Local Classifiers Research Articles

Global Multiclass Classification and Dataset Construction via Heterogeneous Local Experts

In the domains of dataset construction and crowdsourcing, a notable challenge is to aggregate labels from a heterogeneous set of labelers, each of whom is potentially an expert in some subset of tasks (and less reliable in others). To reduce costs of hiring human labelers or training automated labeling systems, it is of interest to minimize the number of labelers while ensuring the reliability of the resulting dataset. We model this as the problem of performing $K$-class classification using the predictions of smaller classifiers, each trained on a subset of $[K]$, and derive bounds on the number of classifiers needed to accurately infer the true class of an unlabeled sample under both adversarial and stochastic assumptions. By exploiting a connection to the classical set cover problem, we produce a near-optimal scheme for designing such configurations of classifiers which recovers the well known one-vs.-one classification approach as a special case. Experiments with the MNIST and CIFAR-10 datasets demonstrate the favorable accuracy (compared to a centralized classifier) of our aggregation scheme applied to classifiers trained on subsets of the data. These results suggest a new way to automatically label data or adapt an existing set of local classifiers to larger-scale multiclass problems.

A new fuzzy k-nearest neighbor classifier based on the Bonferroni mean

Abstract We present a new generalized version of the fuzzy k-nearest neighbor (FKNN) classifier that uses local mean vectors and utilizes the Bonferroni mean. We call the proposed new method Bonferroni-mean based fuzzy k-nearest neighbor (BM-FKNN) classifier. The BM-FKNN classifier can be easily fitted for various contexts and applications, because the parametric Bonferroni mean allows for problem-based parameter value fitting. The BM-FKNN classifier can perform well also in situations where clear imbalances in class distributions of data are found. The performance of the proposed classifier is tested with six real-world data sets and with one artificial data set. The results are benchmarked with classification results obtained with the classical k-nearest neighbor-, the local mean-based k-nearest neighbor-, the fuzzy k-nearest neighbor- and other three selected classifiers. In addition to this, an enhancement of the local mean-based k-nearest neighbor classifier by using the Bonferroni means is also proposed and tested. The results show that the proposed new BM-FKNN classifier has the potential to outperform the benchmarks in classification accuracy and confirm the usefulness of using the Bonferroni mean in the learning part of classifiers.

A new globally adaptive k-nearest neighbor classifier based on local mean optimization

The k-nearest neighbor (KNN) rule is a simple and effective nonparametric classification algorithm in pattern classification. However, it suffers from several problems such as sensitivity to outliers and inaccurate classification decision rule. Thus, a local mean-based k-nearest neighbor classifier (LMKNN) was proposed to address these problems, which assigns the query sample with a class label based on the closest local mean vector among all classes. It is proven that the LMKNN classifier achieves better classification performance and is more robust to outliers than the classical KNN classifier. Nonetheless, the unreliable nearest neighbor selection rule and single local mean vector strategy in LMKNN classifier severely have negative effect on its classification performance. Considering these problems in LMKNN, we propose a globally adaptive k-nearest neighbor classifier based on local mean optimization, which utilizes the globally adaptive nearest neighbor selection strategy and the implementation of local mean optimization to obtain more convincing and reliable local mean vectors. The corresponding experimental results conducted on twenty real-world datasets demonstrated that the proposed classifier achieves better classification performance and is less sensitive to the neighborhood size $$k$$ compared with other improved KNN-based classification methods.

Misbehavior-Aware On-Demand Collaborative Intrusion Detection System Using Distributed Ensemble Learning for VANET

Vehicular ad hoc networks (VANETs) play an important role as enabling technology for future cooperative intelligent transportation systems (CITSs). Vehicles in VANETs share real-time information about their movement state, traffic situation, and road conditions. However, VANETs are susceptible to the cyberattacks that create life threatening situations and/or cause road congestion. Intrusion detection systems (IDSs) that rely on the cooperation between vehicles to detect intruders, were the most suggested security solutions for VANET. Unfortunately, existing cooperative IDSs (CIDSs) are vulnerable to the legitimate yet compromised collaborators that share misleading and manipulated information and disrupt the IDSs’ normal operation. As such, this paper proposes a misbehavior-aware on-demand collaborative intrusion detection system (MA-CIDS) based on the concept of distributed ensemble learning. That is, vehicles individually use the random forest algorithm to train local IDS classifiers and share their locally trained classifiers on-demand with the vehicles in their vicinity, which reduces the communication overhead. Once received, the performance of the classifiers is evaluated using the local testing dataset in the receiving vehicle. The evaluation values are used as a trustworthiness factor and used to rank the received classifiers. The classifiers that deviate much from the box-and-whisker plot lower boundary are excluded from the set of the collaborators. Then, each vehicle constructs an ensemble of weighted random forest-based classifiers that encompasses the locally and remotely trained classifiers. The outputs of the classifiers are aggregated using a robust weighted voting scheme. Extensive simulations were conducted utilizing the network security laboratory-knowledge discovery data mining (NSL-KDD) dataset to evaluate the performance of the proposed MA-CIDS model. The obtained results show that MA-CIDS performs better than the other existing models in terms of effectiveness and efficiency for VANET.

Predicting essential genes of 41 prokaryotes by a semi-supervised method

Essential genes are vitally important to the survival and reproduction of organisms. Many machine learning methods have been widely employed to predict essential genes and have obtained satisfactory results. However, most of these methods are supervised methods and may not obtain the desired result when the labeled data are insufficient. In this paper, we proposed a learning with local and global consistency (LGC) method-based classifier, which was employed to predict the essential genes of 41 prokaryotes. LGC is a graph-based semi-supervised learning method that can construct a prediction model using finite label and constraint information. The performance of the proposed classifier was evaluated by employing intra-organism prediction and leave-one-species-out validation. The average AUC value of 41 organisms in intra-organisms prediction was 0.723 when the labeled sample ratio was 0.5. The results of this study indicate that the proposed method can achieve acceptable prediction performance with limited labeled data. Additionally, the results demonstrate that this method has good universality.

Development of a global 30 m impervious surface map using multisource and multitemporal remote sensing datasets with the Google Earth Engine platform

Abstract. The amount of impervious surface is an important indicator in the monitoring of the intensity of human activity and environmental change. The use of remote sensing techniques is the only means of accurately carrying out global mapping of impervious surfaces covering large areas. Optical imagery can capture surface reflectance characteristics, while synthetic-aperture radar (SAR) images can be used to provide information on the structure and dielectric properties of surface materials. In addition, nighttime light (NTL) imagery can detect the intensity of human activity and thus provide important a priori probabilities of the occurrence of impervious surfaces. In this study, we aimed to generate an accurate global impervious surface map at a resolution of 30 m for 2015 by combining Landsat 8 Operational Land Image (OLI) optical images, Sentinel-1 SAR images and Visible Infrared Imaging Radiometer Suite (VIIRS) NTL images based on the Google Earth Engine (GEE) platform. First, the global impervious and nonimpervious training samples were automatically derived by combining the GlobeLand30 land-cover product with VIIRS NTL and MODIS enhanced vegetation index (EVI) imagery. Then, the local adaptive random forest classifiers, allowing for a regional adjustment of the classification parameters to take into account the regional characteristics, were trained and used to generate regional impervious surface maps for each 5∘×5∘ geographical grid using local training samples and multisource and multitemporal imagery. Finally, a global impervious surface map, produced by mosaicking numerous 5∘×5∘ regional maps, was validated by interpretation samples and then compared with five existing impervious products (GlobeLand30, FROM-GLC, NUACI, HBASE and GHSL). The results indicated that the global impervious surface map produced using the proposed multisource, multitemporal random forest classification (MSMT_RF) method was the most accurate of the maps, having an overall accuracy of 95.1 % and kappa coefficient (one of the most commonly used statistics to test interrater reliability; Olofsson et al., 2014) of 0.898 as against 85.6 % and 0.695 for NUACI, 89.6 % and 0.780 for FROM-GLC, 90.3 % and 0.794 for GHSL, 88.4 % and 0.753 for GlobeLand30, and 88.0 % and 0.745 for HBASE using all 15 regional validation data. Therefore, it is concluded that a global 30 m impervious surface map can accurately and efficiently be generated by the proposed MSMT_RF method based on the GEE platform. The global impervious surface map generated in this paper is available at https://doi.org/10.5281/zenodo.3505079 (Zhang and Liu, 2019).

Local Binary Ensemble based Self-training for Semi-supervised Classification of Hyperspectral Remote Sensing Images

Supervised classification of hyperspectral remote sensing images is still challenging due to the scarcity of enough labelled samples. Semi-supervised methods have been adopted to handle this issue. Self-training is a popular semi-supervised technique which is widely used for training a classifier with limited labelled data and a large quantity of unlabeled data. However, traditional self-training approaches often give poor classification results in high dimensional data. In the current work, a novel efficient self-training approach for handling the deficiency of labelled samples for semi-supervised classification of hyperspectral remote sensing images is proposed. The proposed method first trains an ensemble of locally specialized supervised binary classifiers independently by using the dimensionally reduced spectral feature vectors of few available labelled samples. The trained local binary classifiers are then used to extend the labelled set by iterative addition of highly informative unlabeled samples to it by exploiting both the spectral and spatial information of the hyperspectral image. The classifiers are then retrained with the extended dataset in a batchwise manner and the procedure is repeated until adequate quantity of labelled samples are generated. Finally, a supervised multiclass classifier is trained on the extended dataset to produce the final classification map. Experimental results on two benchmark hyperspectral image datasets prove the effectiveness of the proposed method over supervised and traditional self-training based semi-supervised pixelwise classification approach in terms of different classification measures.

Uncertainty-aware integration of local and flat classifiers for food recognition

Food image recognition has recently attracted the attention of many researchers, due to the challenging problem it poses, the ease collection of food images, and its numerous applications to health and leisure. In real applications, it is necessary to analyze and recognize thousands of different foods. For this purpose, we propose a novel prediction scheme based on a class hierarchy that considers local classifiers, in addition to a flat classifier. In order to make a decision about which approach to use, we define different criteria that take into account both the analysis of the Epistemic Uncertainty estimated from the ‘children’ classifiers and the prediction from the ‘parent’ classifier. We evaluate our proposal using three Uncertainty estimation methods, tested on two public food datasets. The results show that the proposed method reduces parent-child error propagation in hierarchical schemes and improves classification results compared to the single flat classifier, meanwhile maintains good performance regardless the Uncertainty estimation method chosen.

Image Copy Detection Based on Local Binary Pattern and SVM Classifier

Abstract Due to the availability of a large number of image editing software, it is very easy to find duplicate copies of original images. In such a situation, there is a need to develop a robust technique that can be used for the identification of duplicate copies apart from differentiating it from different images. In this paper, we have proposed an image hashing technique based on uniform Local Binary Pattern (LBP). Here, the input image is initially pre-processed before calculating the Local Binary Pattern (LBP) which is used for image identification. Experiments show that proposed hashing gives excellent performance against the Histogram equalization attack. The Receiver Operating Curve (ROC) indicates that the proposed hashing also performs better in terms of robustness and discrimination. Support Vector Machine (SVM) classifier shows that generated features can easily classify images into a set of similar and different images, and can classify new data with a high level of accuracy.

Learning neural networks for text classification by exploiting label relations

Text classification has been a fundamental problem in the realm of Natural Language Processing (NLP), and a variety of approaches have been proposed with the development of deep learning. Despite recent progress, most existing approaches deal with the problem of multi-class text classification in a flat way, assuming that text labels are semantically independent. This assumption doesn’t always hold in realistic settings, since there usually exist hierarchical or within-layer dependencies in the latent label space, especially when we consider larger label sets. In this paper, we propose a label clustering algorithm to exploit the underlying structure of label relations, and express the stacked concept relationships in the form of a two-layer label space. Next, we present two different neural network structures to capture inter-layer and intra-layer label relations. The first model HSNN organizes a group of local classifiers in a hierarchical way according to the exploited label space, while the second model LSNN takes advantages of text representations in different granularity levels and the bidirectional inferences with recurrent connections to make predictions. Finally, we evaluate our methods on three benchmark datasets. The results empirically demonstrate that both models are capable of leveraging the exploited label relations to improve text classification performance.

Hierarchical multilabel classifier for gene ontology annotation using multihead and multiend deep CNN model

AbstractGene ontology annotation is known to be a very complicated multilabel classification task, and the hierarchical multilabel classification (HMC) approaches with local classifiers have been shown to be effective for the task. In a traditional HMC method, a set of hierarchically organized simple local classifiers are usually used, each of which for one hierarchical level separately. In this paper, we propose a novel hierarchical multilabel classifier implementing the whole set of hierarchically organized local classifiers in one deep convolution neural network (CNN) model with multiple heads and multiple ends (MHME). The proposed MHME CNN model consists of three parts: the body part of a deep CNN model shared by different local classifiers for feature extraction and feature mapping; the multiend part of a set of autoencoders performing feature fusion transforming the input vectors of different local classifiers to feature vectors with the same length so as to share the feature mapping part; and the multihead part of a set of linear multilabel classifiers. Furthermore, a sophisticated recursive algorithm is designed to train the MHME CNN model to realize the functions of a set of hierarchically organized local classifiers. In this way, by sharing a deep CNN with multiple local classifiers, we are able to construct more powerful local classifiers for each level with limited training samples, and to achieve better classification performance. Experiment results on various benchmark datasets show that the proposed deep CNN‐based model has better performance than the state‐of‐the‐art traditional models. Moreover, it gives rather good performance even under a transfer learning. © 2020 Institute of Electrical Engineers of Japan. Published by John Wiley & Sons, Inc.

Locally lateral manifolds of normalised Gabor features for face recognition

Due to inherent characteristics of multiscale and orientation, normalised Gabor features have been successfully used in face recognition. Various previous works have showcased the strength and feasibility of this approach, especially on its robustness against local variations. However, the projected features are numerous and substantial in dimension, which is largely due to the convolution of multiscale and orientation of wavelets. Such features, when used in practical face recognition, would require relatively lengthy classification process, particularly when it involves computationally extensive local classifier or experts, such as ensembles of local cosine similarity (ELCS) classifier. The authors address this issue by simultaneously reducing the size of Gabor features laterally and locally using a manifold learning method called locally linear embedding (LLE). This method is thus denoted as locally lateral normalised local Gabor feature vector with LLE (LGFV/LN/LLE). Results on several publicly available face datasets reveal the superiority of the authors’ approach in terms of improvements in feature compression of LGFV features by up to a reduction of 95% of total dimensionality while increasing the average classification accuracy by 26%. Altogether, the authors show that their LGFV/LN/LLE augmented by ELCS classifiers delivers equivalent result when compared against the state‐of‐the‐art.

Quantitative Assessment of Upper-Limb Motor Function for Post-Stroke Rehabilitation Based on Motor Synergy Analysis and Multi-Modality Fusion.

Functional assessment is an essential part of rehabilitation protocols after stroke. Conventionally, the assessment process relies heavily on clinical experience and lacks quantitative analysis. In order to objectively quantify the upper-limb motor impairments in patients with post-stroke hemiparesis, this study proposes a novel assessment approach based on motor synergy quantification and multi-modality fusion. Fifteen post-stroke hemiparetic patients and fifteen age-matched healthy persons participated in this study. During different goal-directed tasks, kinematic data and surface electromyography(sEMG) signals were synchronously collected from these participants, and then motor features extracted from each modal data could be fed into the respective local classifiers. In addition, kinematic synergies and muscle synergies were quantified by principal component analysis (PCA) and k weighted angular similarity ( k WAS) algorithm to provide in-depth analysis of the coactivated features responsible for observable movement impairments. By integrating the outputs of local classifiers and the quantification results of motor synergies, ensemble classifiers can be created to generate quantitative assessment for different modalities separately. In order to further exploit the complementarity between the evaluation results at kinematic and muscular levels, a multi-modal fusion scheme was developed to comprehensively analyze the upper-limb motor function and generate a probability-based function score. Under the proposed assessment framework, three types of machine learning methods were employed to search the optimal performance of each classifier. Experimental results demonstrated that the classification accuracy was respectively improved by 4.86% and 2.78% when the analysis of kinematic and muscle synergies was embedded in the assessment system, and could be further enhanced to 96.06% by fusing the characteristics derived from different modalities. Furthermore, the assessment result of multi-modality fusion framework exhibited a significant correlation with the score of standard clinical tests ( R = - 0.87, P = 1.98e - 5 ). These promising results show the feasibility of applying the proposed method to clinical assessments for post-stroke hemiparetic patients.

Optimized training and scalable implementation of Conditional Deep Neural Networks with early exits for Fog-supported IoT applications

The incoming IoT big data era requires efficient and resource-constrained mining of large sets of distributed data. This paper explores a possible approach to this end, combining the two emerging paradigms of Conditional Neural Networks with early exits and Fog Computing. Apart from describing the general framework, we provide four specific contributions. First, after reviewing the basic architectures of CDNNs with early exits and characterizing their computational capacity, we consider three basic algorithms for their supervised training (namely the End-to-End, Layer-Wise and Classifier-Wise training algorithms), and, then, formally characterize and compare the resulting tradeoffs in a Fog-supported implementation. Second, after presenting a reference architecture for the local classifiers equipping the considered CDNNs, we develop an optimized framework for the parallel and distributed setting of their decision thresholds. Third, we propose a greedy algorithm for placing the early exits efficiently on the considered CDNNs and prove its linear scaling complexity. Fourth, we analytically characterize in closed-form and analyze the energy performance of the optimal CDNN-onto-Fog mapping. Finally, extensive numerical tests are presented, in order to test and compare the energy-vs.-implementation complexity-vs.-accuracy performance of the resulting optimized CDNN-over-Fog platforms under the IoT-oriented SVHN and FER-2013 datasets.

Self-propagating video segmentation using patch matching and enhanced Onecut

In this paper, a self-propagating video segmentation approach based on patch matching and enhanced Onecut is proposed, which takes full advantage of the target’s color feature, shape feature, and motion information. Firstly, an interactive key frame segmentation is performed to obtain the accurate initial contour of target. Secondly, some sampling patches are uniformly selected along the target contour of previous frame to initialize the localized classifiers. Afterwards, the patch matching is used to pass this contour to the current frame. Simultaneously, the localized classifiers are moved to current frame similarly, and their corresponding positions and parameters are also updated. Eventually, the foreground and background probability maps of current frame are calculated through the localized classifiers as well as global probability models, and then the enhanced Onecut model is constructed to obtain its segmentation result. Compared with the state-of-the-art video segmentation methods, our proposed approach performs outstandingly on the DAVIS dataset.

Are coalitions needed when classifiers make decisions?

Abstract Cooperation and coalitions’ formation are usually the preferred behavior when conflict situation occurs in real life. The question arises: is this approach should also be used when an ensemble of classifiers makes decisions? In this paper different approaches to classification based on dispersed knowledge are analysed and compared. The first group of approaches does not generate coalitions. Each local classifier generate a classification vector based on the local table, and then one of the most popular fusion methods is used (the sum method or the maximum method). In addition, the approach in which the final classification is made by the strongest classifier is analysed. The second group of approaches uses a coalitions creating method. The final classification is generated based on the coalitions’ predictions by using the two, mentioned above, fusion methods. In addition, the approach is analysed in which the final classification is made by the strongest coalition. For both groups of approaches, with and without coalitions, methods based on the maximum correlation and methods based on the covering rules are considered. The main conclusion that is made in this article is as follows. When classifiers generate fair and rational classification vectors, it is better to consider a coalition-based approach and the fusion method that collectively takes into account all vectors generated by classifiers.

A Face Spoofing Detection Method Based on Domain Adaptation and Lossless Size Adaptation

In this paper, a face spoofing detection method called the Fully Convolutional Network with Domain Adaptation and Lossless Size Adaptation (FCN-DA-LSA) is proposed. As its name suggests, the FCN-DA-LSA includes a lossless size adaptation preprocessor followed by an FCN based pixel-level classifier embedded with a domain adaptation layer. The FCN local classifier makes full use of the basic properties of face spoof distortion namely ubiquitous and repetitive. The domain adaptation (DA) layer improves generalization across different domains. The lossless size adaptation (LSA) preserves the high-frequent spoof clues caused by the face recapture process. The ablation study shows that both DA and the LSA are necessary for high-accuracy face spoofing detection. The FCN-LSA obtains competitive performance among the state-of-the-art methods. With the help of small-sample external data in the target domain (2/50, 2/50, and 1/20 subjects for CASIA-FASD, Replay-Attack, and OULU-NPU respectively), the FCN-DA-LSA further improves the performance and outperforms the existing methods.

Face Spoofing Detection Based on Local Ternary Label Supervision in Fully Convolutional Networks

Face verification systems are prone to spoofing attacks on photos, videos, and 3D masks. Face spoofing detection, i.e., face anti-spoofing, face liveness detection, or face presentation attack detection, is an important task for securing face verification systems in practice and presents many challenges. In this paper, a state-of-the-art face spoofing detection method based on a depth-based Fully Convolutional Network (FCN) is revisited. Different supervision schemes, including global and local label supervisions, are comprehensively investigated. A generic theoretical analysis and associated simulation are provided to demonstrate that local label supervision is more suitable than global label supervision for local tasks with insufficient training samples, such as the face spoofing detection task. Based on the analysis, the Spatial Aggregation of Pixel-level Local Classifiers (SAPLC), which is composed of an FCN part and an aggregation part, is proposed. The FCN part predicts the pixel-level ternary labels, which include the genuine foreground, the spoofed foreground, and the undetermined background. Then, these labels are aggregated together to yield an accurate image-level decision. Furthermore, to quantitatively evaluate the proposed SAPLC, experiments are carried out on the CASIA-FASD, Replay-Attack, OULU-NPU, and SiW datasets. The experiments show that the proposed SAPLC outperforms the representative deep networks, including two globally supervised CNNs, one depth-based FCN, two FCNs with binary labels, and two FCNs with ternary labels, and achieves competitive performances close to some state-of-the-art method performances under various common protocols. Overall, the results empirically verify the advantage of the proposed pixel-level local label supervision scheme.

Merged Local Neighborhood Difference Pattern for Facial Expression Recognition

Facial expression based emotion recognition is one of the popular research domains in the computer vision field. Many machine vision-based feature extraction methods are available to increase the accuracy of the Facial Expression Recognition (FER). In feature extraction, neighboring pixel values are manipulated in different ways to encode the texture information of muscle movements. However, defining the robust feature descriptor is still a challenging task to handle the external factors. This paper introduces the Merged Local Neighborhood Difference Pattern (MLNDP) to encode and merge the two-level of representation. At the first level, each pixel is encoded with respect to center pixel, and at the second level, encoding is carried out based on the relationship with the closest neighboring pixel. Finally, two levels of encodings are logically merged to retain only the texture that is positively encoded from the two levels. Further, the feature dimension is reduced using chi-square statistical test, and the final classification is carried out using multiclass SVM on two datasets namely, CK+ and MMI. The proposed descriptor compared against other local descriptors such as LDP, LTP, LDN, and LGP. Experimental results show that our proposed feature descriptor is outperformed other descriptors with 97.86% on CK+ dataset and 95.29% on MMI dataset. The classifier comparison confirms the results that the combination of MLNDP with multiclass SVM performs better than other combinations in terms of local descriptor and classifier.

Local manifold sparse model for image classification

How to discriminate the types of images is very important for image understanding. A large number of classifiers have been designed for the automatic classification of image. In recent years, sparse representation has widely used in the field of image classification. However, most of the sparse classification methods are based on the sparse reconstruction which ignores the intrinsic structure of data. Therefore, in this paper, we proposed a local manifold sparse classifier (LMSC) on the basis of the sparse manifold assumption and the local structure of data. The proposed method uses the sparse manifold assumption and the local neighbors of data to construct a sparse representation model. Then, we can obtain the sparse coefficients with the proposed sparse model. Finally, we can calculate the probability of the unknown image in each class and assign the class label of the maximization probability to this image. LMSC can reveal the intrinsic structure of data and improve the accuracy of image classification. Experiments on two handwritten digit image data sets (MNIST and Semeion) and a hyperspectral remote sensing image data set (Pavia university) show that the proposed method can achieve better representation and classification accuracy compared to some state-of-the-art classification methods.