• The Journal of the Acoustical Society of Korea
• /
• v.29 no.2
• /
• pp.133-140
• /
• 2010

We improve the performance of cardiac disorder classification by adding new temporal features extracted from continuous heart sound signals. We add three kinds of novel temporal features to a conventional feature based on mel-frequency cepstral coefficients (MFCC): Heart sound envelope, murmur probabilities, and murmur amplitude variation. In cardiac disorder classification and detection experiments, we evaluate the contribution of the proposed features to classification accuracy and select proper temporal features using the sequential feature selection method. The selected features are shown to improve classification accuracy significantly and consistently for neural network-based pattern classifiers such as multi-layer perceptron (MLP), support vector machine (SVM), and extreme learning machine (ELM).

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.13 no.3
• /
• pp.1434-1449
• /
• 2019

Video object segmentation is a significant task in computer vision, but its performance is not very satisfactory. A method of video object segmentation using weakly temporal information is presented in this paper. Motivated by the phenomenon in reality that the motion of the object is a continuous and smooth process and the appearance of the object does not change much between adjacent frames in the video sequences, we use a feed-forward architecture with motion estimation to predict the mask of the current frame. We extend an additional mask channel for the previous frame segmentation result. The mask of the previous frame is treated as the input of the expanded channel after processing, and then we extract the temporal feature of the object and fuse it with other feature maps to generate the final mask. In addition, we introduce multi-mask guidance to improve the stability of the model. Moreover, we enhance segmentation performance by further training with the masks already obtained. Experiments show that our method achieves competitive results on DAVIS-2016 on single object segmentation compared to some state-of-the-art algorithms.

• Proceedings of the IEEK Conference
• /
• 2004.08c
• /
• pp.764-767
• /
• 2004

The performance of a speech recognition system is generally degraded in telephone environment because of distortions caused by background noise and various channel characteristics. In this paper, data-driven temporal filters are investigated to improve the performance of a specific recognition task such as telephone speech. Three different temporal filtering methods are presented with recognition results for Korean connected-digit telephone speech. Filter coefficients are derived from the cepstral domain feature vectors using the principal component analysis.

• ETRI Journal
• /
• v.32 no.3
• /
• pp.490-492
• /
• 2010

This letter proposes a robust feature extraction method using a spatially normalized mean for temporal ordinal measurement. Before computing a rank matrix from the mean values of non-overlapped blocks, each block mean is normalized so that it obeys the invariance property against linear additive and subtractive noise effects and is insensitive against multiplied and divided noise effects. Then, the temporal ordinal measures of spatially normalized mean values are computed for the feature matching. The performance of the proposed method showed about 95% accuracy in both precision and recall rates on various distortion environments, which represents the 2.7% higher performance on average compared to the temporal ordinal measurement.

• Korean Journal of Remote Sensing
• /
• v.29 no.5
• /
• pp.569-579
• /
• 2013

In this study, land-cover classification was implemented using features extracted from multi-temporal SAR data through 3D wavelet transform and the applicability of the 3D wavelet transform as a feature extraction approach was evaluated. The feature extraction stage based on 3D wavelet transform was first carried out before the classification and the extracted features were used as input for land-cover classification. For a comparison purpose, original image data without the feature extraction stage and Principal Component Analysis (PCA) based features were also classified. Multi-temporal Radarsat-1 data acquired at Dangjin, Korea was used for this experiment and five land-cover classes including paddy fields, dry fields, forest, water, and built up areas were considered for classification. According to the discrimination capability analysis, the characteristics of dry field and forest were similar, so it was very difficult to distinguish these two classes. When using wavelet-based features, classification accuracy was generally improved except built-up class. Especially the improvement of accuracy for dry field and forest classes was achieved. This improvement may be attributed to the wavelet transform procedure decomposing multi-temporal data not only temporally but also spatially. This experiment result shows that 3D wavelet transform would be an effective tool for feature extraction from multi-temporal data although this procedure should be tested to other sensors or other areas through extensive experiments.

• Phonetics and Speech Sciences
• /
• v.7 no.1
• /
• pp.3-10
• /
• 2015

In this paper, different frequency scales in cepstral feature extraction are evaluated for the text-independent speaker recognition. To this end, mel-frequency cepstral coefficients (MFCCs), linear frequency cepstral coefficients (LFCCs), and bilinear warped frequency cepstral coefficients (BWFCCs) are applied to the speaker recognition experiment. In addition, the spectro-temporal features extracted by the cepstral-time matrix (CTM) are examined as an alternative to the delta and delta-delta features. Experiments on the NIST speaker recognition evaluation (SRE) 2004 task are carried out using the Gaussian mixture model-universal background model (GMM-UBM) method and the joint factor analysis (JFA) method, both based on the ALIZE 3.0 toolkit. Experimental results using both the methods show that BWFCC with appropriate warping factor yields better performance than MFCC and LFCC. It is also shown that the feature set including the spectro-temporal information based on the CTM outperforms the conventional feature set including the delta and delta-delta features.

• The KIPS Transactions:PartB
• /
• v.18B no.6
• /
• pp.365-374
• /
• 2011

Recently, malicious video classification and filtering techniques are of practical interest as ones can easily access to malicious multimedia contents through the Internet, IPTV, online social network, and etc. Considerable research efforts have been made to developing malicious video classification and filtering systems. However, the malicious video classification and filtering is not still being from mature in terms of reliable classification/filtering performance. In particular, the most of conventional approaches have been limited to using only the spatial features (such as a ratio of skin regions and bag of visual words) for the purpose of malicious image classification. Hence, previous approaches have been restricted to achieving acceptable classification and filtering performance. In order to overcome the aforementioned limitation, we propose new malicious video classification framework that takes advantage of using both the spatial and temporal features that are readily extracted from a sequence of video frames. In particular, we develop the effective temporal features based on the motion periodicity feature and temporal correlation. In addition, to exploit the best data fusion approach aiming to combine the spatial and temporal features, the representative data fusion approaches are applied to the proposed framework. To demonstrate the effectiveness of our method, we collect 200 sexual intercourse videos and 200 non-sexual intercourse videos. Experimental results show that the proposed method increases 3.75% (from 92.25% to 96%) for classification of sexual intercourse video in terms of accuracy. Further, based on our experimental results, feature-level fusion approach (for fusing spatial and temporal features) is found to achieve the best classification accuracy.

• Journal of Internet Computing and Services
• /
• v.16 no.3
• /
• pp.21-32
• /
• 2015

Since spatio-temporal local features for video representation have become an important issue of modeless bottom-up approaches in action recognition, various methods for feature extraction and description have been proposed in many papers. In particular, BoF(bag of features) has been promised coherent recognition results. The most important part for BoF is how to represent dynamic information of actions in videos. Most of existing BoF methods consider the video as a spatio-temporal volume and describe neighboring 3D interest points as complex volumetric patches. To simplify these complex 3D methods, this paper proposes a novel method that builds BoF representation as a way to learn 2D interest points directly from video data. The basic idea of proposed method is to gather feature points not only from 2D xy spatial planes of traditional frames, but from the 2D time axis called spatio-temporal frame as well. Such spatial-temporal features are able to capture dynamic information from the action videos and are well-suited to recognize human actions without need of 3D extensions for the feature descriptors. The spatio-temporal BoF approach using SIFT and SURF feature descriptors obtains good recognition rates on a well-known actions recognition dataset. Compared with more sophisticated scheme of 3D based HoG/HoF descriptors, proposed method is easier to compute and simpler to understand.

• Journal of Korea Multimedia Society
• /
• v.18 no.4
• /
• pp.460-472
• /
• 2015

Most of previous visual attention system finds attention regions based on saliency map which is combined by multiple extracted features. The differences of these systems are in the methods of feature extraction and combination. This paper presents a new system which has an improvement in feature extraction method of color and motion, and in weight decision method of spatial and temporal features. Our system dynamically extracts one color which has the strongest response among two opponent colors, and detects the moving objects not moving pixels. As a combination method of spatial and temporal feature, the proposed system sets the weight dynamically by each features' relative activities. Comparative results show that our suggested feature extraction and integration method improved the detection rate of attention region.

• Journal of Computing Science and Engineering
• /
• v.12 no.1
• /
• pp.1-11
• /
• 2018

Human activities recognition is a challenging task due to its complexity of human movements and the variety performed by different subjects for the same action. This paper presents a recognition algorithm by using skeleton information generated from depth maps. Concatenating motion features and temporal constraint feature produces feature vector. Reducing dictionary scale proposes an improved fast classifier based on sparse representation. The developed method is shown to be effective by recognizing different activities on the UTD-MHAD dataset. Comparison results indicate superior performance of our method over some existing methods.