对SMQT特征做了很好的优化，效果非常好，有利于人脸检测的实时Top EI Features for Face DeteconO SMQT Feat urEs-Left to Right Tcp to Botom. ERR阻F盟西凸 ILHFLE回平■园3Fig 1. 60 Most Important Features for Face DetectionFig 4. 60 Least Important Fcaturcs for Charactcr RccognitionHmm即 Fet psnr Face detectFurthermore, if we save the lookup table in terms ofint8 rather than a float or double we could further improveI EIthe compression Table i presents the memory needed tostore lookup table for various configurationsFeatures( F)IndoubleⅪ"hE巴133 KB103KB3020 KB156KB39 KB311KB3和4T512330KB2.7MBFig 2. 60 Least Important Features for Face DetectionTable 1. Size of the classifier lookup table for local area size D)3x3 and SMQT level(L)-1 and P-648 pixelslRT而Bdm5. RESULTS品出:Face detection and character recognition are two commonasks investigated by computer vision researchers. TheILDprocedure described in Section 4 is tested on both of these5.1 Face detection2Images are collected from the well-known CMU+MIT和40frontal face database and a private face database. We haveFig 3. 60 Most Important Features for Character Recognitionused 1700 face images from the private face database totrain SNow classifier lookup tables. For testing purposesDefined by Eq9, Figures I and 2 show the most and we have used the CMU+MIT frontal view image databaseleast important features for face detection. Similarly, figures which contains 144 images with 534 face images. We3 and 4 show the most and least important features for generated non-face images by sampling the non-face regionscharacter recognition. As is seen in Figures 1 and 3, the within the same images. In this way we generated 977 nonmost important features for both tasks are those features that face images for CMU+mit datr osr and 2200 for theseem to be able to capture relevant object structure, while as private image databaseseen in Figures 2 and 4, the least important features areIn this face detection study, performance betweenthose features that are more"noise like". This intuitively different SNow configurations is compared using theAttractive result justifies storing only the most important receiver operation characteristic curve(ROC). Area underfeatures to reduce the storage burden of the SNow classifier receiver characteristic curve typically represents the qualityusing SMQT features. By not using the features that are of the classifier. Performance of the efficient SNowassociated with more"noise like"structure in the image, classifier using local SMQT features of various sizes can bevery little class distinguishing information is lost, while seen in Fig 4. Table 1 illustrates the quantitative area undergreatly reducing the storage burdencurve (auc) value for different configurations We achievethe highest performance, 0.79, when we use all 512 features original dense snow classifier while reducing the storagein our classification. However, performance is comparable, requirements considerably. In Section 5, experiments using0.75, when we use only 60 features in our SNow the well-known CMU-MIT image database and a privateclassification scheme. Thus comparable performance is character database are used to demonstrate the effectivenessachieved while requiring only 2.5% of the original storage of the proposed methodburden. Therefore, this could be useful in applications whereROCs Icr Reduced Set of Snow Featureswe have to worry about the size of our lookup table斗二斗0.90.8---in8F60-+-int8F30int8-F20-intB F60I09609660970975099099099095Accurac0.10.20.3040.50Fig 6. Character Recognition results using a reduced set of0.8teaturesFig 5. Detection results on MIT+ CMU (120 images, 534 face7. REFERENCESimages and 1511 patches to analyze)database1 B. Froba, A Ernst, Face Detection with modified censusdouble int8 int8-F60 int8-T30 int8-F20 transform, "in JEFE International Conference on Automatic FaceAUC0.7920.7920.750.6920.692and Gesture Recognition (AFGR), May 2004, pp 91-96Table 2. Area under receiver operating characteristic curve (AUC)[2 P. Viola, M. Jones, Rapid object Detection using a boostedcascade of simple features", CVPR, 2001, vol. 1, pp 511-5185.2 Character recognition3]D. Roth, M. Yang, N. Ahuja, A snow-based face detector, inA similar process was conducted for a character recognition Advances in Neural Information Processing Systems(NIPS), pplask using a private database derived from license plate 855-861, MIT Press 2000images. In this problem a one-versus-all strategy was usedfor the multi-class detection task. All other aspects were [4]M. Nilsson, J. Nordberg, I. Claesson,"Face Detection usingvery similar to what is described above. Figure 6 depicts the Local SMQT Features and Split up Snow Classifier. "in IEEEresults of applying the reduced feature set technique to thisInternational Conference on Acoustic, speech and Signaltask. Note that in this graph, yield is defined as the total Processing(ICASSP), 2004, vol 2, pp589-592number of decisions made compared to the total number ofopportunities, while accuracy is defined as the number of [5]O. Lahdenoja, M. Laiho, A. Paasio, Reducing the featurevector length in local binary pattern based face recognition, "incorrect decisions made divided by the total number of IEEE International Conference on Image Processing(cIPdecisions made. This is done, since in this problem, a third September 2005, vol 2, pp.914-917category of"no decision" is defined, which reduces theyield, but with proper design, increases the accuracy. [6] M.H. Yang, D. Kriegman, N. Ahuja, "Detecting facesaccuracy encompasses errors due to false positives andimages: A survey, IEEE Transactions on Pattern Analysisfalse negatives. note that here using only 30 features storedMachine intelligence(PAmi), vol 24, no l, pp 34-58, 2002as integers achieves comparable performance to using 512features stored as doubles. Thus comparable performance 7J.Li,"Face Detection Using SURF Cascade",ICCV, 2011was achieved while requiring only 1. 5% of the storage[8 X. Zhu, D. Ramanan, "Face Detection, Pose Estimation, andburdenLandmark localization in the wild, CVPr 20126. CONCLUSIONS9 Nilsson, M, Dahl, M, and Claesson, I, "The successive mThis paper presents an efficient local SMQT features basedquantization transform,Proc. ICASSP 2005, 429-432(2105 canSNoW classification method. In this study we show that aSNow classifier with only a limited number ofdistinguishing SMQt features performs as well as the