J. Kittler, M. Hatef, R.P.W. Duin and J. Mates, On combining classifiers, IEEE Transactions on Pattern Analysis and Machine Intelligence 20(3) (1998) 226239.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....dictionary includes, for the considered memory under test, the most relevant failure patterns. Fig. 3 shows an example of the dictionary, with four different failure patterns. In these conditions, while a subset of errors is captured from the BIST execution, a Hough transformation (HT) is applied [13 14], in the space indicated by the models of the fault dictionary. III spot cluster 1 1 1 1 1 1 colu m n row Figure 3: An example of a fault dictionary. Hough transformation allows mapping the features recognized in an image space, to a set of points into a parameter space. In the ....

J. Illingworth, J. Kittler, The Adaptive Hough Transform, IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol. PAMI-9, No. 5, September 1987, pp. 690- 698

....detectors and that information is needed for our feature combination technique. Combining the features can be viewed as the classical problem of combining the results of multiple classifiers which has already received considerate attention in the pattern recognition and in the retrieval community [5, 3, 2, 6]. Combination techniques can be classified into techniques using fixed rules and methods including training on the outputs of the first stage classifiers. Training the combining classifier has been reported to considerately improve the performance of the total system in some cases, especially if ....

....measurement vector x and Q i (x) the output of the combining classifier. We did not formulate the rules in a probabilistic way, since we cannot know if the output of the classifiers quantifies the posterior probabilities of the class memberships. The two most well known fixed combination rules are [5, 2]: The product rule If the feature representations are statistically independent and the classifiers produce posterior probabilities, then this rule will produce an optimal classifier in the sense that it quantifies the true likelihood of the measurements given class k. Unfortunately, this is ....

[Article contains additional citation context not shown here]

J. Kittler, M. Hatef, and J. Matas. On combining classifiers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 20(3):226--239, March 1998.

....probability of the positive class (i.e. the genuine class) given the original pattern. This interpretation is related to the selected transformations. The Mean rule assumes that the posterior probabilities computed by the matchers do not deviate dramatically from the prior probabilities [8]. The Product rule assumes the independence of the matchers scores [8] In the case of the Logistic transformation, the assumption for computing the weights 2 1 0 , w w w is critical: given the above score interpretation, our choice is to maximise the likelihood according to the Bernoulli model ....

....given the original pattern. This interpretation is related to the selected transformations. The Mean rule assumes that the posterior probabilities computed by the matchers do not deviate dramatically from the prior probabilities [8] The Product rule assumes the independence of the matchers scores [8]. In the case of the Logistic transformation, the assumption for computing the weights 2 1 0 , w w w is critical: given the above score interpretation, our choice is to maximise the likelihood according to the Bernoulli model (the classes genuine and impostor are viewed as random variables ....

J. Kittler, M. Hatef, R.P.W. Duin, and J. Matas, On Combining Classifiers, IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 20 num. 3 (1998) 226-239.

.... meta classification combination strategy has been successfully applied to identify unknown people in the video stream [8] In this paper, we investigate how a meta classification strategy performs in the realm of video classification, and compare it with other strategies in a probability framework [7], which has also shown performance improvements in tasks such as biometric identity verification and handwritten digit recognition. While video has rich content, most features exploited by video classifiers in previous studies are audio visual features [15] such as color, motion, pitch, and ....

....them together. The combination strategy decides how judgments from several classifiers are combined to make the final decision. In this section we describe probability based and metaclassification strategies. 3. 1 Probability Based Strategies Kittler et al. proposed a probability based framework [7] to explain various combination strategies. Assume there are p classes w . Wp , and k classifiers to be combined, and the feature vector that the i th classifier observes is xi, i = 1, k. Without any further information, we will select the class wj with the highest posterior probability ....

Kittler, J., Hater, M., Duin, R.P.W. and Mates, J. On Combining Classifiers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 20(3), 1998.

....subspaces of the D dimensional feature space, seemed to be applicable. Hence, we finally present results of combined classifications derived from classifiers trained on different portions of the feature space. The considered combination strategies are those recently suggested by Kittler [6][7], namely product, sum, maximum and majority rules. Combination turned out to bring significant improvement, as individual classifiers seem to form a so called mixture of experts in our case. The task of discriminating two forest classes, three classes of agricultural area, two classes of built up ....

Josef Kittler, Mohamad Hatef, Robert P.W. Duin, and Jiri Matas. On Combining Classifiers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 20(3):226--239, March 1998.

....between dense and fatty breast types. In fulfilling these classification objectives we extract four groups of texture features from segmented digital mammograms. To improve the performance and robustness of the classifiers we use classification combination rules proposed by Kittler et al. [6]. The remainder of this paper is organised as follows; Section 2 describes the experimental method used including feature extraction, data dimensionality reduction, classifier training, testing and combination; Section 3 details the experimental results and finally conclusions from the study are ....

....2. 4 Classifier Training Testing and Combination To improve the performance and robustness of our developed system, we choose to implement a classifier combination paradigm such that we combine the decision of n component classifiers on test using combination rules proposed by Kittler et al. [6]. The assumption of conditional independence between each component classifier is an important aspect of the combination strategy. The combination framework is based on the constraint that each classifier uses its own representation of the input space. This is achieved by training the set of ....

[Article contains additional citation context not shown here]

J Kittler, M. Hatef, R. P. W Duin, and J. Matas. On combining classifiers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 20(3):226--239, 1998.

....and demonstrate both improved performance and better statistical justification for the integration step. First, we compare the performance of several common data fusion strategies on our task, and develop statistical interpretations of each. Following the theoretical framework presented in [6], we compare MAX, MIN, MEAN, and PRODUCT rules for combining classifier outputs. We assess the underlying assumptions for each model, and empirically evaluate which ones are appropriate for the task of face and gait integration. Second, we explore the effect of early versus late temporal ....

....and pen gesture. In [5] a framework for integrating multiple biometric cues in a large database application was provided, where a portion of the cues were used for retrieval and the remainder for verification, and also explored late vs. early fusion in the context of fingerprint recognition. In [6] a theoretical framework was developed for combining independent classifiers, and different sets of simple assumptions were shown to lead to a range of commonly used combination heuristics. We followed in our experiments. 3. Integrated recognition by face and gait We shall briefly review the ....

[Article contains additional citation context not shown here]

J. Kittler, M. Hatef, R. Duin, and J. Matas. On Combining Classifiers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 20(3):226--239, Mar. 1998.

....approximations for the optimal classi er. Because of the di culty in nding a single optimal classi er, another method for improving accuracy of a classi er is to use an ensemble of classi ers (a set of individually trained classi ers whose decisions are combined in some ways) Recent research [16] [9] has shown that an ensemble is often more accurate than any of the single classi ers in the ensemble. There are lots of appeals for using such method. First of all, most learning algorithms work in a very large hypothesis space, then the training data might not provide su cient information ....

Josef Kittler, Mohamad Hatef, Rober P. W. Duin, and Jiri Matas. On combining classiers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 20(3):226239, 1998.

....classifier by its performance on the training data. When a measure of belief, confidence, certainty or other about the classification is available along with the class label, then a number of di#erent rules for combining these measures have been suggested, like Sum, Min, Max, Prod and Median. [7] is an interesting study of these rules. Stacked Generalization [13] also known as Stacking in the literature, is a method that combines multiple classifiers by learning the way that their output correlates with the true class on an independent set of instances. Chan and Stolfo [2] applied the ....

Josef Kittler, Mohamad Hatef, Robert P. W. Duin, and Jiri Matas. On combining classifiers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 20(3):226--238, March 1998.

....ordered in time. 2.1 Late integration strategies A number of methods are applicable to the situation where multiple observations of the same person s face are accumulated over time. This is an instance of the more general problem of fusion of evidence from multiple measurements. Kittler et al..in [9] present a statistical interpretation of a number of common methods for cross modal fusion, such as the product, maximum , and majority rules, which are also appropriate for late integration over a set of observations from a single modality. For example, it can be shown that under the assumption ....

Josef Kittler, Mohamad Hatef, Robert P.W. Duin, and Jiri Matas. On Combining Classifiers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 20(3):226--239, March 1998.

....label x in 2 . For the minimum rule (x) 0:3; 0:4] for the product rule (x) 0:09; 0:14] and for the average rule (x) 0:47; 0:53] Here all the aggregation rules agree on labeling x in 2 but this need not be the case for a di erent example. The above models are derived in [11] as di erent estimates of P ( i jx) under the assumption of conditionally independent D 1 ; DL . B. Decision templates Using decision templates (DT) for combining classi ers is proposed in [15] Given L (trained) classi ers in D, c decision templates are calculated from the data, one ....

J. Kittler, M. Hatef, R.P.W. Duin, and J. Matas. On combining classiers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 20(3):226-239, 1998.

....Prior performances of individual classifiers can be embedded into the combination function in the form of estimates of correctness probabilities conditioned on the individual decisions. Simpler combination rules that do not take into account the classifiers prior performance were studied in [17], where a justification was given in support of the sum rule that chooses the class maximizing the sum of individual estimates of posterior probabilities. The justification is from the sum rule s relative insensitivity to local estimation errors when compared to the product rule. There are also ....

J. Kittler, M. Hatef, R.P.W. Duin, J. Matas, On combining classifiers, IEEE Transactions on Pattern Analysis and Machine Intelligence, PAMI-20, 3, March 1998, 226-239.

....data dependent combination weights. In classifiers fusion schemes, classifiers outputs are combined to achieve a group decision . The most often used classifiers fusion schemes with the combination weights expressed over the entire data space are: the majority vote [24 ] the probability schemes [5]; the weighted averaging [4, 6 9] the Borda count [10] the Bayes approach [2, 3] fuzzy connectives [11] combination through order statistics [12] combination by a neural network [13] and the fuzzy integral [14 16] Combination with data dependent weights attempts to predict which group of ....

Kittler, J., Hatef, M., Duin, R.P.W., Matas, J., On combining classifiers, IEEE Transaction on Pattern Analysis and Machine Intelligence, vol. 20, NO. 3, 1998: pp. 226 --239.

....these approaches is that the integration is static in the sense that we cannot change the weighting of responses dynamically depending, for example, on their usefulness or the environmental conditions. Combining different classifiers is a standard problem in pattern recognition (see for example [15, 8, 6]) Typically different classifiers are trained individually and their parameters are fixed thereafter. It is the combiner s task to learn and choose the appropriate combination mechanisms depending on particular situations. In that sense only the combiner itself may be able to increase the ....

J. Kittler, M. Hatef, R. Duin, and J. Matas. On combining classifiers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 20(3):226--239, March 1998.

....of many algorithms in order to hope for a better average performance than any of the combined methods. It has already been shown in many research papers that combining biometric veri cation systems enables to achieve better performance than techniques based on only one biometric modality [2, 9, 26, 25]. More speci cally, audio visual biometric veri cations systems (based on the face and the voice of an individual) have also been extensively studied [3, 4] Most classi cation machine learning algorithms can be used for fusion purposes. A good introduction to machine learning algorithms can be ....

J. Kittler, M. Hatef, R.P.W. Duin, and J. Matas. On combining classiers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 20(3):226239, 1998.

....X m=1 pm (kjx) 2) Although Eq. 2) is widely accepted to yield state of the art results in many applications, Kittler assumed in his derivation of the sum rule that the posterior probabilities pm (kjx) computed by the di erent classi ers do not di er much from the prior probabilities p(k) [9]. In other words, the derivation of C C C Combination Output 1 2 M Observation C Combination Observation Virtual Observations Output 1 2 M 1 2 M Fig. 2. Classi er combination (left) vs. the virtual test sample method (right) the sum rule for classi er combination is based on the ....

.... the good performance of the sum rule could possibly be explained by its error tolerance: Using the sum rule, errors in estimating the real (and therefore usually unknown) posterior probabilities are dampened, while for instance in the case of the product rule, these estimation errors are ampli ed [9]. If no set of classi ers exists for combination, techniques like bagging [1] or boosting [13] exist, which generate a variety of classi ers using di erent subsets (bagging) respectively di erently weighted versions of the training data (boosting) for training. Here, it is assumed that the ....

Kittler, J.: On Combining Classiers. IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol. 20, No. 3, pp. 226-239, March 1998.

....combination schemes. 1. INTRODUCTION When multiple independently trained pattern classifiers are combined, the resulting accuracy is often better than any of the individual classifiers. This has been demonstrated for automatic speech recognition (ASR) 7, 10, 18] and for pattern classification [12, 13, 20, 29]. Classifier combination can fuse together different information sources to utilize their complementary information. The sources can be multi modal, such as speech and vision, but can also be transformations [18] or (e.g. spectral) partitions [5, 25, 24] of the same signal. In each case, ....

....operate directly on classifier probabilities. One method (the mean rule) computes a weighted average of classifier outputs. Another method (the product rule) multiplies and then renormalizes these probabilities. Other techniques compute the maximum, minimum, or median of the classifier outputs [20]. Other methodologies [27, 19] jointly train separate classifiers which are combined in various ways. In ASR, classifier combination can occur at different levels including the feature stream [2, 16, 10, 17] the HMM state [18] or at higher levels such as at the syllable [31] or sentence [7] It ....

[Article contains additional citation context not shown here]

J. Kittler, M. Hataf, R.P.W. Duin, and J. Matas. On combining classifiers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 20(3):226-- 239, 1998.

....the test data indicate that no method alone can identify buildings with more than 80 confidence. Structural mapping was as effective as the best of the shape only techniques (scalars) The next stage of the work will be to combine results from each classifier together using the sum decision rule [Kittler 1998] which should give a more reliable classification than using one method in isolation. ....

J. Kittler: On Combining Classifiers, IEEE Transactions on Pattern Analysis and machine Intelligence, 20, 226-238, March 1998.

....(x) have been discussed in [1] a normal distribution with mean p and variance 2 ( varied between 0.1 and 1) and a uniform distribution spanning the interval [p b; p b] b varied from 0. 1 to 1) Simple fusion methods are the most obvious choice when constructing a multiple classi er system [2], 3] 6] 7] 8] i.e. the support for class i , d i (x) yielded by the team is d i (x) F(d 1;i (x) d L;i (x) 1) where F is the chosen fusion method. Here we study the fusion methods compared in [1] except the product, i.e. F stands for minimum maximum average ....

J. Kittler, M. Hatef, R. Duin, and J. Matas. On combining classiers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 20(3):226-239, 1998.

....a considerable extension of previous work on probabilistic active fusion [3, 4] The probabilistic version is compared to a possibilistic and an evidence theoretic approach. Furthermore, probabilistic product fusion is compared to the sum rule which has recently been re advocated by Kittler et al. [15]. The classification module is based upon a modified version of the appearance based object recognition approach suggested by Murase and Nayar [20] This method was chosen because it does not only result in object classification but also gives reasonable pose estimations (a prerequisite for ....

....interpretations. In the Bayesian case these considerations have been confirmed by a recently performed comparative analysis of the error tolerance of probabilistic product fusion and fusion 14 by averaging. Theoretical and experimental evidence has clearly favored the average operation in [15]. We will, therefore, compare the product fusion approach to the averaging scheme in experiments with reduced classification performance and increased outlier rate (e.g. because of segmentation errors) Probabilistic product fusion, possibilistic minimum fusion and Dempster s rule of combination ....

J. Kittler, M. Hatef, R.P.W. Duin, and J. Matas. On Combining Classifiers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 20(3):226--238, 1998.

....interpretations. In the Bayesian case these 14 considerations have been confirmed by a recently performed comparative analysis of the error tolerance of probabilistic product fusion and fusion by averaging. Theoretical and experimental evidence has clearly favored the average operation in [13]. In our experiments, however, the advantage of increased noise resistance of averaging schemes is of little value since, for active recognition attention to the worst results obtained is a necessary ingredient of the fusion scheme. The fact remains, that the choice of the right fusion operator ....

....rule is always associative and commutatitve. In probability theory, both associativity and commutativity are consequences of the assumption of conditional independence which is commonly applied in Bayesian classifier fusion because of the lack of knowledge of the relevant conditional probabilities [13]. In possibility theory (or fuzzy logic) various non associative and or non commutative aggregation operators exist [7] However, apart from very general guidelines [3] the proper choice of one of these operators still remains a matter of trial and error. In this paper we have chosen a standard ....

J. Kittler, M. Hatef, R.P.W. Duin, and J. Matas. On Combining Classifiers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 20(3):226--238, 1998. 23

....have proposed in [22] an acoustic labial speaker verification method. Their approach is based on a lip tracker using visual features, and on a text dependent speech expert. The fused score is computed as the weighted sum of the scores generated by the two experts. Kittler et al. have proposed in [23] a multimodal person verification system, using three experts: frontal face, face profile, and voice. The best combination results are obtained for a simple sum rule. Hong and Jain have proposed in [18] a multi modal personal identification system which integrates two different biometrics (face ....

J. Kittler, M. Hatef, R. P. W. Duin, and J. Matas. On combining classifiers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 20(3):226--239, March 1998.

No context found.

Josef Kittler, Mohamad Hatef, Rober P. W. Duin, and Jiri Matas. On combining classi ers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 20(3):226-239, March 1998.

No context found.

J. Kittler, M. Hatef, R. Duin, and J. Matas. On combining classifiers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 20(3):226--239, 1998.

No context found.

J. Kittler, M. Hatef, R. Duin, and J. Matas. On combining classifiers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 20(3):226--239, 1998.

No context found.

J. Kittler, M. Hatef, R. P. W. Duin and J. Matas. On combining classifiers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 20(3):226--239, 1998.

No context found.

Josef Kittler, Mohamad Hatef, Robert P. W. Duin, and Jiri Matas. On combining classifiers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 20(3):226--238, March 1998.

.... among classifiers prior to combining forms the basis of many strategies, including bagging, arcing, boosting and correlation control [6, 31] Approaches to pooling classifiers can be separated into two main categories: i) simple combiners, e.g. voting [3] Bayesian based weighted product rule [22], or averaging [24, 30] and, ii) meta learners, such as arbitration [7] or stacking [34] The simple combining methods are best suited for problems where the individual classifiers perform the same task, and have comparable success. However, such combiners are more susceptible to outliers and to ....

J. Kittler, M. Hatef, R.P.W. Duin, and J. Matas. On combining classifiers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 20(3):226--239, 1998.

....variable j (classifier) This is hardly ever the case. An example may be found by two classifiers computed for different feature spaces that are entirely unrelated, e.g. based on face images and voices assuming that within a class the feature distributions in the two spaces are independent. See [13, 30]. The rule assumes noise free and reliable confidence estimates. It fails if these estimates may be accidentally zero or very small. 2. the sum rule: 8) This is equivalent to the product rule for small deviations in the classifier outcomes (still assuming independent classifiers) An entirely ....

J. Kittler, M. Hatef, R.P.W. Duin, and J. Matas, On combining classifiers, IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 20, no. 3, 1998, 226-239.

....1. Introduction Let 5506 918 be a set (pool) of classifiers such that , where # 34306 6 , assigns ) a class label , The majority vote method of combining classifier decisions, one of many methods in this important research area [2, 3, 4, 5, 6, 7, 8, 9], is to assign the class label , to ) that is supported by the majority of the classifiers . Finding independent classifiers is one aim of classifier fusion methods for the following reason. Let L be odd, 0 21 , and all classifiers have the same classification accuracy ....

J. Kittler, M. Hatef, R. Duin, and J. Matas. On combining classifiers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 20(3):226--239, 1998.

....in several feature spaces X k , k = 1. R. Each object can be a target object, labeled # T , or an outlier object #O (although during the training of one class classifiers we assume example outlier objects are not available) In each feature space di#erent one class classifiers are trained. In [7] and in [8] a theoretical framework for combining (estimated posterior probabilities from) normal classifiers is developed. For di#erent types of combination rules derivations are obtained. When classifiers are applied on (almost) identical data representations X 1 = X 2 = XR , the ....

J. Kittler, R.P.W. Duin, and J. Matas. On combining classifiers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 20(4):226--239, 1998.

....2. then determine their peak and send the results back to the master. 3.1.4. Algorithm 4: iterative refinement Although Algorithm 3 showed good speedup, it may be argued that the actual times for the uniprocessor program were unacceptably slow. Following the work of Illingworth and Kittler [18] and Wallace [19] a different kind of algorithm was developed to improve performance over those which have been described above. Such an algorithm relies on an iterative technique of refinement based on dynamic parameter range reduction. This approach reduces the amount of redundant computation ....

....done on areas of the accumulator which are likely to possess the highest peaks. The aim of the technique is to lower the amount of work done by starting off the computation at a very coarse accumulator quantization and then gradually focusing in on a few selected parameter ranges at each iteration [11,18]. Very small accumulators are used but since the parameter range is reduced at each iteration, the quantization is quite fine by the final iteration. The accumulator size chosen for this experiment is 4 4 and the maximum number of iterations is also four. The farmer begins the processing by ....

J. Illingworth, J. Kittler, Adaptive Hough transform, IEEE Transactions on Pattern Analysis and Machine Intelligence 9 (5) (1987) 690-- 698.

No context found.

J. Illingworth, J. Kittler, The adaptive Hough transform, IEEE Transactions on Pattern Analysis and Machine Intelligence 9 (1987) 690--698.

....[7, 26, 25] of the same signal. Combining independently trained classifiers often produces appreciable gains, even when individual classifiers exhibit widely varying accuracies. This has been demonstrated for automatic speech recognition (ASR) 11, 12, 21] and for pattern classification [15, 16, 22, 30]. Combination rules often operate directly on classifier probabilities. One method (the mean rule) computes a weighted average of classifier outputs. Another method (the product rule) multiplies and then renormalizes these probabilities. Other techniques compute the maximum, minimum, or median of ....

....operate directly on classifier probabilities. One method (the mean rule) computes a weighted average of classifier outputs. Another method (the product rule) multiplies and then renormalizes these probabilities. Other techniques compute the maximum, minimum, or median of the classifier outputs [22]. Other methodologies combine using statistical models [5] or jointly train separate classifiers [28, 14] In ASR, classifier combination can occur at different levels including the feature stream [3, 19, 12, 20] the HMM state [21] or at higher levels such as at the syllable [32] or sentence ....

[Article contains additional citation context not shown here]

J. Kittler, M. Hataf, R.P.W. Duin, and J. Matas. On combining classifiers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 20(3):226--239, 1998.

No context found.

J. Kittler, M. Hatef, R.P.W. Duin and J. Mates, On combining classifiers, IEEE Transactions on Pattern Analysis and Machine Intelligence 20(3) (1998) 226239.

No context found.

J. Kittler, M. Hatef, Robert P.W. Duin, and J. Matas. On combining classifiers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 20(3):226-- 239, 1998.

No context found.

Josef Kittler, Mohamad Hatef, Rober P. W. Duin, and Jiri Matas. On combining classifiers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 20(3):226--239, March 1998.

No context found.

J. Kittler, M. Hatef, R. P. W. Duin, and J. Matas. On combining classifiers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 20(3):226--239, March 1998.

No context found.

J. Kittler, M. Hatef, and J. Matas. On combining classifiers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 20(3):226--239, March 1998.

No context found.

J. KITTLER,M.HATEF, R.P.W. DUIN,J.MATAS,On Combining Classifiers, IEEE Transactions on Pattern Analysis and Machine Intelligence, 20 (3) (1998), 226--239.

No context found.

Kittler, J., Hatef, M., Duin, R.P.W. & Matas, J. On Combining Classifiers. IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 20(3), pp. 226-239, 1998.

No context found.

J. Kittler, M. Hatef, R.P.W. Duin, and J. Matas, On Combining Classifiers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 20 (3) 226-239, 1998.

No context found.

J. Kittler, M. Hatef, R. P. Duin, and J. Matas. On combining classifiers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 20(3):226--239, March 1998.

No context found.

Josef Kittler, Mohamad Hatef, Robert P.W. Duin, and Jiri Matas. On combining classifiers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 1998.

No context found.

J. Kittler, M. Hatef, R. P. W. Duin, and J. Matas. On combining classifiers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 20(3):226--239, March 1998.

No context found.

J. Kittler, M. Hatef, R. P. Duin, and J. Matas. On combining classifiers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 20:226-- 239, March 1998.

No context found.

J. Kittler, M. Hatef, R. Duin, and J. Matas. On combining classiers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 20(3):226239, 1998.

No context found.

J. Kittler, M. Hatef, R. P. Duin, and J. Matas. On combining classifiers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 20(3):226--239, Mar. 1998.

No context found.

J. Kittler, M. Hatef, and J. Matas. On combining classifiers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 20(3):226--239, March 1998.

No context found.

J. Kittler, M. Hatef, R. Duin, J. Matas, On combining classiers, IEEE Transactions on Pattern Analysis and Machine Intelligence 20 (3) (1998) 226 239.

First 50 documents  Next 50

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC