• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.16 no.1
• /
• pp.1-15
• /
• 2022

The MobileNetV3 is specially designed for mobile devices with limited memory and computing power. To reduce the network parameters and improve the network inference speed, a new lightweight network is proposed based on MobileNetV3. Firstly, to reduce the computation of residual blocks, a partial residual structure is designed by dividing the input feature maps into two parts. The designed partial residual structure is used to replace the residual block in MobileNetV3. Secondly, a dual-path feature extraction structure is designed to further reduce the computation of MobileNetV3. Different convolution kernel sizes are used in the two paths to extract feature maps with different sizes. Besides, a transition layer is also designed for fusing features to reduce the influence of the new structure on accuracy. The CIFAR-100 dataset and Image Net dataset are used to test the performance of the proposed partial residual structure. The ResNet based on the proposed partial residual structure has smaller parameters and FLOPs than the original ResNet. The performance of improved MobileNetV3 is tested on CIFAR-10, CIFAR-100 and ImageNet image classification task dataset. Comparing MobileNetV3, GhostNet and MobileNetV2, the improved MobileNetV3 has smaller parameters and FLOPs. Besides, the improved MobileNetV3 is also tested on CPU and Raspberry Pi. It is faster than other networks

• Journal of the Institute of Convergence Signal Processing
• /
• v.24 no.4
• /
• pp.186-191
• /
• 2023

Data processing techniques play a critical role in decision-making, including handling missing and outlier data, prediction, and recommendation models. This requires a clear explanation of the validity, reliability, and accuracy of all processes and results. In addition, it is necessary to solve data problems through explainable models using decision trees, inference, etc., and proceed with model lightweight by considering various types of learning. The multi-layer mining classification method that applies the sixth principle is a method that discovers multidimensional relationships between variables and attributes that occur frequently in transactions after data preprocessing. This explains how to discover significant relationships using mining on transactions and model the data through regression analysis. It develops scalable models and logistic regression models and proposes mining techniques to generate class labels through data cleansing, relevance analysis, data transformation, and data augmentation to make explanatory decisions.

• KIPS Transactions on Software and Data Engineering
• /
• v.9 no.5
• /
• pp.169-176
• /
• 2020

Recently, deep neural networks (DNNs) have brought revolutions to many mobile and embedded devices by providing human-level machine intelligence for various applications. However, high inference accuracy of such DNNs comes at high computational costs, and, hence, there have been significant efforts to reduce computational overheads of DNNs either by compressing off-the-shelf models or by designing a new small footprint DNN architecture tailored to resource constrained devices. One notable recent paradigm in designing small footprint DNN models is sharing parameters in several layers. However, in previous approaches, the parameter-sharing techniques have been applied to large deep networks, such as ResNet, that are known to have high redundancy. In this paper, we propose a parameter-sharing method for already parameter-efficient small networks such as ShuffleNetV2. In our approach, small templates are combined with small layer-specific parameters to generate weights. Our experiment results on ImageNet and CIFAR100 datasets show that our approach can reduce the size of parameters by 15%-35% of ShuffleNetV2 while achieving smaller drops in accuracies compared to previous parameter-sharing and pruning approaches. We further show that the proposed approach is efficient in terms of latency and energy consumption on modern embedded devices.

• Journal of the Institute of Electronics Engineers of Korea SC
• /
• v.42 no.5 s.305
• /
• pp.27-34
• /
• 2005

Ventricular fibrillation(VF) is generally caused by chaotic behavior of electrical propagation in heart and may result in sudden cardiac death. In this study, we proposed a ventricular fibrillation detection algorithm based on support vector machine classifier, which could offer benefits to reduce the teaming costs as well as good classification performance. Before the extraction of input features, raw ECG signal was applied to preprocessing procedures, as like wavelet transform based bandpass filtering, R peak detection and segment assignment for feature extraction. We selected input features which of some are related to the rhythm information and of others are related to wavelet coefficients that could describe the morphology of ventricular fibrillation well. Parameters for SVM classifier, C and ${\alpha}$, were chosen as 10 and 1 respectively by trial and error experiments. Each average performance for normal sinus rhythm ventricular tachycardia and VF, was 98.39%, 96.92% and 99.88%. And, when the VF detection performance of SVM classifier was compared to that of multi-layer perceptron and fuzzy inference methods, it showed similar or higher values. Consequently, we could find that the proposed input features and SVM classifier would one of the most useful algorithm for VF detection.

• Journal of the Korean Geotechnical Society
• /
• v.15 no.1
• /
• pp.99-112
• /
• 1999

By using 92 values of lateral earth pressure coefficient(K) measured in Korea, the tendency of K with varying depth is analyzed and compared with the range of K defined by Hoek and Brown. The horizontal stress is generally larger than the vertical stress in Korea : About 84 % of K values are above 1. In this study, the theory of elasto-plasticity is applied to analyze the variation of K values, and the results are compared with those of numerical analysis. This reveals that the erosion, sedimentation and weathering of earth crust are important factors in the determination of K values. Surface erosion, large lateral pressure and good rock mass increase the K values, but sedimentation decreases the K values. This study enable us to analyze the effects of geological processes on the K values, especially at shallow depth where underground excavation takes place. A neural network expert system using multi-layer back-propagation algorithm is developed to predict the K values. The neural network model has a correlation coefficient above 0.996 when it is compared with measured data. The comparison with 9 measured data which are not included in the back-propagation learning has shown an average inference error of 20% and the correlation coefficient above 0.95. The expert system developed in this study can be used for reliable determination of K values.

• Journal of the Institute of Electronics Engineers of Korea CI
• /
• v.39 no.3
• /
• pp.18-31
• /
• 2002

In this study, we introduce a concept of advanced neurofuzzy polynomial networks(ANFPN), a hybrid modeling architecture combining neurofuzzy networks(NFN) and polynomial neural networks(PNN). These networks are highly nonlinear rule-based models. The development of the ANFPN dwells on the technologies of Computational Intelligence(Cl), namely fuzzy sets, neural networks and genetic algorithms. NFN contributes to the formation of the premise part of the rule-based structure of the ANFPN. The consequence part of the ANFPN is designed using PNN. At the premise part of the ANFPN, NFN uses both the simplified fuzzy inference and error back-propagation learning rule. The parameters of the membership functions, learning rates and momentum coefficients are adjusted with the use of genetic optimization. As the consequence structure of ANFPN, PNN is a flexible network architecture whose structure(topology) is developed through learning. In particular, the number of layers and nodes of the PNN are not fixed in advance but is generated in a dynamic way. In this study, we introduce two kinds of ANFPN architectures, namely the basic and the modified one. Here the basic and the modified architecture depend on the number of input variables and the order of polynomial in each layer of PNN structure. Owing to the specific features of two combined architectures, it is possible to consider the nonlinear characteristics of process system and to obtain the better output performance with superb predictive ability. The availability and feasibility of the ANFPN are discussed and illustrated with the aid of two representative numerical examples. The results show that the proposed ANFPN can produce the model with higher accuracy and predictive ability than any other method presented previously.

• The Journal of Korea Institute of Information, Electronics, and Communication Technology
• /
• v.12 no.5
• /
• pp.467-474
• /
• 2019

In this paper, we propose a novel method to improve FPS while maintaining the accuracy of YOLO v2 model in NVIDIA Jetson TX1. In general, in order to reduce the amount of computation, a conversion to an integer operation or reducing the depth of a network have been used. However, the accuracy of recognition can be deteriorated. So, we use methods to reduce computation and memory consumption through adjustment of the filter size and integrated computation of the network The first method is to replace the $3{\times}3$ filter with a $1{\times}1$ filter, which reduces the number of parameters to one-ninth. The second method is to reduce the amount of computation through CBR (Convolution-Add Bias-Relu) among the inference acceleration functions of TensorRT, and the last method is to reduce memory consumption by integrating repeated layers using TensorRT. For the simulation results, although the accuracy is decreased by 1% compared to the existing YOLO v2 model, the FPS has been improved from the existing 3.9 FPS to 11 FPS.

• Journal of the Korea Society of Computer and Information
• /
• v.27 no.12
• /
• pp.59-68
• /
• 2022

This paper presents a method for 1:1 verification by comparing the similarity between the given real product image and the drawing image. The proposed method combines two existing CNN-based deep learning models to construct a Siamese Network. After extracting the feature vector of the image through the FC (Fully Connected) Layer of each network and comparing the similarity, if the real product image and the drawing image (front view, left and right side view, top view, etc) are the same product, the similarity is set to 1 for learning and, if it is a different product, the similarity is set to 0. The test (inference) model is a deep learning model that queries the real product image and the drawing image in pairs to determine whether the pair is the same product or not. In the proposed model, through a comparison of the similarity between the real product image and the drawing image, if the similarity is greater than or equal to a threshold value (Threshold: 0.5), it is determined that the product is the same, and if it is less than or equal to, it is determined that the product is a different product. The proposed model showed an accuracy of about 71.8% for a query to a product (positive: positive) with the same drawing as the real product, and an accuracy of about 83.1% for a query to a different product (positive: negative). In the future, we plan to conduct a study to improve the matching accuracy between the real product image and the drawing image by combining the parameter optimization study with the proposed model and adding processes such as data purification.

• Journal of IKEEE
• /
• v.28 no.3
• /
• pp.474-478
• /
• 2024

In this paper, we proposes a deep learning network for quality inspection in a multi-camera inline inspection system for pharmaceutical containers. The proposed deep learning network is specifically designed for pharmaceutical containers by using data produced in real manufacturing environments, leading to more accurate quality inspection. Additionally, the use of an inline-capable deep learning network allows for an increase in inspection speed. The development of the deep learning network for quality inspection in the multi-camera inline inspection system consists of three steps. First, a dataset of approximately 10,000 images is constructed from the production site using one line camera for foreign substance inspection and three area cameras for dimensional inspection. Second, the pharmaceutical container data is preprocessed by designating regions of interest (ROI) in areas where defects are likely to occur, tailored for foreign substance and dimensional inspections. Third, the preprocessed data is used to train the deep learning network. The network improves inference speed by reducing the number of channels and eliminating the use of linear layers, while accuracy is enhanced by applying PReLU and residual learning. This results in the creation of four deep learning modules tailored to the dataset built from the four cameras. The performance of the proposed deep learning network for quality inspection in the multi-camera inline inspection system for pharmaceutical containers was evaluated through experiments conducted by a certified testing agency. The results show that the deep learning modules achieved a classification accuracy of 99.4%, exceeding the world-class level of 95%, and an average classification speed of 0.947 seconds, which is superior to the world-class level of 1 second. Therefore, the effectiveness of the proposed deep learning network for quality inspection in a multi-camera inline inspection system for pharmaceutical containers has been demonstrated.

• Korean Journal of Ecology and Environment
• /
• v.38 no.1 s.110
• /
• pp.45-53
• /
• 2005

The environmental change of Yong-nup in Mt. Dae-Am, which is located at the northern part of Kangwon-Do, Korea, was assesed with peat sedimentary carbon and nitrogen isotope analysis. The surface layer of the peat (0 ${\sim}$ 5 cm) was 190 year BP, and the middle layers (30 ${\sim}$ 35 cm and 50 ${\sim}$ 55 cm) were 870 year BP and 1900 year BP, respectively. Bulk sedimentation rate was estimated to be about 0.4 mm $year^{-1}$ for 0 cm to 30 cm and 0.15 mm $year^{-1}$ for 35 cm to 50 cm. The $^{14}C$ age of the bottom sediment (75 ${\sim}$ 80 cm) collected and measured in this study was about 1900 year BP, although it was measured that the $^{14}C$ of the lowest bottom sediment in Yong-nup was 4105 ${\pm}$ 175 year BP (GX-23200). Since the $^{14}C$ ages for 50 ${\sim}$ 55 cm and 75 ${\sim}$ 80 cm layers were almost the same as 1890 ${\pm}$ 80 fear BP (NUTA 5364) and 1850 ${\pm}$ 90 year BP (NUTA 5462), respectively, we have estimated that the deep layers (55 ${\sim}$ 80 cm) in the high moor were the original forest soil. The low organic C and N contents in the deeper layers supported the inference. The sediment of 50 ${\sim}$ 55 cm layer contains much sandy material and showed very low organic content, suggesting the erosion (flooding) from the surrounding area. In this context, the Yong-nup, high moor, of Mt. Dae-Am, might have developed to the sampling site at about 1900 year BP. The ${\delta}^{13}C$ values of organic carbon and the ${\delta}^{15}N$ values of total nitrogen in the peat sediments fluctuated with the depths. The profile of ${\delta}^{13}C$ may indicate that the Yong-nup of Mt. Dae-Am have experienced the dry-wet and cool-warm period cycles during the development of the high moor. The ${\delta}^{15}N$ may indicate that the nitrogen cycling in the Yong-nup have changed from the closed (regeneration depending) system to the open (rain $NO_3\;^-$ and $N_2$ fixation depending) system during the development of the high moor.