Zero-shot learning (ZSL) aims to classify images of unseen classes by learning relationship between visual and semantic features. Existing works have been improving recognition accuracy from various approaches, but they employ computationally intensive algorithms that require iterative optimization. In this work, we revisit the primary approach of the pattern recognition, ı.e., nearest neighbor classifiers, to solve the ZSL task by an extremely simple and fast way, called SimpleZSL. Our algorithm consists of the following three simple techniques: (1) just averaging feature vectors to obtain visual prototypes of seen classes, (2) calculating a pseudo-inverse matrix via singular value decomposition to generate visual features of unseen classes, and (3) inferring unseen classes by a nearest neighbor classifier in which cosine similarity is used to measure distance between feature vectors. Through the experiments on common datasets, the proposed method achieves good recognition accuracy with drastically small computational costs. The execution time of the proposed method on a single CPU is more than 100 times faster than those of the GPU implementations of the existing methods with comparable accuracies.

Encrypted traffic identification is to predict traffic types of encrypted traffic. A deep residual convolution network is proposed for this task. The Softmax classifier is fused with its angular variant, which sets an angular margin to achieve better discrimination. The proposed method improves representation learning and reaches excellent results on the public dataset.

This letter proposes a gradient-enhanced softmax supervisor for face recognition (FR) based on a deep convolutional neural network (DCNN). The proposed supervisor conducts the constant-normalized cosine to obtain the score for each class using a combination of the intra-class score and the soft maximum of the inter-class scores as the objective function. This mitigates the vanishing gradient problem in the conventional softmax classifier. The experiments on the public Labeled Faces in the Wild (LFW) database denote that the proposed supervisor achieves better results when compared with those achieved using the current state-of-the-art softmax-based approaches for FR.

Since the development of food diaries could enable people to develop healthy eating habits, food image recognition is in high demand to reduce the effort in food recording. Previous studies have worked on this challenging domain with datasets having fixed numbers of samples and classes. However, in the real-world setting, it is impossible to include all of the foods in the database because the number of classes of foods is large and increases continually. In addition to that, inter-class similarity and intra-class diversity also bring difficulties to the recognition. In this paper, we solve these problems by using deep convolutional neural network features to build a personalized classifier which incrementally learns the user's data and adapts to the user's eating habit. As a result, we achieved the state-of-the-art accuracy of food image recognition by the personalization of 300 food records per user.

Cell phones with GPS function as well as GPS loggers are widely used and we can easily obtain users' geographic information. Now classifying the measured GPS positions into indoor/outdoor positions is one of the major challenges. In this letter, we propose a robust indoor/outdoor detection method based on sparse GPS measured positions utilizing machine learning. Given a set of clusters of measured positions whose center position shows the user's estimated stayed position, we calculate the feature values composed of: positioning accuracy, spatial features, and temporal feature of measured positions included in every cluster. Then a random forest classifier learns these feature values of the known data set. Finally, we classify the unknown clusters of measured positions into indoor/outdoor clusters using the learned random forest classifier. The experiments demonstrate that our proposed method realizes the maximum F1 measure of 1.000, which classifies measured positions into indoor/outdoor ones with almost no errors.

Intrusion detection system (IDS) is a device or software to monitor a network system for malicious activity. In terms of detection results, there could be two types of false, namely, the false positive (FP) which incorrectly detects normal traffic as abnormal, and the false negative (FN) which incorrectly judges malicious traffic as normal. To protect the network system, we expect that FN should be minimized as low as possible. However, since there is a trade-off between FP and FN when IDS detects malicious traffic, it is difficult to reduce the both metrics simultaneously. In this paper, we propose a sequential classifiers combination method to reduce the effect of the trade-off. The single classifier suffers a high FN rate in general, therefore additional classifiers are sequentially combined in order to detect more positives (reduce more FN). Since each classifier can reduce FN and does not generate much FP in our approach, we can achieve a reduction of FN at the final output. In evaluations, we use NSL-KDD dataset, which is an updated version of KDD Cup'99 dataset. WEKA is utilized as a classification tool in experiment, and the results show that the proposed approach can reduce FN while improving the sensitivity and accuracy.

As the data size of Web-related multi-label classification problems continues to increase, the label space has also grown extremely large. For example, the number of labels appearing in Web page tagging and E-commerce recommendation tasks reaches hundreds of thousands or even millions. In this paper, we propose a graph partitioning tree (GPT), which is a novel approach for extreme multi-label learning. At an internal node of the tree, the GPT learns a linear separator to partition a feature space, considering approximate k-nearest neighbor graph of the label vectors. We also developed a simple sequential optimization procedure for learning the linear binary classifiers. Extensive experiments on large-scale real-world data sets showed that our method achieves better prediction accuracy than state-of-the-art tree-based methods, while maintaining fast prediction.

This paper discusses the effect of pre-grouping on vector classification based on the self-organizing map (SOM). The SOM is an unsupervised learning neural network, and is used to form clusters of vectors using its topology preserving nature. The use of SOMs for practical applications, however, may pose difficulties in achieving high recognition accuracy. For example, in image recognition, the accuracy is degraded due to the variation of lighting conditions. This paper considers the effect of pre-grouping of feature vectors on such types of applications. The proposed pre-grouping functionality is also based on the SOM and introduced into a new parallel configuration of the previously proposed SOM-Hebb classifers. The overall system is implemented and applied to position identification from images obtained in indoor and outdoor settings. The system first performs the grouping of images according to the rough representation of the brightness profile of images, and then assigns each SOM-Hebb classifier in the parallel configuration to one of the groups. Recognition parameters of each classifier are tuned for the vectors belonging to its group. Comparison between the recognition systems with and without the grouping shows that the grouping can improve recognition accuracy.

Word sense disambiguation (WSD) is to identify the right sense of ambiguous words via mining their context information. Previous studies show that classifier combination is an effective approach to enhance the performance of WSD. In this paper, we systematically review state-of-the-art methods for classifier combination based WSD, including probability-based and voting-based approaches. Furthermore, a new classifier combination based WSD, namely the probability weighted voting method with dynamic self-adaptation, is proposed in this paper. Compared with existing approaches, the new method can take into consideration both the differences of classifiers and ambiguous instances. Exhaustive experiments are performed on a real-world dataset, the results show the superiority of our method over state-of-the-art methods.

Anomaly detection is one approach in intrusion detection systems (IDSs) which aims at capturing any deviation from the profiles of normal network activities. However, it suffers from high false alarm rate since it has impediment to distinguish the boundaries between normal and attack profiles. In this paper, we propose an effective anomaly detection approach by hybridizing three techniques, i.e. particle swarm optimization (PSO), ant colony optimization (ACO), and genetic algorithm (GA) for feature selection and ensemble of four tree-based classifiers, i.e. random forest (RF), naive bayes tree (NBT), logistic model trees (LMT), and reduces error pruning tree (REPT) for classification. Proposed approach is implemented on NSL-KDD dataset and from the experimental result, it significantly outperforms the existing methods in terms of accuracy and false alarm rate.

Self-paced e-learning provides much more freedom in time and locale than traditional education as well as diversity of learning contents and learning media and tools. However, its limitations must not be ignored. Lack of information on learners' states is a serious issue that can lead to severe problems, such as low learning efficiency, motivation loss, and even dropping out of e-learning. We have designed a novel e-learning support system that can visually observe learners' non-verbal behaviors and estimate their learning states and that can be easily integrated into practical e-learning environments. Three pairs of internal states closely related to learning performance, concentration-distraction, difficulty-ease, and interest-boredom, were selected as targets of recognition. In addition, we investigated the practical problem of estimating the learning states of a new learner whose characteristics are not known in advance. Experimental results show the potential of our system.

This paper proposes a deep quasi-linear kernel for support vector machines (SVMs). The deep quasi-linear kernel can be constructed by using a pre-trained deep neural network. To realize this goal, a multilayer gated bilinear classifier is first designed to mimic the functionality of the pre-trained deep neural network, by generating the gate control signals using the deep neural network. Then, a deep quasi-linear kernel is derived by applying an SVM formulation to the multilayer gated bilinear classifier. In this way, we are able to further implicitly optimize the parameters of the multilayer gated bilinear classifier, which are a set of duplicate but independent parameters of the pre-trained deep neural network, by using an SVM optimization. Experimental results on different data sets show that SVMs with the proposed deep quasi-linear kernel have an ability to take advantage of the pre-trained deep neural networks and outperform SVMs with RBF kernels.

As the number of surveillance cameras keeps increasing, the demand for automated traffic-monitoring systems is growing. In this paper, we propose a practical vehicle detection method for such systems. In the last decade, vehicle detection mainly has been performed by employing an image scan strategy based on sliding windows whereby a pre-trained appearance model is applied to all image areas. In this approach, because the appearance models are built from vehicle sample images, the normalization of the scales and aspect ratios of samples can significantly influence the performance of vehicle detection. Thus, to successfully apply sliding window schemes to detection, it is crucial to select the normalization sizes very carefully in a wise manner. To address this, we present a novel vehicle detection technique. In contrast to conventional methods that determine the normalization sizes without considering given scene conditions, our technique first learns local region-specific size models based on scene-contextual clues, and then utilizes the obtained size models to normalize samples to construct more elaborate appearance models, namely local size-specific classifiers (LSCs). LSCs can provide advantages in terms of both accuracy and operational speed because they ignore unnecessary information on vehicles that are observable in faraway areas from each sliding window position. We conduct experiments on real highway traffic videos, and demonstrate that the proposed method achieves a 16% increased detection accuracy with at least 3 times faster operational speed compared with the state-of-the-art technique.

In this paper we propose a novel classification method for the multiple k-nearest neighbor (MkNN) classifier and show its practical application to medical image processing. The proposed method performs fine classification when a pair of the spatial coordinate of the observation data in the observation space and its corresponding feature vector in the feature space is provided. The proposed MkNN classifier uses the continuity of the distribution of features of the same class not only in the feature space but also in the observation space. In order to validate the performance of the present method, it is applied to the tissue characterization problem of coronary plaque. The quantitative and qualitative validity of the proposed MkNN classifier have been confirmed by actual experiments.

Nearest-neighbor-search classifiers are attractive but they have high intrinsic computational demands which limit their practical application. In this paper, we propose a coprocessor for k (k with k≥1) nearest neighbor (kNN) classification in which squared Euclidean distances (SEDs) are mapped into the clock domain for realizing high search speed and energy efficiency. The minimal SED searching is carried out by weighted frequency dividers that drastically reduce the normally exponential increase of the worst-case search-clock number with the bit width of vector components to only a linear increase. This also results in low power dissipation and high area-efficiency in comparison to the traditional method using large numbers of adders and comparators. The kNN classifier determines the class of an unknown input sample with a majority decision among the k nearest reference samples. The required majority-decision circuit is integrated with the clock-mapping-based minimal-SED searching architecture and proceeds with the classification immediately after identification of each of the k nearest references. A test chip in 180 nm CMOS technology, which can process 8 dimensions of 32 reference vectors in parallel, achieves low power dissipation of 40.32 mW (at 51.21 MHz clock frequency and 1.8 V supply voltage). Significantly, the distance search circuit consumes only 5.99 mW. Feature vectors with different dimensionality up to 2048 dimensions can be handled by the designed coprocessor due to a dimension extension circuit, enabling large flexibility for usage in different application.

In this paper, a one-class Naïve Bayesian classifier (One-NB) for detecting toll frauds in a VoIP service is proposed. Since toll frauds occur irregularly and their patterns are too diverse to be generalized as one class, conventional binary-class classification is not effective for toll fraud detection. In addition, conventional novelty detection algorithms have struggled with optimizing their parameters to achieve a stable detection performance. In order to resolve the above limitations, the original Naïve Bayesian classifier is modified to handle the novelty detection problem. In addition, a genetic algorithm (GA) is employed to increase efficiency by selecting significant variables. In order to verify the performance of One-NB, comparative experiments using five well-known novelty detectors and three binary classifiers are conducted over real call data records (CDRs) provided by a Korean VoIP service company. The experimental results show that One-NB detects toll frauds more accurately than other novelty detectors and binary classifiers when the toll frauds rates are relatively low. In addition, The performance of One-NB is found to be more stable than the benchmark methods since no parameter optimization is required for One-NB.

Multi-label classification allows a sample to belong to multiple classes simultaneously, which is often the case in real-world applications such as text categorization and image annotation. In multi-label scenarios, taking into account correlations among multiple labels can boost the classification accuracy. However, this makes classifier training more challenging because handling multiple labels induces a high-dimensional optimization problem. In this paper, we propose a scalable multi-label method based on the least-squares probabilistic classifier. Through experiments, we show the usefulness of our proposed method.

Aiming to alleviate the visual tracking problem of drift which reduces the abilities of almost all online visual trackers, a robust visual tracker (called CCMM tracker) is proposed with a coupled-classifier based on multiple representative appearance models. The coupled-classifier consists of root and head classifiers based on local sparse representation. The two classifiers collaborate to fulfil a tracking task within the Bayesian-based tracking framework, also to update their templates with a novel mechanism which tries to guarantee an update operation along the “right” orientation. Consequently, the tracker is more powerful in anti-interference. Meanwhile the multiple representative appearance models maintain features of the different submanifolds of the target appearance, which the target exhibited previously. The multiple models cooperatively support the coupled-classifier to recognize the target in challenging cases (such as persistent disturbance, vast change of appearance, and recovery from occlusion) with an effective strategy. The novel tracker proposed in this paper, by explicit inference, can reduce drift and handle frequent and drastic appearance variation of the target with cluttered background, which is demonstrated by the extensive experiments.

This paper proposes a novel method for object tracking by combining local feature and global template-based methods. The proposed algorithm consists of two stages from coarse to fine. The first stage applies on-line classifiers to match the corresponding keypoints between the input frame and the reference frame. Thus a rough motion parameter can be estimated using RANSAC. The second stage employs kernel-based global representation in successive frames to refine the motion parameter. In addition, we use the kernel weight obtained during the second stage to guide the on-line learning process of the keypoints' description. Experimental results demonstrate the effectiveness of the proposed technique.

The objective of pool-based incremental active learning is to choose a sample to label from a pool of unlabeled samples in an incremental manner so that the generalization error is minimized. In this scenario, the generalization error often hits a minimum in the middle of the incremental active learning procedure and then it starts to increase. In this paper, we address the problem of early labeling stopping in probabilistic classification for minimizing the generalization error and the labeling cost. Among several possible strategies, we propose to stop labeling when the empirical class-posterior approximation error is maximized. Experiments on benchmark datasets demonstrate the usefulness of the proposed strategy.