Coronary Vessel Segmentation by Coarse-to-Fine Strategy using Otsu Algorithm and Decimation-Free Directional Filter Bank

Tan Dat Trinh,Pham The Bao,Le Nhi Lam Thuy,Ikuko Shimizu,김진영,Pham The Bao 한국전기전자학회 2019 전기전자학회논문지 Vol.23 No.2

In this study, a novel hierarchical approach is investigated to extract coronary vessel from X-ray angiogram. First, we propose to combine Decimation-free Directional Filter Bank (DDFB) and Homographic Filtering (HF)in order to enhance X-ray coronary angiographic image for segmentation purposes. Because the blood vesselensures that blood flows in only one direction on vessel branch, the DDFB filter is suitable to be used toenhance the vessels at different orientations and radius. In the combination with HF filter, our method cansimultaneously normalize the brightness across the image and increases contrast. Next, a coarse-to-finestrategy for iterative segmentation based on Otsu algorithm is applied to extract the main coronary vessels indifferent sizes. Furthermore, we also propose a new approach to segment very small vessels. Specifically,based on information of the main extracted vessels, we introduce a new method to extract junctions on thevascular tree and level of nodes on the tree. Then, the window based segmentation is applied to locate andextract the small vessels. Experimental results on our coronary X-ray angiography dataset demonstrate thatthe proposed approach can outperform standard method and attain the accuracy of 71.34%.

Improved Running Gaussian Average for Background Subtraction in Thermal Imagery

Tan Dat Trinh,Xinjie Ma,Jin Young Kim 한국정보기술학회 2017 한국정보기술학회논문지 Vol.15 No.7

In this paper, we propose a new method for background subtraction in thermal videos by improving the running Gaussian average technique (RGA). First, we propose a new background modeling even in the presence of moving objects in scene using region-based robust principle component analysis (RPCA). To enhance the performance or reduce computation cost of the RGA, we incorporate selectivity and random spatial subsampling techniques into background updating scheme. In addition, we also propose a new technique to deal with intensity sudden change problem by detecting corrupted frame based on skewness value of histogram followed by intensity enhancement using histogram matching method. Finally, we reduce number of ROI regions for human detection step by using a candidate extraction using morphology operator. Experiment results with our thermal database confirm that the proposed method significantly outperforms the baseline RGA and frame difference methods. Specially, the recall rate, precision rate and F value of the proposed method are 82.02%, 75.08% and 73.20% in comparison to 76.12%, 42.80% and 39.64%, of the baseline RGA, respectively.

Cheering Event Detection in Basketball Audio Stream Using Adaptive GMM Model and Low Rank Matrix Recovery

Tan Dat Trinh,Xinjie Ma,Hak-Jae Lee(이학재),Jin Young Kim(김진영),Seung-Ho Choi(최승호) 한국정보기술학회 2016 한국정보기술학회논문지 Vol.14 No.10

In this paper, we propose a new method to detect cheering events in basketball audio streams by combining short time Fourier transform (STFT) bin strengths, adaptive Gaussian mixture model (GMM) and low rank matrix recovery (LRR) approach. First, we apply the STFT and then calculate pre-defined frequency bins based on a specific frequency range of cheering sounds. An adaptive GMM model is used as a classifier to detect cheering events. In addition, we also propose to apply a post processing approach based on the LRR and power spectral density (PSD) within specified frequency interval to reduce false alarms and to improve the performance of the system. The experimental results on Korean basketball audio database demonstrate that our proposed method can outperform other well-known methods and achieve high accuracy. Specifically, recall rate, precision rate and F value are, respectively, 92.38%, 91.29% and 91.83%.

Multi-scale Pedestrian Detection in Thermal Imaging Using Deep Convolutional Neural Network and Adaptive NMS

Tan Dat Trinh,Jin Young Kim 한국정보기술학회 2018 한국정보기술학회논문지 Vol.16 No.9

In this paper, we propose a new method for pedestrian detection in thermal image/video by improving the non-maxima suppression(NMS) algorithm. We apply the sliding window detector based deep convolutional neural networks(DNN) to extract pedestrian candidates. A sliding window combined with image pyramids is used to identify pedestrians at varying scales and locations in the image via Convolutional Neural Network(CNN) based binary classification. To improve the performance, we propose an adaptive NMS algorithm to remove false alarms. The proposed NMS uses adaptive overlap-thresholds to overcome the drawback of the standard NMS and improve performance of detection system. It is automatically adjusting the overlap-thresholds based on the density of overlapping windows to improve the accuracy of system. Pedestrian detection experiment results with our thermal database and OSU thermal pedestrian database confirm that the proposed method outperforms the baseline method.

Robust Speaker Verification Using Low-Rank Matrix Recovery and Weighted Sparse Representation Under Total Variability Space

Tan Dat Trinh,Jin Young Kim,Pham The Bao,Seung Ho Choi,Keeseong Cho 한국정보기술학회 2016 한국정보기술학회논문지 Vol.14 No.3

In this paper, we propose a method to enhance performance of speaker verification by combining strengths of low-rank matrix recovery (LRR) based Fisher discriminant and weighed sparse representation (WSR) with auxiliary dictionary under total variability space. We use LRR based Fisher discriminant to make the training data more discriminate. Weighted sparse representation with auxiliary dictionary is applied to provide both sparsity and data locality structure, more robust and less sensitivity to outlier. Experiment results on utterances from Korean movie (“You Who Came From the Stars”) show that our proposed approach can significantly improve the performance of speaker verification and outperform the baseline sparse representation (SR)-ivector, LRR-SR-ivector and the other standard approaches in noisy environments.

A Modified Adaptive GMM Approach Based GMM Supervector and I-vector Using NMF Decomposition for Robust Speaker Verification

Tan Dat Trinh,Min Kyung Park(박민경),Jin Young Kim(김진영),Kyong Rok Lee(이경록),Seung Ho Choi(최승호),Keeseong Cho(조기성) 한국정보기술학회 2015 한국정보기술학회논문지 Vol.13 No.7

We propose a new method to enhance performance of speaker verification by investigating a novel modification of adaptive Gaussian Mixture Model (GMM) training. This model is trained using a modified Expectation Maximization (EM) algorithm, combined with a modified Maximum A Posteriori (MAP) estimation based weight factor of observation probabilities, called the observation confidence. The observation confidence is calculated based on the SNR estimation. Based on this modified adaptive GMM training algorithm, we propose to construct GMM supervectors and i-vectors, which are considered as input feature vectors for SVM. Besides, the discriminant features for speaker verification are also exploited by using non-negative matrix factorization (NMF) in the GMM-supervector and i-vector space. Experiment results on utterances from Korean drama (“You came from the stars”) show that our proposed methods significantly outperform the baseline GMM-UBM, GMM-supervector and i-vector based SVM under various noisy conditions.

Enhanced Face Recognition by Fusion of Global and Local Features under Varying Illumination

Tan Dat Trinh,Jin Young Kim(김진영),Pham The Bao 한국정보기술학회 2014 한국정보기술학회논문지 Vol.12 No.12

In this paper, we propose a new method to enhance the performance of face recognition under varying lighting condition. We try to combine strengths of illumination normalization, global and local features, feature-level and score-level fusion. Specially, we introduce two main contributions: 1) Firstly, we propose a feature-level fusion based on global and local Local binary patterns (LBP) features. Kernel PCA (KPCA) is used to reduce the dimension of the combined features. Then these features are used as input of SVM classifier; and 2) we further improve significantly the performance of face recognition by applying score-level fusion between global and local LBP features based SVM. An optimal method based Particle Swarm Optimization (PSO) is used to find optimal weights to fuse the aforementioned information at score-level. The experiment results on Korean face database demonstrate that our proposed methods outperform standard global feature, local feature and other well-know methods. Specifically, the best recognition rate is 100% for indoor images and 94.5% for outdoor images.

Coronary Vessel Segmentation by Coarse-to-Fine Strategy using Otsu Algorithm and Decimation-Free Directional Filter Bank

Trinh, Tan Dat,Tran, Thieu Bao,Thuy, Le Nhi Lam,Shimizu, Ikuko,Kim, Jin Young,Bao, Pham The Institute of Korean Electrical and Electronics Eng 2019 전기전자학회논문지 Vol.23 No.2

In this study, a novel hierarchical approach is investigated to extract coronary vessel from X-ray angiogram. First, we propose to combine Decimation-free Directional Filter Bank (DDFB) and Homographic Filtering (HF) in order to enhance X-ray coronary angiographic image for segmentation purposes. Because the blood vessel ensures that blood flows in only one direction on vessel branch, the DDFB filter is suitable to be used to enhance the vessels at different orientations and radius. In the combination with HF filter, our method can simultaneously normalize the brightness across the image and increases contrast. Next, a coarse-to-fine strategy for iterative segmentation based on Otsu algorithm is applied to extract the main coronary vessels in different sizes. Furthermore, we also propose a new approach to segment very small vessels. Specifically, based on information of the main extracted vessels, we introduce a new method to extract junctions on the vascular tree and level of nodes on the tree. Then, the window based segmentation is applied to locate and extract the small vessels. Experimental results on our coronary X-ray angiography dataset demonstrate that the proposed approach can outperform standard method and attain the accuracy of 71.34%.

Audio Event Classification Using SVM with GMM-UBM Supervectors

Tan Dat Trinh,Ngoc Nam Bui,So Hee Min,Jin Young Kim 한국정보기술학회 2013 한국정보기술학회논문지 Vol.11 No.11

Audio event recognition is a fascinating and challenging research topic in signal processing, audio retrieval, and pattern recognition. In this paper, we investigate GMM supervector based Universal Background Model (UBM) and Support Vector Machine (SVM) with MFCC features and various kernels for audio event recognition. A GMM supervector is obtained by adapting with UBM and cascading all the mean vector components. After that, the supervectors are applied as input features for SVM classifier. Experimental results belonging to our audio event database demonstrates that the proposed approach outperforms standard GMM-UBM baseline. Moreover, when applying SVM with GUMI kernel, error rate significantly decreases from 26.52% to 14.97% for 16 mixtures.

Comparison of Observation Confidence Estimators for Robust Speaker Verification

Xinjie Ma,Tan Dat Trinh,Jin Young Kim 한국정보기술학회 2017 한국정보기술학회논문지 Vol.15 No.7

In this paper, we first explore a modified adaptive Gaussian mixture model (MAGMM) by investigating the confidence value of observation vectors to deal with noise conditions problem. The observation confidence values are estimated by using the frame SNR values calculated between the input noisy speech and the enhanced speech, and the sigmoid function. We compare three speech enhancement techniques, minimum mean square error logarithm shorttime spectral amplitude (MMSE log-STSA), low-rank matrix recovery (LRR) and multiple low-rank representation (MLRR), for observation confidence computation. Furthermore, we also consider the effect of the use of observation confidence value in the GMM-supervector (GSV) and i-vector approaches which are considers as input feature vectors for the Support vector machine (SVM). To verify the accuracy of the speaker system, we use utterances from a Korean drama “You came from the stars.” The experimental results show that our proposed approaches achieve better performance than the baseline systems under noisy environments.