Johannes Merkle (secunet)：開源人臉圖像質量 (OFIQ).pdf

《Johannes Merkle (secunet)：開源人臉圖像質量 (OFIQ).pdf》由會員分享，可在線閱讀，更多相關《Johannes Merkle (secunet)：開源人臉圖像質量 (OFIQ).pdf（19頁珍藏版）》請在三個皮匠報告上搜索。

1、Johannes Merkle01.04.2025Open Source Face Image Quality(OFIQ)https:/de.wikipedia.org/wiki/StyleGAN2Agenda|OFIQ Objectives Algorithms Release Way Forward3Objectives|OFIQ C+software library for facial image quality assessment(FIQA)Checks quality requirements from ISO/IEC 39794-5:2019 Open source publi

2、shed under liberal licences Commercial use possible,no copy-left Support of many plattforms(incl.mobile devices)Evaluation through NIST FATE Quality SIDD and internal test Reference implementation of upcoming revision of ISO/IEC 29794-5 Development funded by BSI4Pre-Processing Algorithms|OFIQ Face D

3、etection Face Landmark Estimation Alignment Segmentations:Landmarked Region Occlusion Segmentation Face Parsing 5Algorithms|OFIQ Unified Quality Score Background Uniformity Illumination Uniformity Moments of the Luminance Distribution Under-Exposure/Over-Exposure Dynamic Range Sharpness No Compressi

4、on Artifacts Natural Colour Single Face Present Eyes Open Mouth Closed Eyes Visible Mouth Occlusion Prevention Face Occlusion Prevention Inter-Eye Distance Head Size Crop of the Face Head Pose Expression Neutrality No Head Coverings6Algorithms-Unified Quality Score|OFIQNot limited to certain quality

5、 defectsCNN MagFace(iResNet 50 model)Excellent results in FATE Quality 1st out of 52 algorithms1Good prediction of face recognition scores1Measured by FNMR after removal of 5%lowest quality images7Algorithms-Sharpness|OFIQRandom Forest classifier Several features:Sobel-Filter Laplace filter Differen

6、ce of image from mean-filtered imageRestricted to landmarked region Trained on synthetic and real blur8Algorithms-SharpnessGood results in FATE Quality 5th out of 34 Only synthetic blur Internal evaluation on FRGCv2(real blur)Accuracy high but not very high Challenging9Algorithms-No Compression Arti

7、facts|OFIQCNN trained by secunet Trained to predict the PSNR between compressed and uncompressed image Trained on JPEG and JPEG2000 artifacts Also trained on scaled and rotated artifacts10Algorithms-No Compression Artifacts|OFIQResults in FATE Quality not bad 10th out of 20 Only JPEG tested No scali

8、ng or rotation after compressionInternal evaluation of test sets based on FFHQ Test set 1:Only scaling after compression Test set2:Scaling and rotation after compression11Algorithms-Eyes Open and Mouth Closed|OFIQAlgorithms based on landmarksMaximum distance between lids/lipsNormalized by distance T

9、 between eyes midpoint and chin12Algorithms-Eyes Open and Mouth Closed|OFIQVery good results in FATE Quality 1st out of 22(Eyes Open 2)2nd out of 19(Mouth Open 2)Very good results in internal evaluation No ethnic bias found for Eyes Open No errors on CAS-PEAL-R113Algorithms-Face Occlusion Prevention

10、|OFIQBased on face occlusion segmentationCNN from repository FaceExtraction1Fraction of un-occluded landmarked regionSimilar approach used for Mouth Occlusion Prevention and Eyes Visible1 https:/ Face Occlusion|OFIQExcellent performance in FATE Quality 1st of 20High accuracy on COFW test set15Algori

11、thms-Head Pose|OFIQCNN from repository 3DDFA_V21Results in FATE Quality vary strongly among test sets Yaw:19th/4th/15th Pitch:4th/34th/16th out of 36 algorithms Roll:10thResults heavily tainted by errors of face detection1 https:/ of OFIQ|OFIQ Release 1.0.0 published in November 2024https:/ Implemen

12、tation&Evaluation Report published 1 Regular maintenance Current release is 1.0.11 https:/ Forward|OFIQ Development of OFIQ 2.0 has begun Objectives remain unchanged Planned to be completed by end of 2027 Close alignment with revision of ISO/IEC 29794-5 Input from community is very welcome OFIQ user

13、 group virtual meeting:https:/eab.org/events/program/37418Way Forward Potential Improvements|OFIQ Computational performance E.g.Landmark Estimation,Face Parsing Accuracy E.g.Background Uniformity Reduction of demographic bias Under Exposure Prevention Additional quality checks Motion Blur,Eyes Looking to the Camera More information:https:/