• Journal of the Korea Academia-Industrial cooperation Society
• /
• v.16 no.5
• /
• pp.3378-3384
• /
• 2015

The problem of fast search for raman spectrum has attracted much attention recently. By far the most simple and widely used method is to calculate and compare the Euclidean distance between the given spectrum and the spectra in a database. But it is non-trivial problem because of the inherent high dimensionality of the data. One of the most serious problems is the high computational complexity of searching for the closet codeword. To overcome this problem, The fast codeword search algorithm based on the mean pyramids of codewords is currently used in image coding applications. In this paper, we present three new methods for the fast algorithm to search for the closet codeword. the proposed algorithm uses two significant features of a vector, mean values and variance, to reject many unlikely codewords and save a great deal of computation time. The Experiment results show about 42.8-55.2% performance improvement for the 1DMPS+PDS. The results obtained confirm the effectiveness of the proposed algorithm.

• Journal of Broadcast Engineering
• /
• v.26 no.4
• /
• pp.368-376
• /
• 2021

Artificial intelligence clouds help to efficiently develop the autonomous things integrating artificial intelligence technologies and control technologies by sharing the learned models and providing the execution environments. The existing autonomous things development technologies only take into account for the accuracy of artificial intelligence models at the cost of the increment of the complexity of the models including the raise up of the number of the hidden layers and the kernels, and they consequently require a large amount of computation. Since resource-constrained computing environments, could not provide sufficient computing resources for the complex models, they make the autonomous things violate time criticality. In this paper, we propose a digital twin software development framework that selects artificial intelligence models optimized for the computing environments. The proposed framework uses a load estimation DNN model to select the optimal model for the specific computing environments by predicting the load of the artificial intelligence models with digital twin data so that the proposed framework develops the control software. The proposed load estimation DNN model shows up to 20% of error rate compared to the formula-based load estimation scheme by means of the representative CNN models based experiments.

• Korean Journal of Remote Sensing
• /
• v.34 no.6_3
• /
• pp.1427-1443
• /
• 2018

Object-based image analysis (OBIA) allows higher computation efficiency and usability of information inherent in the image, as it reduces the complexity of the image while maintaining the image properties. Superpixel methods oversegment the image with a smaller image unit than an ordinary object segment and well preserve the edges of the image. SLIC (Simple linear iterative clustering) is known for outperforming the previous superpixel methods with high image segmentation quality. Although the input parameter for SLIC, number of superpixels has considerable influence on image segmentation results, impact analysis for SLIC parameter has not been investigated enough. In this study, we performed optimal parameter analysis and evaluation of change detection for SLIC-based superpixel techniques using KOMPSAT data. Forsuperpixel generation, three superpixel methods (SLIC; SLIC0, zero parameter version of SLIC; SNIC, simple non-iterative clustering) were used with superpixel sizes in ranges of $5{\times}5$ (pixels) to $50{\times}50$ (pixels). Then, the image segmentation results were analyzed for how well they preserve the edges of the change detection reference data. Based on the optimal parameter analysis, image segmentation boundaries were obtained from difference image of the bi-temporal images. Then, DBSCAN (Density-based spatial clustering of applications with noise) was applied to cluster the superpixels to a certain size of objects for change detection. The changes of features were detected for each superpixel and compared with reference data for evaluation. From the change detection results, it proved that better change detection can be achieved even with bigger superpixel size if the superpixels were generated with high regularity of size and shape.

• Sleep Medicine and Psychophysiology
• /
• v.6 no.1
• /
• pp.52-60
• /
• 1999

Objectives: With the object of finding the appropriate conditions and algorithms for dimensional analysis of human EEG, we calculated correlation dimensions in the various condition of sampling rate and data aquisition time and improved the computation algorithm by taking advantage of bit operation instead of log operation. Methods: EEG signals from 13 scalp lead of a man were digitized with A-D converter under the condition of 12 bit resolution and 1000 Hertz of sampling rate during 32 seconds. From the original data, we made 15 time series data which have different sampling rate of 62.5, 125, 250, 500, 1000 hertz and data acqusition time of 10, 20, 30 second, respectively. New algorithm to shorten the calculation time using bit operation and the Least Trimmed Squares(LTS) estimator to get the optimal slope was applied to these data. Results: The values of the correlation dimension showed the increasing pattern as the data acquisition time becomes longer. The data with sampling rate of 62.5 Hz showed the highest value of correlation dimension regardless of sampling time but the correlation dimension at other sampling rates revealed similar values. The computation with bit operation instead of log operation had a statistically significant effect of shortening of calculation time and LTS method estimated more stably the slope of correlation dimension than the Least Squares estimator. Conclusion: The bit operation and LTS methods were successfully utilized to time-saving and efficient calculation of correlation dimension. In addition, time series of 20-sec length with sampling rate of 125 Hz was adequate to estimate the dimensional complexity of human EEG.

• Journal of the Korean Institute of Telematics and Electronics S
• /
• v.36S no.11
• /
• pp.102-111
• /
• 1999

In this paper, we propose a fast full search block matching algorithm using the search region subsampling and the difference of adjacent pixels in current block. In the proposed algorithm, we calculate the lower bound of mean absolute difference (MAD) at each search point using the MAD value of neighbor search point and the difference of adjacent pixels in current block. After that, we perform block matching process only at the search points that need block matching process using the lower bound of MAD at each search point. To calculate the lower bound of MAD at each search point, we need the MAD value of neighbor search point. Therefore, the search points are subsampled at the factor of 4 and the MAD value at the subsampled search points are calculated by the block matching process. And then, the lower bound of MAD at the rest search points are calculated using the MAD value of the neighbor subsampled search point and the difference of adjacent pixels in current block. Finally, we discard the search points that have the lower bound of MAD value exceed the reference MAD which is the minimum MAD value of the MAD values at the subsampled search points and we perform the block matching process only at the search points that need block matching process. By doing so, we can reduce the computation complexity drastically while the motion compensated error performance is kept the same as that of full search block matching algorithm (FSBMA). The experimental results show that the proposed method has a much lower computational complexity than that of FSBMA while the motion compensated error performance of the proposed method is kept same as that of FSBMA.

• Journal of the Korea Academia-Industrial cooperation Society
• /
• v.18 no.12
• /
• pp.126-134
• /
• 2017

There have been numerous studies on extracting the R-peak from electrocardiogram (ECG) signals. However, most of the detection methods are complicated to implement in a real-time portable electrocardiograph device and have the disadvantage of requiring a large amount of calculations. R-peak detection requires pre-processing and post-processing related to baseline drift and the removal of noise from the commercial power supply for ECG data. An adaptive filter technique is widely used for R-peak detection, but the R-peak value cannot be detected when the input is lower than a threshold value. Moreover, there is a problem in detecting the P-peak and T-peak values due to the derivation of an erroneous threshold value as a result of noise. We propose a robust R-peak detection algorithm with low complexity and simple computation to solve these problems. The proposed scheme removes the baseline drift in ECG signals using an adaptive filter to solve the problems involved in threshold extraction. We also propose a technique to extract the appropriate threshold value automatically using the minimum and maximum values of the filtered ECG signal. To detect the R-peak from the ECG signal, we propose a threshold neighborhood search technique. Through experiments, we confirmed the improvement of the R-peak detection accuracy of the proposed method and achieved a detection speed that is suitable for a mobile system by reducing the amount of calculation. The experimental results show that the heart rate detection accuracy and sensitivity were very high (about 100%).

• Journal of Intelligence and Information Systems
• /
• v.24 no.1
• /
• pp.205-225
• /
• 2018

Convolutional Neural Network (ConvNet) is one class of the powerful Deep Neural Network that can analyze and learn hierarchies of visual features. Originally, first neural network (Neocognitron) was introduced in the 80s. At that time, the neural network was not broadly used in both industry and academic field by cause of large-scale dataset shortage and low computational power. However, after a few decades later in 2012, Krizhevsky made a breakthrough on ILSVRC-12 visual recognition competition using Convolutional Neural Network. That breakthrough revived people interest in the neural network. The success of Convolutional Neural Network is achieved with two main factors. First of them is the emergence of advanced hardware (GPUs) for sufficient parallel computation. Second is the availability of large-scale datasets such as ImageNet (ILSVRC) dataset for training. Unfortunately, many new domains are bottlenecked by these factors. For most domains, it is difficult and requires lots of effort to gather large-scale dataset to train a ConvNet. Moreover, even if we have a large-scale dataset, training ConvNet from scratch is required expensive resource and time-consuming. These two obstacles can be solved by using transfer learning. Transfer learning is a method for transferring the knowledge from a source domain to new domain. There are two major Transfer learning cases. First one is ConvNet as fixed feature extractor, and the second one is Fine-tune the ConvNet on a new dataset. In the first case, using pre-trained ConvNet (such as on ImageNet) to compute feed-forward activations of the image into the ConvNet and extract activation features from specific layers. In the second case, replacing and retraining the ConvNet classifier on the new dataset, then fine-tune the weights of the pre-trained network with the backpropagation. In this paper, we focus on using multiple ConvNet layers as a fixed feature extractor only. However, applying features with high dimensional complexity that is directly extracted from multiple ConvNet layers is still a challenging problem. We observe that features extracted from multiple ConvNet layers address the different characteristics of the image which means better representation could be obtained by finding the optimal combination of multiple ConvNet layers. Based on that observation, we propose to employ multiple ConvNet layer representations for transfer learning instead of a single ConvNet layer representation. Overall, our primary pipeline has three steps. Firstly, images from target task are given as input to ConvNet, then that image will be feed-forwarded into pre-trained AlexNet, and the activation features from three fully connected convolutional layers are extracted. Secondly, activation features of three ConvNet layers are concatenated to obtain multiple ConvNet layers representation because it will gain more information about an image. When three fully connected layer features concatenated, the occurring image representation would have 9192 (4096+4096+1000) dimension features. However, features extracted from multiple ConvNet layers are redundant and noisy since they are extracted from the same ConvNet. Thus, a third step, we will use Principal Component Analysis (PCA) to select salient features before the training phase. When salient features are obtained, the classifier can classify image more accurately, and the performance of transfer learning can be improved. To evaluate proposed method, experiments are conducted in three standard datasets (Caltech-256, VOC07, and SUN397) to compare multiple ConvNet layer representations against single ConvNet layer representation by using PCA for feature selection and dimension reduction. Our experiments demonstrated the importance of feature selection for multiple ConvNet layer representation. Moreover, our proposed approach achieved 75.6% accuracy compared to 73.9% accuracy achieved by FC7 layer on the Caltech-256 dataset, 73.1% accuracy compared to 69.2% accuracy achieved by FC8 layer on the VOC07 dataset, 52.2% accuracy compared to 48.7% accuracy achieved by FC7 layer on the SUN397 dataset. We also showed that our proposed approach achieved superior performance, 2.8%, 2.1% and 3.1% accuracy improvement on Caltech-256, VOC07, and SUN397 dataset respectively compare to existing work.