Traditional Gaussian Mixture Model Research Articles

Improvements on Gaussian mixture model and its application in identifying aerosol types in two major cities in the Yangtze River Delta, China

Accurately identifying the authentic local aerosol types is one of the fundamental tasks in studying aerosol radiative effects and model assessment. In this paper, improvements were made to the traditional Gaussian Mixture Model, leading to the following results: 1) This study introduces several improvements to the traditional Gaussian Mixture Model (GMM), referred to as M-GMMs. These improvements include the incorporation of multivariate kurtosis coefficients, Mahalanobis distance instead of Euclidean distance, and weights of variables. The M-GMMs overcome the issues related to dimensional units and correlations among multiple parameters, thereby enhancing the estimation of the covariance matrix. 2) The proposed M-GMMs model was evaluated for its clustering performance using machine-generated data with known classifications and real iris flower data. The results demonstrated that the classification performance of M-GMMs was superior to other models. Furthermore, compared to the slightly less effective K-means algorithm (which requires manual definition of the number of aerosol types), the M-GMMs model was able to automatically iterate and produce consistent classification results based on similar characteristics. 3) There is still a significant disparity between the characteristics of real stations and typical aerosols. Directly evaluating local aerosols using the characteristics of typical aerosols results in substantial errors. However, the M-GMMs model can effectively reflect the authentic aerosol characteristics at the local level. 4) The M-GMMs model was utilized to perform cluster analysis on the Xuzhou and Nanjing stations of AERONET. This analysis yielded quantitative proportions, temporal distribution characteristics, and spectral distribution features of aerosol types in the two regions. The improved M-GMMs model presented in this paper enables more accurate and continuous characterization of aerosol type variations. Its findings hold significant theoretical and practical value in reassessing aerosol radiative effects.

Residual networks for text-independent speaker identification: Unleashing the power of residual learning

The human voice, a dynamic signal, conveys valuable information for speaker identification, encompassing gender, age, emotions, and language. In the biometrics industry, identifying voices in real-time amidst diverse accents, tones, and noisy backgrounds is a challenging task. Voice biometry, a complex aspect of speaker identification, is gaining importance in various applications, such as user authentication, attendance systems, forensics, and banking operations, as it eliminates the need for traditional credentials like cards or passwords. Recent advancements in Human–Computer Interaction technology have made conversational tasks technically feasible. Deep Neural Learning approaches, especially Convolutional Deep Neural Networks (CDNN), have emerged as a powerful tool in the field of speech processing, surpassing traditional Speaker Identification methods. This paper introduces a novel approach using 1-Dimensional Convolutional Residual Blocks for audio classification and Speaker Identification, specifically focusing on speaker recognition from spoken Hindi language. The proposed Residual architecture significantly enhances speaker identification, even in low Signal Noise Ratio environments, achieving an impressive accuracy rate of 86.02%. This outperforms traditional Gaussian Mixture Model (GMM) and Feed Forward Back-propagation Network (FFBN) model for the same set of speakers. Future research directions may explore the classification of audio and speaker identification using various acoustic features derived from speech signals.

An Improved Semi-Supervised Gaussian Mixture Model (I-SGMM)

In the era of data-driven decision-making, the Gaussian Mixture Model (GMM) stands as a cornerstone in statistical modeling, particularly in clustering and density estimation. The Improved GMM presents a robust solution to a fundamental problem in clustering: the determination of the optimal number of clusters. Unlike its predecessor, it does not rely on a predetermined cluster count but employs model selection criteria, such as the Bayesian Information Criterion (BIC) or Akaike Information Criterion (AIC), to automatically identify the most suitable cluster count for the given data. This inherent adaptability is a hallmark of the Improved GMM, making it a versatile tool in a broad spectrum of applications, from market segmentation to image processing. Furthermore, the Improved GMM revolutionizes parameter estimation and model fitting. It leverages advanced optimization techniques, such as the Expectation-Maximization (EM) algorithm or variational inference, to achieve convergence to more favorable local optima. This results in precise and reliable parameter estimates, including cluster means, covariances, and component weights. The Improved GMM is particularly invaluable when dealing with data of varying complexities, non-standard data distributions, and clusters with differing shapes and orientations. It excels at capturing the nuanced relationships within the data, providing a powerful framework for understanding complex systems. One of the key differentiators of the Improved GMM is its accommodation of full covariance matrices for each component. This feature empowers the model to account for intricate interdependencies between variables, which is essential for modeling real-world data effectively. It is capable of handling data that exhibits non-spherical or irregular cluster shapes, a significant limitation of the traditional GMM.

Effective Quantization Evaluation Method of Functional Movement Screening with Improved Gaussian Mixture Model

Background: Functional movement screening (FMS) allows for the rapid assessment of an individual’s physical activity level and the timely detection of sports injury risk. However, traditional functional movement screening often requires on-site assessment by experts, which is time-consuming and prone to subjective bias. Therefore, the study of automated functional movement screening has become increasingly important. Methods: In this study, we propose an automated assessment method for FMS based on an improved Gaussian mixture model (GMM). First, the oversampling of minority samples is conducted, the movement features are manually extracted from the FMS dataset collected with two Azure Kinect depth sensors; then, we train the Gaussian mixture model with different scores (1 point, 2 points, 3 points) of feature data separately; finally, we conducted FMS assessment by using a maximum likelihood estimation. Results: The improved GMM has a higher scoring accuracy (improved GMM: 0.8) compared to other models (traditional GMM = 0.38, AdaBoost.M1 = 0.7, Naïve Bayes = 0.75), and the scoring results of improved GMM have a high level of agreement with the expert scoring (kappa = 0.67). Conclusions: The results show that the proposed method based on the improved Gaussian mixture model can effectively perform the FMS assessment task, and it is potentially feasible to use depth cameras for FMS assessment.

Gaussian Mixture Modeling Extensions for Improved False Discovery Rate Estimation in GC-MS Metabolomics.

The ability to reliably identify small molecules (e.g., metabolites) is key toward driving scientific advancement in metabolomics. Gas chromatography-mass spectrometry (GC-MS) is an analytic method that may be applied to facilitate this process. The typical GC-MS identification workflow involves quantifying the similarity of an observed sample spectrum and other features (e.g., retention index) to that of several references, noting the compound of the best-matching reference spectrum as the identified metabolite. While a deluge of similarity metrics exist, none quantify the error rate of generated identifications, thereby presenting an unknown risk of false identification or discovery. To quantify this unknown risk, we propose a model-based framework for estimating the false discovery rate (FDR) among a set of identifications. Extending a traditional mixture modeling framework, our method incorporates both similarity score and experimental information in estimating the FDR. We apply these models to identification lists derived from across 548 samples of varying complexity and sample type (e.g., fungal species, standard mixtures, etc.), comparing their performance to that of the traditional Gaussian mixture model (GMM). Through simulation, we additionally assess the impact of reference library size on the accuracy of FDR estimates. In comparing the best performing model extensions to the GMM, our results indicate relative decreases in median absolute estimation error (MAE) ranging from 12% to 70%, based on comparisons of the median MAEs across all hit-lists. Results indicate that these relative performance improvements generally hold despite library size; however FDR estimation error typically worsens as the set of reference compounds diminishes.

An improved Gaussian Mixture Model with post-processing for multiple object detection in surveillance video analytics

Gaussian Mixture Model (GMM) is an effective method for extracting foreground objects from video sequences. However, GMM fails to detect the object in challenging scenarios like the presence of shadow, occlusion, complex backgrounds, etc. To handle these challenges, intrinsic and extrinsic enhancement is required in traditional GMM. This paper presents a novel framework that combines improved GMM with postprocessing for multiple object detection. In the proposed system, GMM with parameter initialization is considered an intrinsic improvement. Video preprocessing and postprocessing are considered extrinsic improvements. Integration of morphological operation with GMM helps for better segmentation than traditional GMM, and it also helps to increase detection performance by reducing false positives. Video preprocessing is the process of noise removal that prepares input video ready for further processing. In the final step gradient of morphological operations is used for postprocessing. The proposed approach was tested on challenging surveillance video sequences from benchmark datasets such as PETS 2009 and CD 2014(Change Detection). The experimental results are compared using ground truth and performance evaluation metrics. The results show that the proposed approach performs better than GMM, and the method can detect the object effectively even in illumination variation and partial occlusion.

Image Point Set Registration: Phased Gaussian Mixture Model

The introduction of probability model into point set matching has achieved great success. Because the image point set has more outliers and the image has complex transformation, the traditional Gaussian Mixture Model (GMM) does not perform well in image point set matching. In this paper, we propose a Phased Gaussian Mixture Model (PGMM) for point set registration between images. Different from the existing registration algorithms based on GMM, PGMM uses Locality Sensitive Hashing to distinguish the feature reliability of feature points. The initial transformation relation is obtained in the segment with high feature reliability and vector constraint is applied to other segments. The average Hamming distance is used to give different weights to each Gaussian component, which is related to the characteristics between data points and model points. The proposed method has good performance in the case of large image transformation scale.

Efficient Johnson-SB Mixture Model for Segmentation of CT Liver Image.

To overcome the problem that the traditional Gaussian mixture model (GMM) cannot well describe the skewness distribution of the gray-level histogram of a liver CT slice, we propose a novel segmentation method for liver CT images by introducing the Johnson-SB mixture model (JSBMM). The Johnson-SB model not only has a flexible asymmetrical distribution but also covers a variety of other distributions as well. In this article, the parameter optimization formulas for JSBMM were derived by employing the expectation-maximization (EM) algorithm and maximum likelihood. The implementation process of the JSBMM-based segmentation algorithm is provided in detail. To make better use of the skewness of Johnson-SB and improve the segmentation accuracy, we devise an idea to divide the histogram into two parts and calculate the segmentation threshold for each part, respectively, which is called JSBMM-TDH. By analyzing and comparing the segmentation thresholds with different cluster numbers, it is illustrated that the segmentation threshold of JSBMM-TDH will tend to be stable with the increasing of cluster number, while that of GMM is sensitive to different cluster numbers. The proposed JSBMM-TDH is applied to segment four randomly obtained abdominal CT image sequences, and the segmentation results and robustness have been compared between JSBMM-TDH and GMM. It is verified that JSBMM-TDH has preferable segmentation results and better robustness than GMM for the segmentation of liver CT images.

Mixture Representation Learning for Deep Speaker Embedding

How to effectively convert a sequence of variable-length acoustic features to a fixed-dimension representation has always been a research focus in speaker recognition. In state-of-the-art speaker recognition systems, the conversion is implemented by concatenating the mean and the standard deviation of a sequence of frame-level features. However, a single mean and a single standard deviation are limited descriptive statistics for an acoustic sequence even with powerful feature extractors such as convolutional neural networks. In this paper, we propose a novel statistics pooling method that can produce more descriptive statistics through a mixture representation. Our approach is inspired by the expectation–maximization (EM) algorithm in Gaussian mixture models (GMMs). Instead of using traditional GMM style alignment, we novelly leverage modern deep learning tools to produce a more powerful mixture representation. The novelty includes: (1) unlike GMMs, the mixture assignments are determined by an attention network instead of the Euclidean distances between the frame-level features and explicit centers; (2) instead of using a single frame as input to the attention network, contextual frames are included to smooth out attention transition; and (3) soft-attention assignments are replaced by hard-attention assignments via the Gumbel-Softmax with straight-through estimators. With the proposed attention mechanism, we obtained a 13.7% relative improvement over vanilla mean and standard deviation pooling in the VOiCES19-eval set.

Research on Voiceprint Recognition of Camouflage Voice Based on Deep Belief Network

The problem of disguised voice recognition based on deep belief networks is studied. A hybrid feature extraction algorithm based on formants, Gammatone frequency cepstrum coefficients (GFCC) and their different coefficients is proposed to extract more discriminative speaker features from the original voice data. Using mixed features as the input of the model, a masquerade voice library is constructed. A masquerade voice recognition model based on a depth belief network is proposed. A dropout strategy is introduced to prevent overfitting, which effectively solves the problems of traditional Gaussian mixture models, such as insufficient modeling ability and low discrimination. Experimental results show that the proposed disguised voice recognition method can better fit the feature distribution, and significantly improve the classification effect and recognition rate.

Directly Comparing the Listening Strategies of Humans and Machines

Automatic speech recognition (ASR) has reached human performance on many clean speech corpora, but it remains worse than human listeners in noisy environments. This paper investigates whether this difference in performance might be due to a difference in the time-frequency regions that each listener utilizes in making their decisions and how these “important” regions change for ASRs using different acoustic models (AMs) and language models (LMs). We define important regions as time-frequency points in a spectrogram that tend to be audible when the listener correctly recognizes that utterance in noise. The evidence from this study indicates that a neural network AM attends to regions that are more similar to those of humans (capturing certain high-energy regions) than those of a traditional Gaussian mixture model (GMM) AM. Our analysis also shows that the neural network AM has not yet captured all the cues that human listeners utilize, such as certain transitions between silence and high speech energy. We also find that differences in important time-frequency regions tend to track differences in accuracy on specific words in a test sentence, suggesting a connection. Because of this connection, adapting an ASR to attend to the same regions humans use might improve its generalization in noise.

Outdoor object detection for surveillance based on modified GMM and Adaptive Thresholding

This paper presents a modified Gaussian Mixture Model (GMM) and Adaptive Thresholding designed to improve object detection accuracy for the outdoor surveillance. Intrinsic and extrinsic improvements in traditional GMM will handle the outdoor dynamic scenes e.g. tree weaving, gradual illumination changes, partial occlusions and also handles certain amount of shadow. For foreground detection, Adaptive Thresholding is utilized for better classification among the background and foreground objects and it also help to reduces false positives and hence increases the detection accuracy. We tested proposed algorithm on standard datasets consisting of CDnet 2014, PETS 2009 and ViSOR. The robustness of the propose algorithm has been compared with the ground truth and other similar approaches through several performance evaluation metrics. The experimental results conclude that the proposed algorithm efficiently detect objects in dynamic environments as well as handle partial occlusions and certain amount of shadows very efficiently.

An Automatic Recognition Method of Microseismic Signals Based on S Transformation and Improved Gaussian Mixture Model

The microseismic signals in the coal minefield are very complex because of its special environment with a large number of blast vibration signals, and how to effectively identify the microseismic signals is still a big problem. S transform (ST) and Manifold Learning (ML) methods are introduced to extract the characteristics of the microseismic signals, and Gaussian Mixture Model based on the improved Bee Colony optimization algorithm (IBC‐GMM) is established to identify the microseismic signals accurately. Firstly, the time‐frequency characteristics of microseismic signals in coal mine are extracted by ST analysis. It is found that there are obvious time‐frequency differences between rock‐fracturing signals and blast vibration signals. Blast vibration signals have short duration, high frequency, and complex frequency spectrum, and their dominant frequencies are mainly over 100 Hz. However, rock‐fracturing signals are relatively slow, with low frequency and stable spectrum change, and their dominant frequencies are generally below 100 Hz. Then, combining with the microseismic data of Xiashinjie coal mine in Tongchuan, China, the feature dimension reduction is carried out by Manifold Learning (ML) method, and the processed feature vectors are automatically recognized by IBC‐GMM. Field test results show that the method summarizes the characteristics of the microseismic wave which are difficult to emerge as the learning vector, and the features reflect the key features of microseismic signals well. The identification accuracy is as high as 94%, and its recognition effect is superior to other recognition models (such as traditional Gaussian Mixture Model based on Expectation‐Maximum (EM‐GMM), Backpropagation (BP) neural network, Random Forests (RF), Bayes (Bayes) methods, and Logistic Regression (LR) method). Therefore, IBC‐GMM could be used to mine engineering microseismic monitoring waveform recognition to provide the reference.

Clustering in the wireless channel with a power weighted statistical mixture model in indoor scenario

Cluster-based channel model is the main stream of fifth generation mobile communications, thus the accuracy of clustering algorithm is important. Traditional Gaussian mixture model (GMM) does not consider the power information which is important for the channel multipath clustering. In this paper, a normalized power weighted GMM (PGMM) is introduced to model the channel multipath components (MPCs). With MPC power as a weighted factor, the PGMM can fit the MPCs in accordance with the cluster-based channel models. Firstly, expectation maximization (EM) algorithm is employed to optimize the PGMM parameters. Then, to further increase the searching ability of EM and choose the optimal number of components without resort to cross-validation, the variational Bayesian (VB) inference is employed. Finally, 28 GHz indoor channel measurement data is used to demonstrate the effectiveness of the PGMM clustering algorithm.

Moving Target Detection Based on Improved Gaussian Mixture Background Subtraction in Video Images

In recent years, background subtraction techniques have been used in vision and image applications for moving target detection. However, most methods cannot provide fine results due to dynamic backgrounds, noise, etc. The Gaussian mixture model (GMM) is a background modeling method commonly used in moving target detection. The traditional GMM method is vulnerable to noise interference, especially from dynamic backgrounds; thus, its detection performance is not good. Because of the influence of background noise and dynamic effects on moving target detection, we propose a method of moving target detection for dynamic backgrounds based on improved GMM background subtraction. This method can be divided into three stages. First, in the background modeling stage, to facilitate calculation and improve modeling speed, the video frame is blocked, and the background model is reconstructed using the image block averaging method. Second, in the moving target detection stage, the method of combining wavelet semi-threshold function denoising with mathematical morphology closed operation is used for denoising, which effectively eliminates the influence of noise and improves the detection effect. Third, in the background updating stage, the adaptive background updating method is used to update the background to improve detection results. The simulation results show that the improved method can reduce noise and dynamic background interference while improving moving target detection, thereby proving the effectiveness and adaptability of the proposed method.

Speaker identification using multi-modal i-vector approach for varying length speech in voice interactive systems

The development in the interface of smart devices has lead to voice interactive systems. An additional step in this direction is to enable the devices to recognize the speaker. But this is a challenging task because the interaction involves short duration speech utterances. The traditional Gaussian mixture models (GMM) based systems have achieved satisfactory results for speaker recognition only when the speech lengths are sufficiently long. The current state-of-the-art method utilizes i-vector based approach using a GMM based universal background model (GMM-UBM). It prepares an i-vector speaker model from a speaker’s enrollment data and uses it to recognize any new test speech. In this work, we propose a multi-model i-vector system for short speech lengths. We use an open database THUYG-20 for the analysis and development of short speech speaker verification and identification system. By using an optimum set of mel-frequency cepstrum coefficients (MFCC) based features we are able to achieve an equal error rate (EER) of 3.21% as compared to the previous benchmark score of EER 4.01% on the THUYG-20 database. Experiments are conducted for speech lengths as short as 0.25 s and the results are presented. The proposed method shows improvement as compared to the current i-vector based approach for shorter speech lengths. We are able to achieve improvement of around 28% even for 0.25 s speech samples. We also prepared and tested the proposed approach on our own database with 2500 speech recordings in English language consisting of actual short speech commands used in any voice interactive system.

Research and Improvement of Content-Based Image Retrieval Framework

This paper proposed a high-performance image retrieval framework, which combines the improved feature extraction algorithm SIFT (Scale Invariant Feature Transform), improved feature matching, improved feature coding Fisher and improved Gaussian Mixture Model (GMM) for image retrieval. Aiming at the problem of slow convergence of traditional GMM algorithm, an improved GMM is proposed. This algorithm initializes the GMM by using on-line [Formula: see text]-means clustering method, which improves the convergence speed of the algorithm. At the same time, when the model is updated, the storage space is saved through the improvement of the criteria for matching rules and generating new Gaussian distributions. Aiming at the problem that the dimension of SIFT (Scale Invariant Feature Transform) algorithm is too high, the matching speed is too slow and the matching rate is low, an improved SIFT algorithm is proposed, which preserves the advantages of SIFT algorithm in fuzzy, compression, rotation and scaling invariance advantages, and improves the matching speed, the correct match rate is increased by an average of 40% to 55%. Experiments on a recently released VOC 2012 database and a database of 20 category objects containing 230,800 images showed that the framework had high precision and recall rates and less query time. Compared with the standard image retrieval framework, the improved image retrieval framework can detect the moving target quickly and effectively and has better robustness.

A Novel GMM-Based Behavioral Modeling Approach for Smartwatch-Based Driver Authentication.

All drivers have their own distinct driving habits, and usually hold and operate the steering wheel differently in different driving scenarios. In this study, we proposed a novel Gaussian mixture model (GMM)-based method that can improve the traditional GMM in modeling driving behavior. This new method can be applied to build a better driver authentication system based on the accelerometer and orientation sensor of a smartwatch. To demonstrate the feasibility of the proposed method, we created an experimental system that analyzes driving behavior using the built-in sensors of a smartwatch. The experimental results for driver authentication—an equal error rate (EER) of 4.62% in the simulated environment and an EER of 7.86% in the real-traffic environment—confirm the feasibility of this approach.

Supervoxel Segmentation with Voxel-Related Gaussian Mixture Model.

Extended from superpixel segmentation by adding an additional constraint on temporal consistency, supervoxel segmentation is to partition video frames into atomic segments. In this work, we propose a novel scheme for supervoxel segmentation to address the problem of new and moving objects, where the segmentation is performed on every two consecutive frames and thus each internal frame has two valid superpixel segmentations. This scheme provides coarse-grained parallel ability, and subsequent algorithms can validate their result using two segmentations that will further improve robustness. To implement this scheme, a voxel-related Gaussian mixture model (GMM) is proposed, in which each supervoxel is assumed to be distributed in a local region and represented by two Gaussian distributions that share the same color parameters to capture temporal consistency. Our algorithm has a lower complexity with respect to frame size than the traditional GMM. According to our experiments, it also outperforms the state-of-the-art in accuracy.

Deep neural network training for whispered speech recognition using small databases and generative model sampling

State-of-the-art speech recognition solutions currently employ hidden Markov models (HMMs) to capture the time variability in a speech signal and deep neural networks (DNNs) to model the HMM state distributions. It has been shown that DNN–HMM hybrid systems outperform traditional HMM and Gaussian mixture model (GMM) hybrids in many applications. This improvement is mainly attributed to the ability of DNNs to model more complex data structures. However, having sufficient data samples is one key point in training a high accuracy DNN as a discriminative model. This barrier makes DNNs unsuitable for many applications with limited amounts of data. In this study, we introduce a method to produce an excessive amount of pseudo-samples that requires availability of only a small amount of transcribed data from the target domain. In this method, a universal background model (UBM) is trained to capture a parametric estimate of the data distributions. Next, random sampling is used to generate a large amount of pseudo-samples from the UBM. Frame-Shuffling is then applied to smooth the temporal cepstral trajectories in the generated pseudo-sample sequences to better resemble the temporal characteristics of a natural speech signal. Finally, the pseudo-sample sequences are combined with the original training data to train the DNN–HMM acoustic model of a speech recognizer. The proposed method is evaluated on small-sized sets of neutral and whisper data drawn from the UT-Vocal Effort II corpus. It is shown that phoneme error rates (PERs) of a DNN–HMM based speech recognizer are considerably reduced when incorporating the generated pseudo-samples in the training process, with + 79.0 and + 45.6% relative PER improvements for neutral–neutral training/test and whisper–whisper training/test scenarios, respectively.