• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.7 no.7
• /
• pp.1657-1670
• /
• 2013

The increase in the number of music files in smart phone and MP3 player makes it difficult to find the music files which people want. So, Query-by-Singing/Humming (QbSH) systems have been developed to retrieve music from a user's humming or singing without having to know detailed information about the title or singer of song. Most previous researches on QbSH have been conducted using musical instrument digital interface (MIDI) files as reference songs. However, the production of MIDI files is a time-consuming process. In addition, more and more music files are newly published with the development of music market. Consequently, the method of using the more common MPEG-1 audio layer 3 (MP3) files for reference songs is considered as an alternative. However, there is little previous research on QbSH with MP3 files because an MP3 file has a different waveform due to background music and multiple (polyphonic) melodies compared to the humming/singing query. To overcome these problems, we propose a new QbSH method using MP3 files on mobile device. This research is novel in four ways. First, this is the first research on QbSH using MP3 files as reference songs. Second, the start and end positions on the MP3 file to be matched are estimated by using multi-layered perceptron (MLP) prior to performing the matching with humming/singing query file. Third, for more accurate results, four MLPs are used, which produce the start and end positions for dynamic time warping (DTW) matching algorithm, and those for chroma-based DTW algorithm, respectively. Fourth, two matching scores by the DTW and chroma-based DTW algorithms are combined by using PRODUCT rule, through which a higher matching accuracy is obtained. Experimental results with AFA MP3 database show that the accuracy (Top 1 accuracy of 98%, with an MRR of 0.989) of the proposed method is much higher than that of other methods. We also showed the effectiveness of the proposed system on consumer mobile device.

• The KIPS Transactions:PartB
• /
• v.11B no.6
• /
• pp.735-742
• /
• 2004

We present a new use of common graphics hardware to perform a faster artificial neural network. And we examine the use of GPU enhances the time performance of the image processing system using neural network, In the case of parallel computation of multiple input sets, the vector-matrix products become matrix-matrix multiplications. As a result, we can fully utilize the parallelism of GPU. Sigmoid operation and bias term addition are also implemented using pixel shader on GPU. Our preliminary result shows a performance enhancement of about thirty times faster using ATI RADEON 9800 XT board.

• IEMEK Journal of Embedded Systems and Applications
• /
• v.11 no.3
• /
• pp.163-171
• /
• 2016

Lane detection is a widely researched topic. Although simple road detection is easily achieved by previous methods, lane detection becomes very difficult in several complex cases involving noisy edges. To address this, we use a Convolution neural network (CNN) for image enhancement. CNN is a deep learning method that has been very successfully applied in object detection and recognition. In this paper, we introduce a robust lane detection method based on a CNN combined with random sample consensus (RANSAC) algorithm. Initially, we calculate edges in an image using a hat shaped kernel, then we detect lanes using the CNN combined with the RANSAC. In the training process of the CNN, input data consists of edge images and target data is images that have real white color lanes on an otherwise black background. The CNN structure consists of 8 layers with 3 convolutional layers, 2 subsampling layers and multi-layer perceptron (MLP) of 3 fully-connected layers. Convolutional and subsampling layers are hierarchically arranged to form a deep structure. Our proposed lane detection algorithm successfully eliminates noise lines and was found to perform better than other formal line detection algorithms such as RANSAC

• Proceedings of the Korea Inteligent Information System Society Conference
• /
• 2002.11a
• /
• pp.445-453
• /
• 2002

Past and current customer behavior is the best predicator of future customer behavior. This paper introduces a procedure on personalized defection detection and prevention for an online game site. The basic idea for our defection detection and prevention is adopted from the observation that potential defectors have a tendency to take a couple of months or weeks to gradually change their behavior (i.e. trim-out their usage volume) before their eventual withdrawal. For this purpose, we suggest a SOM (Self-Organizing Map) based procedure to determine the possible states of customer behavior from past behavior data. Based on this representation of the state of behavior, potential defectors are detected by comparing their monitored trajectories of behavior states with frequent and confident trajectories of past defectors. The key feature of this study includes a defection prevention procedure which recommends the desirable behavior state for the ext period so as to lower the likelihood of defection. The defection prevention procedure can be used to design a marketing campaign on an individual basis because it provides desirable behavior patterns for the next period. The experiments demonstrate that our approach is effective for defection prevention and efficient for defection detection because it predicts potential defectors without deterioration of prediction accuracy compared to that of the MLP (Multi-Layer Perceptron) neural network.

• The Journal of Korean Institute of Communications and Information Sciences
• /
• v.29 no.12C
• /
• pp.1660-1668
• /
• 2004

This paper proposes a Voice Activity Detection (VAD) algorithm based on Radial Basis Function (RBF) network using dual threshold. The k-means clustering and Least Mean Square (LMS) algorithm are used to upade the RBF network to the underlying speech condition. The inputs for RBF are the three parameters in a Code Exited Linear Prediction (CELP) coder, which works stably under various background noise levels. Dual hangover threshold applies in BRF-VAD for reducing error, because threshold value has trade off effect in VAD decision. The experimental result show that the proposed VAD algorithm achieves better performance than G.729 Annex B at any noise level.

• Smart Structures and Systems
• /
• v.14 no.2
• /
• pp.159-189
• /
• 2014

One of the most important requirements in the evaluation of existing structural systems and ensuring a safe performance during their service life is damage assessment. Damage can be defined as a weakening of the structure that adversely affects its current or future performance which may cause undesirable displacements, stresses or vibrations to the structure. The mass and stiffness of a structure will change due to the damage, which in turn changes the measured dynamic response of the system. Damage detection can increase safety, reduce maintenance costs and increase serviceability of the structures. Artificial Neural Networks (ANNs) are simplified models of the human brain and evolved as one of the most useful mathematical concepts used in almost all branches of science and engineering. ANNs have been applied increasingly due to its powerful computational and excellent pattern recognition ability for detecting damage in structural engineering. This paper presents and reviews the technical literature for past two decades on structural damage detection using ANNs with modal parameters such as natural frequencies and mode shapes as inputs.

• The Journal of Korean Institute of Communications and Information Sciences
• /
• v.24 no.12B
• /
• pp.2447-2454
• /
• 1999

This paper proposes a N-APPEM(Nonlinear A Posteriori Probability Estimation Method) with a frame normalization method to conventional classification network to increase speech recognition ability. It also tests the recognition ability of the classification and prediction neural networks for the Korean isolated digits. From the experimental results, the prediction network with MLP(Multi-Layer Perceptron) achieves the highest recognition ability of 98.0%. The prediction requires very complicated networks increased linearly with the number of incoming speech categories. However, the classification network with the N-APPEM and the normalization improves the recognition ability up to 85.5% with a sin81e network, which is almost 12.0% improvement.

• Proceedings of the KIEE Conference
• /
• 1996.11a
• /
• pp.404-406
• /
• 1996

The objective of this paper is to present a new feature-vector extraction method for the automatic detection and classification of power quality(PQ) disturbances, where FFT, DWT(Discrete Wavelet Transform), and data compression are utilized to extract an appropriate feature vector. In particular, the proposed classifier consists of three parts: i.e., (i) automatic detection of PQ disturbances, where the wavelet transform and signal power estimation method are utilized to detect each disturbance, (ii) feature vector extraction from the detected disturbance, and (iii) automatic classification, where Multi-Layer Perceptron(MLP) is used to classify each disturbance from the corresponding extracted feature vector. To demonstrate the performance and applicability of the proposed classification algorithm, some test results obtained by analyzing 7-class power quality disturbances generated by the EMTP are also provided.

• Journal of Korean Association for Spatial Structures
• /
• v.21 no.2
• /
• pp.99-110
• /
• 2021

This study presents the estimation of crack depth by analyzing temperatures extracted from thermal images and environmental parameters such as air temperature, air humidity, illumination. The statistics of all acquired features and the correlation coefficient among thermal images and environmental parameters are presented. The concrete crack depths were predicted by four different machine learning models: Multi-Layer Perceptron (MLP), Random Forest (RF), Gradient Boosting (GB), and AdaBoost (AB). The machine learning algorithms are validated by the coefficient of determination, accuracy, and Mean Absolute Percentage Error (MAPE). The AB model had a great performance among the four models due to the non-linearity of features and weak learner aggregation with weights on misclassified data. The maximum depth 11 of the base estimator in the AB model is efficient with high performance with 97.6% of accuracy and 0.07% of MAPE. Feature importances, permutation importance, and partial dependence are analyzed in the AB model. The results show that the marginal effect of air humidity, crack depth, and crack temperature in order is higher than that of the others.

• Journal of Positioning, Navigation, and Timing
• /
• v.12 no.1
• /
• pp.1-9
• /
• 2023

The problem of classifying a non-line-of-sight (NLOS) signal in a multipath channel is important to improve global navigation satellite system (GNSS) positioning accuracy in urban areas. Conventional deep learning-based NLOS signal classifiers use GNSS satellite measurements such as the carrier-to-noise-density ratio (CN_0), pseudorange, and elevation angle as inputs. However, there is a computational inefficiency with use of these measurements and the NLOS signal features expressed by the measurements are limited. In this paper, we propose a Convolutional Neural Network (CNN)-based NLOS signal classifier that receives successive Auto-correlation function (ACF) outputs according to a time-series, which is the most primitive output of GNSS signal processing. We compared the proposed classifier to other DL-based NLOS signal classifiers such as a multi-layer perceptron (MLP) and Gated Recurrent Unit (GRU) to show the superiority of the proposed classifier. The results show the proposed classifier does not require the navigation data extraction stage to classify the NLOS signals, and it has been verified that it has the best detection performance among all compared classifiers, with an accuracy of up to 97%.