceFRect; |
| 476 |
|
| 477 |
// compute current image relative rect |
| 478 |
|
| 479 |
QRect rect = QRect(package->image.width() * currentFaceFRect.x(), |
| 480 |
package->image.height() * currentFaceFRect.y(), |
| 481 |
package->image.width() * currentFaceFRect.width(), |
| 482 |
package->image.height() * currentFaceFRect.height()); |
| 483 |
|
| 484 |
// check if rect is already assigned to a face to filter out confirmed and ignored faces |
| 485 |
|
| 486 |
bool found = false; |
| 487 |
FaceTagsIface face; |
| 488 |
|
| 489 |
if (faces.size() > 0) |
| 490 |
{ |
| 491 |
for (int j = 0; j < faces.size(); ++j) |
| 492 |
{ |
| 493 |
if (faces[j].region().intersects(TagRegion(rect), 0.5)) |
| 494 |
{ |
| 495 |
if (faces[j].isConfirmedName() || faces[j].isIgnoredName()) |
| 496 |
{ |
| 497 |
// face is already confirmed or ignored |
| 498 |
|
| 499 |
found = true; |
| 500 |
break; |
| 501 |
} |
| 502 |
else |
| 503 |
{ |
| 504 |
// face is not confirmed or ignored, but already detected |
| 505 |
|
| 506 |
face = faces[j]; |
| 507 |
break; |
| 508 |
} |
| 509 |
} |
| 510 |
} |
| 511 |
} |
| 512 |
|
| 513 |
// if face is not found (confirmed), then extract for classification |
| 514 |
|
| 515 |
if (!found) |
| 516 |
{ |
| 517 |
cv::UMat ualignedFace; |
| 518 |
cv::UMat uface_features; |
| 519 |
cv::Mat face_features; |
| 520 |
|
| 521 |
// extract the face vectors (features) for classification |
| 522 |
|
| 523 |
// get cvMat of the face landmarks |
| 524 |
|
| 525 |
cv::Mat row = detectionResults.row(i); |
| 526 |
|
| 527 |
// convert the face landmarks to the full size image |
| 528 |
|
| 529 |
for (int j = 0 ; j < row.cols ; ++j) |
| 530 |
{ |
| 531 |
row.at<float>(j) = (int)(row.at<float>(j) * reciprocalFactor); |
| 532 |
} |
| 533 |
|
| 534 |
// convert the face landmarks to a UMat |
| 535 |
|
| 536 |
cv::UMat urow = row.getUMat(cv::ACCESS_READ); |
| 537 |
|
| 538 |
// align and crop the face |
| 539 |
|
| 540 |
QMutexLocker lock(&(faceExtractor->mutex)); |
| 541 |
|
| 542 |
faceExtractor->getNet()->alignCrop(cvUImage, urow, ualignedFace); |
| 543 |
|
| 544 |
cv::UMat paddedFace; |
| 545 |
|
| 546 |
if (std::min(ualignedFace.cols, ualignedFace.rows) > 112) |
| 547 |
{ |
| 548 |
// Image should be resized. YuNet image sizes are much more flexible than SSD and YOLO. |
| 549 |
// So we just need to make sure no one bound exceeds the max. No padding needed. |
| 550 |
|
| 551 |
float resizeFactor2 = std::min(static_cast<float>(112) / static_cast<float>(ualignedFace.cols), |
| 552 |
static_cast<float>(112) / static_cast<float>(ualignedFace.rows)); |
| 553 |
|
| 554 |
int newWidth = (int)(resizeFactor2 * ualignedFace.cols); |
| 555 |
int newHeight = (int)(resizeFactor2 * ualignedFace.rows); |
| 556 |
cv::resize(ualignedFace, paddedFace, cv::Size(newWidth, newHeight)); |
| 557 |
} |
| 558 |
else |
| 559 |
{ |
| 560 |
paddedFace = ualignedFace.clone(); |
| 561 |
} |
| 562 |
|
| 563 |
ualignedFace.release(); |
| 564 |
|
| 565 |
// get the face features |
| 566 |
|
| 567 |
faceExtractor->getNet()->feature(paddedFace, uface_features); |
| 568 |
|
| 569 |
face_features = uface_features.getMat(cv::ACCESS_READ); |
| 570 |
|
| 571 |
// normalize the face features if we have any |
| 572 |
|
| 573 |
if (0 < face_features.rows) |
| 574 |
{ |
| 575 |
// normalize the face features |
| 576 |
|
| 577 |
cv::Mat normalized_features; |
| 578 |
normalize(face_features, normalized_features); |
| 579 |
|
| 580 |
// add the face features and face rect to the package |
| 581 |
|
| 582 |
package->featuresList << normalized_features; |
| 583 |
package->faceRects << currentFaceFRect; |
| 584 |
package->faceList << face; |
| 585 |
} |
| 586 |
} |
| 587 |
} |
| 588 |
} |
| 589 |
|
| 590 |
// send the package to the next stage |
| 591 |
|
| 592 |
enqueue(nextQueue, package); |
| 593 |
|
| 594 |
package = nullptr; |
| 595 |
} |
| 596 |
|
| 597 |
/* ========================================================================================= |
| 598 |
* End pipeline stage specific loop |
| 599 |
*/ |
| 600 |
|
| 601 |
MLPIPELINE_LOOP_END(MLPipelineStage::Extractor, "FacePipelineDetectRecognize::extractor"); |
| 602 |
|
| 603 |
/* ========================================================================================= |
| 604 |
* Pipeline stage specific cleanup |
| 605 |
* |
| 606 |
* Use the block from here to MLPIPELINE_STAGE_END to clean up any resources used by the stage. |
| 607 |
*/ |
| 608 |
|
| 609 |
MLPIPELINE_STAGE_END(MLPipelineStage::Extractor, MLPipelineStage::Classifier); |
| 610 |
} |
| 611 |
|
| 612 |
bool FacePipelineDetectRecognize::classifier() |
| 613 |
{ |
| 614 |
ActionThreadBase::setCurrentThreadName(QString::fromLatin1("FaceDetReco%1").arg(QLatin1String(__FUNCTION__))); // To customize thread name |
| 615 |
|
| 616 |
MLPIPELINE_STAGE_START(QThread::LowPriority, MLPipelineStage::Classifier, MLPipelineStage::Writer); |
| 617 |
FacePipelinePackageBase* package = nullptr; |
| 618 |
|
| 619 |
/* ========================================================================================= |
| 620 |
* Pipeline stage specific initialization code |
| 621 |
* |
| 622 |
* Use the block from here to MLPIPELINE_LOOP_START to initialize the stage. |
| 623 |
* The code in this block is run once per stage initialization. The number of instances |
| 624 |
* is at least 1. More instances are created by addMoreWorkers if needed. |
| 625 |
*/ |
| 626 |
|
| 627 |
FaceClassifier* const classifier = FaceClassifier::instance(); |
| 628 |
classifier->setParameters(settings); |
| 629 |
|
| 630 |
MLPIPELINE_LOOP_START(MLPipelineStage::Classifier, thisQueue); |
| 631 |
package = static_cast<FacePipelinePackageBase*>(mlpackage); |
| 632 |
|
| 633 |
/* ========================================================================================= |
| 634 |
* Start pipeline stage specific loop |
| 635 |
* |
| 636 |
* All code from here to MLPIPELINE_LOOP_END is in a try/catch block and loop. |
| 637 |
* This loop is run once per image. |
| 638 |
*/ |
| 639 |
|
| 640 |
{ |
| 641 |
for (int i = 0 ; i < package->featuresList.size() ; ++i) |
| 642 |
{ |
| 643 |
// verify the feature mat is not empty |
| 644 |
|
| 645 |
if (0 < package->featuresList[i].rows) |
| 646 |
{ |
| 647 |
QList<int> exclusionIdentityIds; |
| 648 |
|
| 649 |
for (const auto tagId : package->faceList[i].rejectedFaceTagList()) |
| 650 |
{ |
| 651 |
// add the Identity ID for the rejected face tag to the exclusion list |
| 652 |
|
| 653 |
QMultiMap<QString, QString> attributes = FaceTags::identityAttributes(tagId); |
| 654 |
Identity identity = IdentityProvider::instance()->findIdentity(attributes); |
| 655 |
|
| 656 |
if (!identity.isNull()) |
| 657 |
{ |
| 658 |
exclusionIdentityIds << identity.id(); |
| 659 |
} |
| 660 |
} |
| 661 |
|
| 662 |
// classify the features |
| 663 |
|
| 664 |
package->labelList << classifier->predict(package->featuresList[i], exclusionIdentityIds); |
| 665 |
} |
| 666 |
else |
| 667 |
{ |
| 668 |
package->labelList << FaceClassifier::UNKNOWN_LABEL_ID; |
| 669 |
} |
| 670 |
} |
| 671 |
|
| 672 |
enqueue(nextQueue, package); |
| 673 |
|
| 674 |
package = nullptr; |
| 675 |
} |
| 676 |
|
| 677 |
/* ========================================================================================= |
| 678 |
* End pipeline stage specific loop |
| 679 |
*/ |
| 680 |
|
| 681 |
MLPIPELINE_LOOP_END(MLPipelineStage::Classifier, "FacePipelineDetectRecognize::classifier"); |
| 682 |
|
| 683 |
/* ========================================================================================= |
| 684 |
* Pipeline stage specific cleanup |
| 685 |
* |
| 686 |
* Use the block from here to MLPIPELINE_STAGE_END to clean up any resources used by the stage. |
| 687 |
*/ |
| 688 |
|
| 689 |
MLPIPELINE_STAGE_END(MLPipelineStage::Classifier, MLPipelineStage::Writer); |
| 690 |
} |
| 691 |
|
| 692 |
bool FacePipelineDetectRecognize::writer() |
| 693 |
{ |
| 694 |
ActionThreadBase::setCurrentThreadName(QString::fromLatin1("FaceDetReco%1").arg(QLatin1String(__FUNCTION__))); // To customize thread name |
| 695 |
|
| 696 |
MLPIPELINE_STAGE_START(QThread::LowPriority, MLPipelineStage::Writer, MLPipelineStage::None); |
| 697 |
FacePipelinePackageBase* package = nullptr; |
| 698 |
|
| 699 |
/* ========================================================================================= |
| 700 |
* Pipeline stage specific initialization code |
| 701 |
* |
| 702 |
* Use the block from here to MLPIPELINE_LOOP_START to initialize the stage. |
| 703 |
* The code in this block is run once per stage initialization. The number of instances |
| 704 |
* is at least 1. More instances are created by addMoreWorkers if needed. |
| 705 |
*/ |
| 706 |
|
| 707 |
const IdentityProvider* const idProvider = IdentityProvider::instance(); |
| 708 |
FaceUtils utils; |
| 709 |
ThumbnailLoadThread* const thumbnailLoadThread = ThumbnailLoadThread::defaultThread(); |
| 710 |
const QList<AlbumRootInfo> roots = CoreDbAccess().db()->getAlbumRoots(); |
| 711 |
Q_UNUSED(roots); |
| 712 |
|
| 713 |
MLPIPELINE_LOOP_START(MLPipelineStage::Writer, thisQueue); |
| 714 |
package = static_cast<FacePipelinePackageBase*>(mlpackage); |
| 715 |
|
| 716 |
/* ========================================================================================= |
| 717 |
* Start pipeline stage specific loop |
| 718 |
* |
| 719 |
* All code from here to MLPIPELINE_LOOP_END is in a try/catch block and loop. |
| 720 |
* This loop is run once per image. |
| 721 |
*/ |
| 722 |
|
| 723 |
{ |
| 724 |
switch (settings.alreadyScannedHandling) |
| 725 |
{ |
| 726 |
case FaceScanSettings::Rescan: |
| 727 |
{ |
| 728 |
// remove old unconfirmed face rects |
| 729 |
|
| 730 |
QList<FaceTagsIface> oldEntries = utils.unconfirmedFaceTagsIfaces(package->info.id()); |
| 731 |
utils.removeFaces(oldEntries); |
| 732 |
break; |
| 733 |
} |
| 734 |
|
| 735 |
case FaceScanSettings::ClearAll: |
| 736 |
{ |
| 737 |
// remove all face rects |
| 738 |
|
| 739 |
utils.removeAllFaces(package->info.id()); |
| 740 |
break; |
| 741 |
} |
| 742 |
|
| 743 |
case FaceScanSettings::RecognizeOnly: |
| 744 |
case FaceScanSettings::Skip: |
| 745 |
{ |
| 746 |
// do nothing |
| 747 |
// Skipped images were skipped in the finder stage |
| 748 |
// RecognizeOnly defines the pipeline |
| 749 |
|
| 750 |
break; |
| 751 |
} |
| 752 |
} |
| 753 |
|
| 754 |
// mark the image as scanned |
| 755 |
|
| 756 |
utils.markAsScanned(package->info); |
| 757 |
|
| 758 |
// create thumbnails and write the new face rects to the database |
| 759 |
|
| 760 |
QStringList names; |
| 761 |
|
| 762 |
if (package->faceRects.size()) |
| 763 |
{ |
| 764 |
QList<FaceTagsIface> databaseFaces; |
| 765 |
QList<Identity> identities; |
| 766 |
QList<QRectF> paddedRectList; |
| 767 |
|
| 768 |
for (int i = 0 ; i < package->faceRects.size() ; ++i) |
| 769 |
{ |
| 770 |
// pad the rect by FaceTagsIface::faceThumbnailResizeFactor (1/2 on each side) |
| 771 |
|
| 772 |
QRectF paddedRect = QRectF(qMax(0.0, package->faceRects[i].x() - (package->faceRects[i].width() * (FaceTagsIface::faceThumbnailResizeFactor / 2.0F))), |
| 773 |
qMax(0.0, package->faceRects[i].y() - (package->faceRects[i].height() * (FaceTagsIface::faceThumbnailResizeFactor / 2.0F))), |
| 774 |
qMin(1.0 ,package->faceRects[i].width() + (package->faceRects[i].width() * FaceTagsIface::faceThumbnailResizeFactor)), |
| 775 |
qMin(1.0 ,package->faceRects[i].height() + (package->faceRects[i].height() * FaceTagsIface::faceThumbnailResizeFactor))); |
| 776 |
|
| 777 |
paddedRectList << paddedRect; |
| 778 |
|
| 779 |
QRect faceRect(std::round(package->image.width() * paddedRect.x()), |
| 780 |
std::round(package->image.height() * paddedRect.y()), |
| 781 |
std::round(package->image.width() * paddedRect.width()), |
| 782 |
std::round(package->image.height() * paddedRect.height())); |
| 783 |
|
| 784 |
if (package->labelList[i] != FaceClassifier::UNKNOWN_LABEL_ID) |
| 785 |
{ |
| 786 |
Identity identity = idProvider->identity(package->labelList[i]); |
| 787 |
names << identity.attribute(QStringLiteral("name")); |
| 788 |
identities << identity; |
| 789 |
databaseFaces << FaceTagsIface(FaceTagsIface::Type::UnconfirmedName, |
| 790 |
package->info.id(), |
| 791 |
FaceTags::unconfirmedPersonTagId(), |
| 792 |
TagRegion(faceRect)); |
| 793 |
} |
| 794 |
else |
| 795 |
{ |
| 796 |
identities << Identity(); |
| 797 |
databaseFaces << FaceTagsIface(FaceTagsIface::Type::UnknownName, |
| 798 |
package->info.id(), |
| 799 |
FaceTags::unknownPersonTagId(), |
| 800 |
TagRegion(faceRect)); |
| 801 |
} |
| 802 |
} |
| 803 |
|
| 804 |
// store the thumbnails |
| 805 |
|
| 806 |
if (!package->image.isNull()) |
| 807 |
{ |
| 808 |
utils.storeThumbnails(thumbnailLoadThread, package->info.filePath(), |
| 809 |
databaseFaces, package->image); |
| 810 |
} |
| 811 |
|
| 812 |
// write the new face rects to the database |
| 813 |
|
| 814 |
utils.writeUnconfirmedResults(package->info.id(), |
| 815 |
paddedRectList, |
| 816 |
identities, |
| 817 |
package->image.originalSize()); |
| 818 |
} |
| 819 |
|
| 820 |
QString albumName = CollectionManager::instance()->albumRootLabel(package->info.albumRootId()); |
| 821 |
|
| 822 |
// send a notification that the image was processed |
| 823 |
|
| 824 |
notify(MLPipelineNotification::notifyProcessed, |
| 825 |
package->info.name(), |
| 826 |
albumName + package->info.relativePath(), |
| 827 |
names.join(QLatin1String(", ")), |
| 828 |
package->faceRects.size(), |
| 829 |
package->thumbnailIcon); |
| 830 |
|
| 831 |
// delete the package |
| 832 |
|
| 833 |
delete package; |
| 834 |
|
| 835 |
package = nullptr; |
| 836 |
} |
| 837 |
|
| 838 |
/* ========================================================================================= |
| 839 |
* End pipeline stage specific loop |
| 840 |
*/ |
| 841 |
|
| 842 |
MLPIPELINE_LOOP_END(MLPipelineStage::Writer, "FacePipelineDetectRecognize::writer"); |
| 843 |
|
| 844 |
/* ========================================================================================= |
| 845 |
* Pipeline stage specific cleanup |
| 846 |
* |
| 847 |
* Use the block from here to MLPIPELINE_STAGE_END to clean up any resources used by the stage. |
| 848 |
*/ |
| 849 |
|
| 850 |
MLPIPELINE_STAGE_END(MLPipelineStage::Writer, MLPipelineStage::None); |
| 851 |
} |
| 852 |
|
| 853 |
void FacePipelineDetectRecognize::addMoreWorkers() |
| 854 |
{ |
| 855 |
/* ========================================================================================= |
| 856 |
* Use the performanceProfile metrics to find the slowest stages |
| 857 |
* and add more workers to those stages. |
| 858 |
* |
| 859 |
* For the Face detection and recognition pipeline, the loader is the |
| 860 |
* slowest stage so add 3 more loaders, 2 more extractors, and 1 more classifier. |
| 861 |
*/ |
| 862 |
|
| 863 |
addWorker(Loader); |
| 864 |
addWorker(Loader); |
| 865 |
addWorker(Loader); |
| 866 |
addWorker(Extractor); |
| 867 |
addWorker(Extractor); |
| 868 |
addWorker(Classifier); |
| 869 |
} |
| 870 |
|
| 871 |
} // namespace Digikam |
| 872 |
|
| 873 |
#include "moc_facepipelinedetectrecognize.cpp" |