Nearest-neighbor Classifier Research Articles

IMPROVED NOISE-RESILIENT ISOLATED WORDS SPEECH RECOGNITION USING PIECEWISE DIFFERENTIATION

Speech is the primary method of communication among humans; it is the main form of communication to deliver emotions and thoughts. Intuitively, humans would find it convenient to communicate with machines through speech. The automatic speech recognition model’s primary goal is to transcribe or identify the word sequence represented by the acoustic signals or speech. In this advanced era of digital media and e-learning platforms, the effective use of voice recognition technology to deliver a more individualized and effective educational experience in e-learning initiatives. It helps the students to improve their oral pronunciation skills. Even though Arabic is one of the most spoken languages, the research works related to Arabic speech and text are lacking when compared to other languages. Here, we propose a hybrid model of [Formula: see text]-Nearest Neighbor (KNN) classifier and Dynamic Time Warping (DTW) for implementing a noise-resilient speech recognition system for isolated words in the Arabic language. The model is implemented with Mel-Frequency Cepstral Coefficients (MFCC) and its piecewise first and second derivatives as feature representation models. The proposed model is implemented in Python and simulated using Arabic Speech Corpus for Isolated Words [A. Alalshekmubarak and L. Smith, On improving the classification capability of reservoir computing for Arabic speech recognition, in International Conference on Artificial Neural Networks (Springer, Cham, 2014), pp. 225–232]. The proposed model is implemented with (1) MFCC alone for feature representation, (2) combination of MFCC with Delta coefficients, and (3) combination of MFCC with Delta and Delta–Delta coefficients. The implemented model is evaluated using different test sets of varying sizes of 100, 200, 500, 1000, 1500, and 2000. Evaluation is performed for both noised and noiseless speech using these three feature representation models and performed an evaluation of these three models. For evaluating this model’s performance in noised conditions, both white and babble noises of various signal-to-noise ratio values such as 10 dB, 20 dB, and 30 dB are added into the noiseless speech and estimate the classification accuracy in those situations. The proposed KNN–DTW model with MFCC outperformed all the other models in the literature. MFCC with Delta and Delta–Delta coefficients is effective than other two models using MFCC and MFCC with Delta coefficients. The proposed model could be used to recognize the isolated word recognition of grade 1 textbook vocabulary in classroom noisy environment for providing a better interactive classroom environment to Kuwaiti elementary students to enhance their math learning level. In an interactive computer-based education environment for Mathematics learning, the proposed model could be recognized as the spoken utterances of students.

Plant recognition by AI: Deep neural nets, transformers, and kNN in deep embeddings.

The article reviews and benchmarks machine learning methods for automatic image-based plant species recognition and proposes a novel retrieval-based method for recognition by nearest neighbor classification in a deep embedding space. The image retrieval method relies on a model trained via the Recall@k surrogate loss. State-of-the-art approaches to image classification, based on Convolutional Neural Networks (CNN) and Vision Transformers (ViT), are benchmarked and compared with the proposed image retrieval-based method. The impact of performance-enhancing techniques, e.g., class prior adaptation, image augmentations, learning rate scheduling, and loss functions, is studied. The evaluation is carried out on the PlantCLEF 2017, the ExpertLifeCLEF 2018, and the iNaturalist 2018 Datasets—the largest publicly available datasets for plant recognition. The evaluation of CNN and ViT classifiers shows a gradual improvement in classification accuracy. The current state-of-the-art Vision Transformer model, ViT-Large/16, achieves 91.15% and 83.54% accuracy on the PlantCLEF 2017 and ExpertLifeCLEF 2018 test sets, respectively; the best CNN model (ResNeSt-269e) error rate dropped by 22.91% and 28.34%. Apart from that, additional tricks increased the performance for the ViT-Base/32 by 3.72% on ExpertLifeCLEF 2018 and by 4.67% on PlantCLEF 2017. The retrieval approach achieved superior performance in all measured scenarios with accuracy margins of 0.28%, 4.13%, and 10.25% on ExpertLifeCLEF 2018, PlantCLEF 2017, and iNat2018–Plantae, respectively.

Convolutional Neural Networks-Based Framework for Early Identification of Dementia Using MRI of Brain Asymmetry.

Computer-aided diagnosis of health problems and pathological conditions has become a substantial part of medical, biomedical, and computer science research. This paper focuses on the diagnosis of early and progressive dementia, building on the potential of deep learning (DL) models. The proposed computational framework exploits a magnetic resonance imaging (MRI) brain asymmetry biomarker, which has been associated with early dementia, and employs DL architectures for MRI image classification. Identification of early dementia is accomplished by an eight-layered convolutional neural network (CNN) as well as transfer learning of pretrained CNNs from ImageNet. Different instantiations of the proposed CNN architecture are tested. These are equipped with Softmax, support vector machine (SVM), linear discriminant (LD), or [Formula: see text] -nearest neighbor (KNN) classification layers, assembled as a separate classification module, which are attached to the core CNN architecture. The initial imaging data were obtained from the MRI directory of the Alzheimer's disease neuroimaging initiative 3 (ADNI3) database. The independent testing dataset was created using image preprocessing and segmentation algorithms applied to unseen patients' imaging data. The proposed approach demonstrates a 90.12% accuracy in distinguishing patients who are cognitively normal subjects from those who have Alzheimer's disease (AD), and an 86.40% accuracy in detecting early mild cognitive impairment (EMCI).

Gömülü ve Sarıcı Öznitelik Seçim Yöntemleri Kullanılarak Akciğer Rahatsızlıklarının Tespiti

Son yıllarda biyomedikal sinyal işleme alanındaki gelişmelere rağmen, akciğer rahatsızlıklarının tespiti üzerine hızlı ve yüksek doğrulukta çalışan teşhis sistemlerine duyulan ihtiyaç artmaktadır. Yapılan çalışmada fiziki muayene ile 94 farklı kişiden, solunum döngülerinin otomatik olarak tespit edilmesiyle elde edilen 150 adet normal ve 444 adet normal olmayan akciğer sesleri veri tabanı olarak kullanılmıştır. Sınıflandırma işleminde öznitelik olarak frekans ve zaman bölgesinde 12 farklı yöntem uygulanmıştır. Tüm veriler %80 eğitim %20 test aşamasında kullanılacak şekilde ikiye bölünmüştür. Elde edilen öznitelikler gömülü ve sarıcı öznitelik seçim yöntemleri kullanılarak değerlendirilmiştir. Bu yöntemler; özyinelemeli öznitelik eliminasyonu, uyarlanabilir yapı öğrenimi ile öznitelik seçimi, bağımlılık kılavuzlu denetimsiz öznitelik seçimi, sıralı yerellik ile denetimsiz öznitelik seçimi, içbükey küçültme yoluyla öznitelik seçimi, en küçük mutlak büzülme ve seçim operatörü öznitelik seçim yöntemleri olarak isimlendirilmektedir. İncelenen bu öznitelikler doğrusal destek vektör makineleri, k en yakın komşuluk, karar ağaçları ve naive bayes yöntemleri ile sınıflandırılmıştır. Sonuç olarak öznitelik sayısının sınırlandırılmadığı durum için, özyinelemeli öznitelik eliminasyonu yönteminin k en yakın komşuluk sınıflandırma ile beraber kullanıldığı durum için %97,3 doğruluk değerindeki başarıma ulaşılmaktadır. Öznitelik sayısının üç ile sınırlandırıldığı durumda ise uyarlanabilir yapı öğrenimi ile öznitelik seçimi yönteminin karar ağaçları yöntemi ile beraber kullanılması ile %91,4 değerinde başarıma ulaşılmıştır.

Machine learning approach for classification of mangifera indica leaves using digital image analysis

ABSTRACT There is a wide range of horticulture farming in Asia. Mangifera Indica belongs to the species of flowering plant, also publicly recognized as mango. It has a significant local demand as well as a broad export marketplace throughout the world, and is considered as ‘King of Fruits.’ There are many mango varieties and each has its own business market. Efficient identification of the mango varieties is still difficult because of untrained growers and obsolete farming culture, especially in remote areas of the Asia. The primary purpose of this research study was to discriminate mango varieties with the potential of machine learning techniques by analyzing their leaves. For the purpose, we selected leaves of eight mango varieties, namely: Anwar-Ratul (AR), Chaunsa (CHAUN), Langra (LANG), Sindhri (SIND), Saroli (SARO), Fajri (FAJ), Desi (DESI), Alo-Marghan (ALM). A digital cell phone camera captured these datasets in open atmosphere without any well-equipped lab and infrastructure. Binary, histogram, RST, spectral, and texture features were employed for machine learning (ML)-based mango leaf image discrimination. A k-fold (k = 10) cross-validation method was used for ML classification. The k nearest neighbors (KNN) classifier achieved maximum overall classification accuracy (OCA) from 88.33% to 97%.

Analysis and Classification of Muscular Paralysis Disease using Electromyography Signal with Machine Learning

Electrical activity of the muscles is characterized by Electromyography (EMG) signals. The EMG signal analysis form the basis for the diagnosis of muscular paralysis. The EMG signals from amyotrophic lateral sclerosis (ALS), and Myopathy are considered to analyze the paralysis. The statistical analysis of EMG signals from ALS, Myopathy, and Normal conditions, aid in the analysis and classification of paralysis. This work intends to analyze and classify the paralysis and normal conditions using EMG features extracted in time and frequency domains. Twelve statistical features are extracted from the EMG signals considered. Machine Learning techniques and deep learning techniques (DLT) are employed to perform the classification. multi-layer perceptron (MLP), support vector machine (SVM), random forest (RF), gradient boosting (GB), and nearest neighbor (NN) classifier models are used for the classification. The accuracy of the classifiers is calculated. The accuracy values obtained are 72% for MLP, 73% for SVM, 72% for RF, 71% for GB, and 69% for NN. The performance accuracy is better in SVM compared to other classifiers.

Bearing fault diagnosis using a novel coding-statistic feature combined with NNC

The failures of rolling bearings usually cause the breakdown of rotating machinery. Therefore, bearing fault diagnosis is receiving more and more attentions. In this paper, a new coding-statistic feature is proposed for bearing fault diagnosis. Firstly, a waveform coding matrix (WCM) is drawn from each signal using a coding algorithm then a statistical feature is extracted from the WCM with a pre-defined dictionary. Secondly, all statistical features are processed using two-dimensional principal component analysis (2DPCA) to reduce redundant information and dimensionality. Finally, a nearest neighbor classifier (NNC) is employed to classify the bearing faults. Two bearing fault classification problems are utilized to demonstrate the effectiveness of the proposed scheme. Experimental results show that an excellent performance could be accomplished with the proposed scheme.

An Ensemble and Iterative Recovery Strategy Based kGNN Method to Edit Data with Label Noise

Learning label noise is gaining increasing attention from a variety of disciplines, particularly in supervised machine learning for classification tasks. The k nearest neighbors (kNN) classifier is often used as a natural way to edit the training sets due to its sensitivity to label noise. However, the kNN-based editor may remove too many instances if not designed to take care of the label noise. In addition, the one-sided nearest neighbor (NN) rule is unconvincing, as it just considers the nearest neighbors from the perspective of the query sample. In this paper, we propose an ensemble and iterative recovery strategy-based kGNN method (EIRS-kGNN) to edit data with label noise. EIRS-kGNN first uses the general nearest neighbors (GNN) to expand the one-sided NN rule to a binary-sided NN rule, taking the neighborhood of the queried samples into account. Then, it ensembles the prediction results of a finite set of ks in the kGNN to prudently judge the noise levels for each sample. Finally, two loops, i.e., the inner loop and the outer loop, are leveraged to iteratively detect label noise. A frequency indicator is derived from the iterative processes to guide the mixture approaches, including relabeling and removing, to deal with the detected label noise. The goal of EIRS-kGNN is to recover the distribution of the data set as if it were not corrupted. Experimental results on both synthetic data sets and UCI benchmarks, including binary data sets and multi-class data sets, demonstrate the effectiveness of the proposed EIRS-kGNN method.

On the Asymptotic Evaluation of the Finite-Sample Risk of the Nearest Neighbor Classifier

On the Asymptotic Evaluation of the Finite-Sample Risk of the Nearest Neighbor Classifier

Breast Cancer Diagnosis with Machine Learning Techniques

Cancer deaths are one of the highest rates of death. Although breast cancer is commonly associated with women, it is sometimes seen in men, and the mortality rate for men with breast cancer may be higher. The importance of early detection and treatment of breast cancer cannot be overstated. Cancer is diagnosed at an early stage thanks to expert systems, artificial intelligence, and machine learning approaches, and data analysis makes life easier for healthcare professionals. The nearest neighbor method, principal component analysis, neighborhood component method approaches were employed to detect breast cancer in this study. "Breast Cancer Wisconsin Diagnostic" database was used to create and test the approach. According to the results obtained, the highest success rate with 99.42% was obtained by using neighborhood component analysis and nearest neighbor classification algorithm method.

Federated Nearest Neighbor Classification with a Colony of Fruit-Flies

The mathematical formalization of a neurological mechanism in the fruit-fly olfactory circuit as a locality sensitive hash (Flyhash) and bloom filter (FBF) has been recently proposed and "reprogrammed" for various learning tasks such as similarity search, outlier detection and text embeddings. We propose a novel reprogramming of this hash and bloom filter to emulate the canonical nearest neighbor classifier (NNC) in the challenging Federated Learning (FL) setup where training and test data are spread across parties and no data can leave their respective parties. Specifically, we utilize Flyhash and FBF to create the FlyNN classifier, and theoretically establish conditions where FlyNN matches NNC. We show how FlyNN is trained exactly in a FL setup with low communication overhead to produce FlyNNFL, and how it can be differentially private. Empirically, we demonstrate that (i) FlyNN matches NNC accuracy across 70 OpenML datasets, (ii) FlyNNFL training is highly scalable with low communication overhead, providing up to 8x speedup with 16 parties.

MetaNODE: Prototype Optimization as a Neural ODE for Few-Shot Learning

Few-Shot Learning (FSL) is a challenging task, i.e., how to recognize novel classes with few examples? Pre-training based methods effectively tackle the problem by pre-training a feature extractor and then predicting novel classes via a cosine nearest neighbor classifier with mean-based prototypes. Nevertheless, due to the data scarcity, the mean-based prototypes are usually biased. In this paper, we attempt to diminish the prototype bias by regarding it as a prototype optimization problem. To this end, we propose a novel meta-learning based prototype optimization framework to rectify prototypes, i.e., introducing a meta-optimizer to optimize prototypes. Although the existing meta-optimizers can also be adapted to our framework, they all overlook a crucial gradient bias issue, i.e., the mean-based gradient estimation is also biased on sparse data. To address the issue, we regard the gradient and its flow as meta-knowledge and then propose a novel Neural Ordinary Differential Equation (ODE)-based meta-optimizer to polish prototypes, called MetaNODE. In this meta-optimizer, we first view the mean-based prototypes as initial prototypes, and then model the process of prototype optimization as continuous-time dynamics specified by a Neural ODE. A gradient flow inference network is carefully designed to learn to estimate the continuous gradient flow for prototype dynamics. Finally, the optimal prototypes can be obtained by solving the Neural ODE. Extensive experiments on miniImagenet, tieredImagenet, and CUB-200-2011 show the effectiveness of our method.

Sensitivity of Sparse Codes to Image Distortions.

Sparse coding has been proposed as a theory of visual cortex and as an unsupervised algorithm for learning representations. We show empirically with the MNIST data set that sparse codes can be very sensitive to image distortions, a behavior that may hinder invariant object recognition. A locally linear analysis suggests that the sensitivity is due to the existence of linear combinations of active dictionary elements with high cancellation. A nearest-neighbor classifier is shown to perform worse on sparse codes than original images. For a linear classifier with a sufficiently large number of labeled examples, sparse codes are shown to yield higher accuracy than original images, but no higher than a representation computed by a random feedforward net. Sensitivity to distortions seems to be a basic property of sparse codes, and one should be aware of this property when applying sparse codes to invariant object recognition.

Eyes versus Eyebrows: A Comprehensive Evaluation Using the Multiscale Analysis and Curvature-Based Combination Methods in Partial Face Recognition

This work aimed to find the most discriminative facial regions between the eyes and eyebrows for periocular biometric features in a partial face recognition system. We propose multiscale analysis methods combined with curvature-based methods. The goal of this combination was to capture the details of these features at finer scales and offer them in-depth characteristics using curvature. The eye and eyebrow images cropped from four face 2D image datasets were evaluated. The recognition performance was calculated using the nearest neighbor and support vector machine classifiers. Our proposed method successfully produced richer details in finer scales, yielding high recognition performance. The highest accuracy results were 76.04% and 98.61% for the limited dataset and 96.88% and 93.22% for the larger dataset for the eye and eyebrow images, respectively. Moreover, we compared the results between our proposed methods and other works, and we achieved similar high accuracy results using only eye and eyebrow images.

Identifying Alcohol Use Disorder With Resting State Functional Magnetic Resonance Imaging Data: A Comparison Among Machine Learning Classifiers.

Alcohol use disorder (AUD) is a burden to society creating social and health problems. Detection of AUD and its effects on the brain are difficult to assess. This problem is enhanced by the comorbid use of other substances such as nicotine that has been present in previous studies. Recent machine learning algorithms have raised the attention of researchers as a useful tool in studying and detecting AUD. This work uses AUD and controls samples free of any other substance use to assess the performance of a set of commonly used machine learning classifiers detecting AUD from resting state functional network connectivity (rsFNC) derived from independent component analysis. The cohort used included 51 alcohol dependent subjects and 51 control subjects. Despite alcohol, none of the 102 subjects reported use of nicotine, cannabis or any other dependence or habit formation substance. Classification features consisted of whole brain rsFNC estimates undergoing a feature selection process using a random forest approach. Features were then fed to 10 different machine learning classifiers to be evaluated based on their classification performance. A neural network classifier showed the highest performance with an area under the curve (AUC) of 0.79. Other good performers with similar AUC scores were logistic regression, nearest neighbor, and support vector machine classifiers. The worst results were obtained with Gaussian process and quadratic discriminant analysis. The feature selection outcome pointed to functional connections between visual, sensorimotor, executive control, reward, and salience networks as the most relevant for classification. We conclude that AUD can be identified using machine learning classifiers in the absence of nicotine comorbidity.

Analysis of Muscular Paralysis using EMG Signal with Wavelet Decomposition Approach

Paralysis refers to temporary or permanent loss of voluntary muscle movement in a body part or region. The degree of muscle function loss determines the severity of paralysis. The muscle function is represented by electrical activity of the muscles. Electromyography is a technique concerned with the analysis of myoelectric signals. EMG allows the determination of muscular activity. EMG signal analysis is performed using the features extracted in time domain, frequency domain and time frequency domain. In this work, the EMG of Amyotrophic Lateral Sclerosis (ALS), Myopathy, and Normal conditions are considered, and the time frequency analysis has been carried out to extract the features using wavelet decomposition approach. The classification of normal and paralyzed condition is carried by four classifier models. The classifier models used are Multi-layer Perceptron (MLP), Random Forest (RF), Gradient Boosting (GB), and Nearest Neighbor (NN) models. The standard data set has been used for the purpose. The classification accuracy obtained for MLP is 80%, for RF is 75%, for GB is 79%, and for NN is 69%. MLP show better classification performance over RF, GB, and NN Classifiers.

CARD9: Cardiac Contractility Estimation Based on Left Ventricular Pressure: Comparison of Time Series Classifiers and Graph-based Neural Networks

Background: In heart failure therapy, the real-time estimation of remaining cardiac contractility (RCC) can facilitate patient management, as well as, the performance of physiological controllers for ventricular assist devices (VAD). In this study, the real-time estimation of RCC based on time series data (TSD) of the left ventricular pressure (LVP) was investigated by exploiting two traditional time series classifiers (TSC) and two graph-based neural networks (GNN). All classifiers were assessed with respect to their RCC estimation accuracy and their applicability for real-time VAD control. Methods: For the RCC estimation (Figure 1), LVP TSD was generated using a hybrid mock circulation (HMC). On the HMC instantaneous pressure, volume, and flowrate values were computed by a numerical model of the human cardiovascular system and applied by a hydraulic interface on the inlet and outlet mixed-flow blood pump. The experimental protocol for the acquisition of the LVP TSD included 63 experiments, where preload, afterload, pump speed, and heart rate variations were performed for 9 RCC values. The data was pre-processed by segmenting and normalizing the TSD, producing 6300 cardiac cycle samples, which were used for the TSC approaches. The TSCs comprised the dynamic time warping nearest neighbor (DTW-NN) classifier and the support vector (SVM) classifier. Furthermore, the maturity of two GNN approaches was explored, including a simple custom architecture (C-GNN) and a pretrained architecture (P-GNN) provided by the KERAS library (Inception V3). The C-GNN comprised a convolutional layer with 32 filters of size 3x3, a max pooling layer with filter size 2x2, a dense layer with 128 units, and a dense layer with 9 units for the classification output. For the TSD to be used in the GNN frameworks, two different image encoders (IE), namely direct plot and recurrence plot encoders, were used and assessed. The data was split 80/20 for network training and validation. Results: All four classifiers, DTW-NN, SVM, C-GNN (IE: recurrence plot), and the P-GNN (IE: direct plot) achieved an accuracy of at least 98%. The SVM had the highest accuracy (99.9%) and the shortest prediction time (63 μs/sample). The prediction time of DTW-NN was on average 4.84 s/sample, prohibiting the implementation in real-time VAD control. Both GNN architectures achieve high accuracy and short prediction time, however, their performance does not reach the benchmark set by the SVM. A comparison of the image encoders showed that both the direct plot and the recurrence plot encoder lead to similar classification performances when using the same GNN, but different performances between the two investigated GNNs. Conclusion: All classification methods provided accurate RCC estimation with the SVM showing superior performance and being the most promising for real-time clinical implementation. These estimation approaches could substantially support patient surveillance and physiologic VAD control approaches. Figure 1. Schematic of the sequential steps for the estimation of the remaining cardiac contractility from time-series data (TSD) of the left ventricular pressure (LVP). a) LVP TSD is produced through a hybrid mock circulation, where the human cardiovascular system is simulated and the pressures, volume, and flowrates calculated are applied on the inlet and outlet of a ventricular assist device in real-time via a hydraulic interface. LVP TSD was acquired for various physiologic conditions. b) Pre-processing of the data, including cardiac cycle segmentation and time normalization. c) Translation of the TSD to images through different image encoders. d) Time series classifiers, including dynamic time warping nearest neighbor (DTW-NN) and the support vector (SVM) classifiers. e) Graph-based neural networks (GNN). f) Classified cardiac contractility.

Perbandingan K-Nearest Neighbor dan Random Forest dengan Seleksi Fitur Information Gain untuk Klasifikasi Lama Studi Mahasiswa

<p align="justify">Accreditation is a quality and feasibility assessment form in carrying out higher education. One of the factors that affect accreditation is the length of student study. In this study, the length of student study is classified by using the best attributes resulting from selecting information gain features. In optimizing the classification algorithm, we process the data by converting the original data into data that is ready to be mined. The next step is dividing the data into training and testing data so that the classification algorithm can be applied. This study gives the best four attributes, with <em>K</em>-nearest neighbor (K-NN) classification of 86.67% and random forest classification of 100%.</p><p><strong>Keywords</strong><strong>: </strong>length of study; information gain; <em>K</em>-nearest neighbor; random forest</p>

An active adaptation strategy for streaming time series classification based on elastic similarity measures

In streaming time series classification problems, the goal is to predict the label associated to the most recently received observations over the stream according to a set of categorized reference patterns. In on-line scenarios, data arise from non-stationary processes, which results in a succession of different patterns or events. This work presents an active adaptation strategy that allows time series classifiers to accommodate to the dynamics of streamed time series data. Specifically, our approach consists of a classifier that detects changes between events over streaming time series. For this purpose, the classifier uses features of the dynamic time warping measure computed between the streamed data and a set of reference patterns. When classifying a streaming series, the proposed pattern end detector analyzes such features to predict changes and adapt off-line time series classifiers to newly arriving events. To evaluate the performance of the proposed scheme, we employ the pattern end detection model along with dynamic time warping-based nearest neighbor classifiers over a benchmark of ten time series classification problems. The obtained results present exciting insights into the detection accuracy and latency performance of the proposed strategy.

M3SPCANet: A simple and effective ConvNets with unsupervised predefined filters for face recognition

The principal component analysis network (PCANet) with predefined filters has recognized as a lightweight convolutional neural networks (CNN) baseline. However, there are three shortages in PCANet: (i) single-scale convolution leads to insufficient feature learning, (ii) the spatial layout information and relationship between channel features are neglected, (iii) lack of nonlinearity because there is no activation function. Therefore, we devise a PCANet alternative dubbed Multi-Scale Spatial pyramid Second-order pooling Principle Component Analysis Network (M3SPCANet) in the paper. First, multi-scale PCA filters are learned and applied to obtain feature maps via convolution. Second, the feature maps are activated via nonlinear function Tanh and stacked together. Third, second-order pooling is exploited to extract correlation information between any two channels of the feature maps, moreover, spatial pyramid strategy is employed to learn local and holistic features simultaneously. Finally, whitening PCA based fully connected layer is constructed to remove redundancy information. We utilize cosine similarity based nearest neighbor classifier for face matching. Extensive experiments on four homogeneous and one heterogeneous face databases indicate that our method can achieve superior identification rates, and yield competitive verification performance. Cross database evaluations demonstrate that the multi-scale filters of our method have strong generalization ability. Our implementation code will be available at https://github.com/dxtyut/facerecognition.