Eye Detection Based on Improved AD AdaBoost Algorithm

Eye Detection Based on Improved AD AdaBoost AlgorithmBenke XiangCollege of Computer and Information Science SouthwestUniversityChongqing, P.R Chinaxiangbenke@Xiaoping ChengCollege of Computer and Information Science SouthwestUniversityChongqing, P.R Chinaxpcheng@Abstract—Eye detection is an important step in eye tracking and eye state recognition. An improved AD AdaBoost algorithm for eye detection is proposed to slow the degradation in training step. Weight on negative samples which are classified correctly is released then the other samples' weight is normalized to slow the expansion of weight on difficult samples. The experiment results show that the approach proposed is real time and has a higher detection accuracy.Keywords-Eye Detection; AD AdaBoost; Cascaded ClassifierI.I NTRODUCTIONThe human eye, because of its extremely important information, has become a hot research field of pattern recognition. The research on its tracking and pattern recognition has been widely used in the areas such as fatigue testing, specific user man-machine interface, animation synthesis, automated access control and security monitoring [1]. Human eye detection is the first step of the human eye tracking and human eye state identification, so it has attracted more and more attention.In the detection process, the method which based on statistics, which through training and learning on a large number of target and non-target samples [2], obtained a set of model parameters, and then building a classifier based on the model, or filters to detect targets. It takes full advantage of the weak learning theorem in machine learning [3], making the classification accuracy greatly improved through extensive studies. It has gradually become the mainstream technology in research area of human eye detection.AD AdaBoost algorithm in statistical learning methods [4] is the latest development of AdaBoost, it reduces the error rate of negative samples and optimize the disadvantage of traditional AdaBoost algorithm which can only make the upper bound of classification error rate minimal, so it is more suitable for the target detection problem. However, the training process, as number of iterations in the classifier increases, the weights of difficult samples expanded in the entire sample set, making the rate of simple sample values decrease in the entire sample set, resulting in the classifiers produces bias on the sample [5]. The already-generate accurate classification rules have been destroyed, the entire system recognition rate decreases, resulting in degradation.This article will use AD AdaBoost for the detection of human eye, taking into account the clarity of human eye's own outline, sharp color contrast, using a large number of human eye samples for statistical learning. Releasing correctly classified weight of negative samples in the classifier training process, and through weight normalization to solve the problem of expansion on weights of buffering difficult samples, slowing the degradation in classification process, improving detection accuracy.II.C ASCADE STRUCTURE CLASSIFIER TRAININGA.AD AdaBoost Cascade StructureCascade structure of classifiers which is proposed by Paul Viola [6] is a set of serial classifiers. In the treatment of classification on identifying samples, only the samples which are pointed positive by classifier in the front were sent to classifier in the back to be addressed later, to the contrary, it is seen as negative sample and to be outputted directly. Finally, only the samples which are judged as positive at every level of the classifier are to serve as the final output of the positive samples. Cascade structure as shown in Figure 1.Figure 1 Schematic diagram of cascadeT indicates that this layer is determined as positive, F indicates negative samples. The preceding level classifier structure is simple, using fewer number of features, but can filter out large numbers of negative samples which were quite different with the target samples. The latter class uses more features and more complex structures and thus be able to distinguish between negative samples which has high similarity with the target samples. In the actual target detection problem, because the positive samples to be tested are usually only a small proportion of the overall sample, most of the samples was filtered out in the classifier of cascade structure in the front, and only a small number ofsamples need to go through all levels of the classifiers, and thus cascade structure significantly reduces the computational complexity.B.Enhancing Mechanism in the Training of the SampleWeightsThe basic idea of AD AdaBoost is to integrate multiple weak classifiers into a strong classifier. In the training process, all of the positive and negative samples were given an equal initial weights. When one classifier training is completed, based on their classification result on the training set of samples, we make adjustment on all of the sample weights, the correctly classified samples by previous class, its weight decreases when it enters the next iteration. The mistakenly classified samples by previous class, its weight increases when it enters the next iteration, making the next weak classifier training are more concerned about samples which have been misidentified [4]. The final judgment result of strong classifier weighted sum of all judgment results of all weak classifiers.However, with the increase of the iterations in training, the algorithm will gradually shift its focus to the sample difficult to categorize, which leads to the expansion of the weight of difficult samples within the scope of the entire sample set. When the weight of the difficult sample is too large, the weak classifier will increase emphasis for each loop and the emphasis will improve with the increasing weight; meanwhile, due to the sample weight normalization after each classification, the proportion of simple sample weights will severely reduced in the entire sample, resulting in bias, that the accurate classification rules will be destroyed, the entire system recognition rate decreases, resulting in degradation [5].C.Analysis for the Training of Cascade StructureClassifierThrough the learning of a large number of the positive samples with target type and negative samples with non-target type in the training stage, AdaBoost has accessed to the testing classification rules.Suppose the final cascade structure classifiers obtained a total of N levels (N> 1), samples were concentrated with a negative sample A (xa, -1), its corresponding classification rules is ¾. During training phase, if the first t-level (t <N) classifier correctly classified negative samples of A, the first t-class classifier obtain the rule of ¾, while input the first t +1 level after the reduce of weight of A, which continue to be used for the training of next classifier. Because of the repeat study for correctly classified negative sample of A, the first t +1 level classifier also obtained discrimination rule ¾. During the actual testing phase, if the negative samples A are classified correctly in the t -level, it will be output directly asa negative sample and will not be proceed to the next level for further distinguish. Thus, judge rule ¾ for A acquired by t +1 level class classifier has lost its practical value.To sum up, A is invalid sample for the t +1 level classifier during the training process, which reduces the effective capacity of the training samples while the continuing possession of weight which will compress weight space of other effective samples, resulting in distortions in the distribution of weight. Therefore, the improvement of the AD AdaBoos is marking the correctly classified negative samples, then releasing the sample weights in time, and finally normalizing to buffer the expansion of difficult samples' weights.III.I MPROVED AD A DA B OOST FOR HUMAN EYEDETECTION$ Algorithm ImprovementAlgorithm improvements from the following aspects:Ł for the training samples, set identification state markfi the initial value of +1;ł in iterative process of the algorithm, if a negative samples were correctly classified in a certain level, modify fi =- 1;Ń for the fi =- 1 sample, release its weight, Doing normalization operation after updating weights of remaining samples.ń fi = +1 samples will continue to be used for training in next class classifier.% The Improved AD AdaBoost for Human Eye Detection The specific algorithm is as follows:1. For a given set of training samples for calibration: (x1, y1, f1), (x2, y2, f2 ),...,( xm, ym, fm), where xi∈ X, yi∈Y = (-1, + 1), fi∈F = (-1, +1);2. initialize the sample weight and identification state: D1(i) = 1 / m, fi = +1;3. Circle t = 1,2 ,..., T:3.1 The current distribution of weights of Dt, according to every single-feature, we use fi = +1 samples to train a weak classifier for human eye identification, in these classifiers, we selected the classifier with smallest error rate as a weak classifier ht in this circle;3.2 selected ht, calculate the weighted error rate:)()(iDiityxhtt¦≠=εAnd identify the correct weighted sum of positive samples:)(1)(,1iDpitixhytt¦===3.3 Solving the parameters weak weighted classifier ht:tp t tt e k ⋅+−=1ln[21εαWhere k is a constant, must reduce the upper bound of least error rate in this circle [4], often taking 1 / 120.3.4 Update sample identification condition,circulating t '= 1,2 ,..., m:endiff set y x h if i i i t 11&&1)(−=−=−= 3.5 update the sample weights for the next cycle:°¯°®­≠−==+==×=−+i i t i i i t i i i t tt t yx h e y y x h y y x h e Z i D i D t t )(,1&&)(,01&&)(,)()(1ααWhere Zt is the normalization factor:¦+=⋅⋅−⋅=1)()(i i t i t f x h y tt e i D Z α4. The final strong classifier shall be:»¼º«¬ª−=¦=T t t t Th x h sign x H 1)()(α Where Th is the threshold value which is set manually , the default is 0.For the training samples, add identification mark fi initial value +1, in the subsequent iteration, if a negative sample is correctly classified in this layer set fi =- 1, it will not participate in the next round of training. negative samples which have already been correctly classified are not taken into account when samples are update and normalized.IV.E XPERIMENT$ Training SamplesWhen we do Training on the AD AdaBoost, the positive samples are selected from the Feret Face Database. In the library, 5000 images from 14051 are selected as positive samples. Negative samples were mainly from scenes photo which are downloaded from the Internet, and the others are taken from the plants and animals photos taken by my own. A total of about 8000 photos.% Experimental ResultsExperiment is operated on a P Č 2.4GHz desktop machine, end up with 12 classes of human eye detection, 433 features, with a total of 7 training days. A 300 × 240 image takes about 16ms. testing samples tested were selected from Feret face database and human photos downloaded from the Internet, a total of 323. The test results are shown in Table 1. Experimental results show that the accuracy of the method in this article is better than the previous method. Some test results are shown in Figure 2.Table 1 Test results for the human eye detectionAdaBoostAD AdaBoos t Improved ADAdaBoostRecognitionrate93.8% 96.3% 97.2% False detectionrate 6.2% 3.7%2.8% False Report2012 9Figure 2 examples of test resultsAs can be seen from the table, AD AdaBoost have made considerable improvements than the traditional AdaBoost in the recognition rate, and the Improved AD AdaBoost method in this article, which mark the indentification state of samples, has improvement on recognition rate compare to the traditional AD AdaBoost.V.S UMMARYThis paper presents an improved AD AdaBoost algorithm useing to the human eye detection. Compare to traditional AdaBoost, AD AdaBoost effectively reduces the negative sample error rate while positive sample error rate is relatively low, improving the identification ability of overall classifiers. Also, aiming at the AD AdaBoost’ degradation phenomenon which appears in the experiment, this paper proposed an improved algorithm by marking identification state of samples to slow spread of degradation. So that AD AdaBoost recognition capability has been further improved. This improved algorithm can only mitigate the spread of degradation, because there is no absolute appropriate restrictions on upper bound on the increase of difficult sample weights in iteration, marking on samples can not fundamentally put an end on rise of degradation phenomenon. So, the enhancing mechanism of samples weights is a topic needs further research.R EFERENCES[1]TANG Jin, XU Hai-zhu, WANG Li, “Survey on human eyesdetection in images, ” Application Research of Computers, vol 25, Apr. 2008, pp. 961-965[2]Freund Y and Schapire R E, “A decision-theoretic generalization ofon-line learning and an application to boosting, ” Journal of Computer and System Sciences, vol 55, Jan. 1997, pp. 119-139.[3]Schapire R E, “The strength of weak learnability, ” Machine Learning,vol 5, Feb. 1990, pp. 197-227.[4]LI Chuang, DING Xiao-qing and WU You-shou, “A RevisedAdaBoost Algorithm-AD AdaBoost, ” CHINESE JOURNAL OF COMPUTERS, vol 30, Jan. 2007, pp. 103-109.[5]Quinlan J R, “Bagging,boosting,and C4.5, ” Proc of the ThirteenthNational Conference on Artificial Intelligence, 1996, pp. 725-730. [6]Paul Viola and Michael J.Jones, “Robust Real-time ObjectDetection, ” UK: Cambridge Research Laboratory, Feb. 2001.。