《Rapid Object Detection using a Boosted Cascade of Simple 采用升压级联简单的快速目标检测》由会员分享,可在线阅读,更多相关《Rapid Object Detection using a Boosted Cascade of Simple 采用升压级联简单的快速目标检测(35页珍藏版)》请在金锄头文库上搜索。
1、Rapid Object Detection using a Boosted Cascade of Simple FeaturesOriginal Author Paul Viola & Michael JonesIn: Proc. Conf. Computer Vision and Pattern Recognition. Volume 1., Kauai, HI, USA (2001) 511518Speaker: Speaker: JingJing Ming Ming ChiuanChiuan ( (井民全井民全) )(moving or acting with great speed)
2、 (increase the strength or value of Sth) OutlinenIntroductionnThe Boost algorithm for classifier learningnFeature SelectionnWeak learner constructornThe strong classifiernA tremendously difficult problemnResultnConclusionWhat had we done?nA machine learning approach for visual object detectionnCapab
3、le of processing images extremely rapidlynAchieving high detection ratesnThree key contributionsnA new image representation Integral ImagenA learning algorithm( Based on AdaBoost5)nA combining classifiers method cascade classifiersSelect a small # of visual features from a larger set yield an effici
4、ent classifiersSpeed up the feature evaluationDiscard the background regionsof the imageWorking only with a single grey scale imageA demonstration on face detectionnA frontal face detection systemnThe detector run at 15 frames per second without resorting to image differencing or skin color detectio
5、n Image difference in video sequences384 x 288 on a PentiumIII 700 MHzThe broad practical applicationsfor a extremely fast face detectornUser Interface, Image Databases, TeleconferencingnThe system can be implemented on a small low power devices. Compaq iPaq 2 frame/secTraining process for classifie
6、rnThe attentional operator is trained to detect examples of a particular class - a supervised training processIn the domain of face detectionX is a face False positiveFalse negativeAdaBoostingnPlace the most weight on the examples must often misclassified by the preceding weak rulesnForcing the base
7、 learner to focus its attention on the “hardest” examplesThe Boost algorithm for classifier learningStep 1: Giving example imagesStep 2: Initialize the weights For t = 1, , T1. Normalize the weights,2. For each feature j, train a classifier hj which is restricted to using a singlefeature3. Update th
8、e weights: Weak learner constructor Final strong classifier Selected the weaker classifiersThe Big Picture on testing processAda Boosting LearnerFeature setFeature Select & ClassifierStage 1False (Reject)Ada Boosting LearnerStage 2 PassFalse (Reject)Ada Boosting LearnerStage 3PassFalse (Reject)Rejec
9、t as many negatives as possible (minimize the false negative)100% Detection Rate 50% False PositiveA tremendously difficult problemnHow to determinenThe number of classifier stagesnThe number of features in each stagesnThe threshold of each stageAda Boosting LearnerTraining exampleFeature Select & C
10、lassifierStage 1False (Reject)faceNon-face100% Detection Rate 50% False PositiveAda Boosting LearnerStage 2PassFalse (Reject)ResultnA 38 layer cascaded classifier was trained to detect frontal upright facesnTraining set: nFace: 4916 hand labeled faces with resolution 24x24.nNon-face: 9544 images con
11、tain no face. (350 million subwindows within these non-face images)nFeaturesnThe first five layers of the detector: 1, 10, 25, 25 and 50 featuresnTotal # of features in all layer 6061ResultnEach classifier in the cascade was trainednFace : 4916 + the vertical mirror image 9832 imagesnNon-face sub-wi
12、ndows: 10,000 (size=24x24)Outline ResultnSpeed of the final DetectornImage ProcessingnScanning the DetectornIntegration of Multiple DetectornExperiments on a Real-World Test SetSpeed of the final DetectorResultnThe speed is directly related to the number of features evaluated per scanned sub-window.
13、nMIT+CMU test setnAn average of 10 features out of a total 6061 are evaluated per sub-window.nOn a 700Mhz PentiumIII, a 384 x 288 pixel image in about .067 seconds (using a staring scale of 1.25 and a step size of 1.5)Image ProcessingResultnMinimize the effect of different lighting-conditionsnVarian
14、ce normalized reference: http:/www.ic.sunysb.edu/Stu/sewang/papers/Fingerprint%20Classification%20by%20Directional%20Fields.pdf Scanning the DetectorResultnThe final detector is scanned across the image at multiple scale and locationsnGood results are obtained using a set of scales a factor of 1.25
15、apartnLocations are obtained by shifting the window some pixels nIf the current scale is s, the window is shifted by Scale is achieved by scaling the detector itself rather than the imageis the rounding operationIntegration of Multiple DetectorResultnMultiple detections will usually occur around eac
16、h face and some types of false positives. nA post-process to detected sub-windows in order to combine overlapping detections into a single detectionnTwo detections are in the same subset if their bounding regions overlapExperiments on a Real-World Test SetResultThe MIT+CMU frontal face test set consists of 130 images with 507 labeled frontal facesDetection rates for various numbers of false positives on the MIT+ CMU test set containing 130 im