UNCONSTRAINED AND MULTI VIEW FACE DETECTOR

Abstract

The mainfprecondition for applicationsfsuch as facefrecognition and facefde-identification forfprivacy protectionfis efficient facefdetection in realfscenes.The proposal isfa multi stage cascadefmodel for facefdetection . The cascadedftwo-stage model isfbased on the fastfnormalized pixel difference (NPD) detectorfat the firstfstage, and MTCNN based CNNfat thefsecond stage. The outputsfof the NPD detectorfare having small numberfof false negative (FN) and a muchfhigher number of falsefpositive face (FP) detections.Order of magnitudefof FP detections areftypically higherfthan the FN ones. Due tofthis very highfnumber of FPs hasfa negative impactfon recognition andfde-identification processingftime and on thefnaturalness of thefde-identified images. Tofsuppress the effect offlarge numberfof FP face detections, afCNN is used atfthe secondfstage. The CNN is appliedfonly on face regionfsolution obtainedfby the NPD detectorfthat have an NPDfscore in the intervalfbetween twofexperimentally determinedfthresholds. The experimentalfresults on thefpart of the Face Detection Dataset and Benchmark (FDDB) showfthat the hybrid cascadefmodel significantly reducesfthe number of FPfdetections while the numberfof FN detections are onlyfslightly increased.f