• Journal of the Korea Computer Graphics Society
• /
• v.28 no.2
• /
• pp.1-9
• /
• 2022

In this paper, we propose a vector-based augmentation technique that can generate data required for crack detection and a ConvNet(Convolutional Neural Network) technique that can learn it. Detecting cracks quickly and accurately is an important technology to prevent building collapse and fall accidents in advance. In order to solve this problem with artificial intelligence, it is essential to obtain a large amount of data, but it is difficult to obtain a large amount of crack data because the situation for obtaining an actual crack image is mostly dangerous. This problem of database construction can be alleviated with elastic distortion, which increases the amount of data by applying deformation to a specific artificial part. In this paper, the improved crack pattern results are modeled using ConvNet. Rather than elastic distortion, our method can obtain results similar to the actual crack pattern. By designing the crack data augmentation based on a vector, rather than the pixel unit used in general data augmentation, excellent results can be obtained in terms of the amount of crack change. As a result, in this paper, even though a small number of crack data were used as input, a crack database can be efficiently constructed by generating various crack directions and patterns.

• Journal of Semiconductor Engineering
• /
• v.1 no.1
• /
• pp.31-37
• /
• 2020

We present an approximate DRAM architecture for energy-efficient deep learning. Our key premise is that by bounding memory errors to non-critical information, we can significantly reduce DRAM refresh energy without compromising recognition accuracy of deep neural networks. To validate the key premise, we make extensive Monte-Carlo simulations for several well-known convolutional neural networks such as LeNet, ConvNet and AlexNet with the input of MINIST, CIFAR-10, and ImageNet, respectively. We assume that the highest-order 8-bits (in single precision) and 4-bits (in half precision) are protected from retention errors under the proposed architecture and then, randomly inject bit-errors to unprotected bits with various bit-error-rates. Here, recognition accuracies of the above convolutional neural networks are successfully maintained up to the 10^-5-order bit-error-rate. We simulate DRAM energy during inference of the above convolutional neural networks, where the proposed architecture shows the possibility of considerable energy saving up to 10 ~ 37.5% of total DRAM energy.

• Korean Journal of Remote Sensing
• /
• v.38 no.5_2
• /
• pp.793-810
• /
• 2022

Synthetic Aperture Radar (SAR) is considered to be suitable for near real-time inundation monitoring. The distinctly different intensity between water and land makes it adequate for waterbody detection, but the intrinsic speckle noise and variable intensity of SAR images decrease the accuracy of waterbody detection. In this study, we suggest two modules, named 'morphology module' and 'edge-enhanced module', which are the combinations of pooling layers and convolutional layers, improving the accuracy of waterbody detection. The morphology module is composed of min-pooling layers and max-pooling layers, which shows the effect of morphological transformation. The edge-enhanced module is composed of convolution layers, which has the fixed weights of the traditional edge detection algorithm. After comparing the accuracy of various versions of each module for U-Net, we found that the optimal combination is the case that the morphology module of min-pooling and successive layers of min-pooling and max-pooling, and the edge-enhanced module of Scharr filter were the inputs of conv9. This morphologic and edge-enhanced U-Net improved the F1-score by 9.81% than the original U-Net. Qualitative inspection showed that our model has capability of detecting small-sized waterbody and detailed edge of water, which are the distinct advancement of the model presented in this research, compared to the original U-Net.

• IEMEK Journal of Embedded Systems and Applications
• /
• v.11 no.4
• /
• pp.219-225
• /
• 2016

Visual object tracking is one of the key tasks in computer vision. Robust trackers should address challenging issues such as fast motion, deformation, occlusion and so on. In this paper, we therefore propose a visual object tracking method that exploits inter-frame correlations of convolutional feature maps in Convolutional Neural Net (ConvNet). The proposed method predicts the location of a target by considering inter-frame spatial correlation between target location proposals in the present frame and its location in the previous frame. The experimental results show that the proposed algorithm outperforms the state-of-the-art work especially in hard-to-track sequences.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.17 no.9
• /
• pp.2458-2482
• /
• 2023

Medical image segmentation techniques based on convolution neural networks indulge in feature extraction triggering redundancy of parameters and unsatisfactory target localization, which outcomes in less accurate segmentation results to assist doctors in diagnosis. In this paper, we propose a multi-level semantic-rich encoding-decoding network, which consists of a Pooling-Conv-Former (PCFormer) module and a Cbam-Dilated-Transformer (CDT) module. In the PCFormer module, it is used to tackle the issue of parameter explosion in the conservative transformer and to compensate for the feature loss in the down-sampling process. In the CDT module, the Cbam attention module is adopted to highlight the feature regions by blending the intersection of attention mechanisms implicitly, and the Dilated convolution-Concat (DCC) module is designed as a parallel concatenation of multiple atrous convolution blocks to display the expanded perceptual field explicitly. In addition, MultiHead Attention-DwConv-Transformer (MDTransformer) module is utilized to evidently distinguish the target region from the background region. Extensive experiments on medical image segmentation from Glas, SIIM-ACR, ISIC and LGG demonstrated that our proposed network outperforms existing advanced methods in terms of both objective evaluation and subjective visual performance.

• Journal of KIISE
• /
• v.44 no.1
• /
• pp.44-50
• /
• 2017

Pedestrian detection, an important component of autonomous driving and driving assistant system, has been extensively studied for many years. In particular, image based pedestrian detection methods such as Hierarchical classifier or HOG and, deep models such as ConvNet are well studied. The evaluation score has increased by the various methods. However, pedestrian detection requires high sensitivity to errors, since small error can lead to life or death problems. Consequently, further reduction in pedestrian detection error rate of autonomous systems is required. We proposed a new method to detect pedestrians and reduce the error rate by using the Faster R-CNN with new developed pedestrian training data sets. Finally, we compared the proposed method with the previous models, in order to show the improvement of our method.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.16 no.7
• /
• pp.2407-2424
• /
• 2022

Recently, Transformer has made great progress in video retrieval tasks due to its high representation capability. For the structure of a Transformer, the cascaded self-attention modules are capable of capturing long-distance feature dependencies. However, the local feature details are likely to have deteriorated. In addition, increasing the depth of the structure is likely to produce learning bias in the learned features. In this paper, an improved Transformer structure named TransDCS (Transformer with Dynamic Convolution and Shortcut) is proposed. A Multi-head Conv-Self-Attention module is introduced to model the local dependencies and improve the efficiency of local features extraction. Meanwhile, the augmented shortcuts module based on a dual identity matrix is applied to enhance the conduction of input features, and mitigate the learning bias. The proposed model is tested on MSRVTT, LSMDC and Activity-Net benchmarks, and it surpasses all previous solutions for the video-text retrieval task. For example, on the LSMDC benchmark, a gain of about 2.3% MdR and 6.1% MnR is obtained over recently proposed multimodal-based methods.

• Nuclear Engineering and Technology
• /
• v.52 no.7
• /
• pp.1417-1428
• /
• 2020

The design of a nuclear reactor core requires basic thermal-hydraulic information concerning the heat transfer regime at which onset of nucleate boiling (ONB) will occur, the pressure drop and flow rate through the reactor core, the temperature and power distributions in the reactor core, the departure from nucleate boiling (DNB), the condition for onset of flow instability (OFI), in addition to, the critical velocity beyond which the fuel elements will collapse. These values depend on coolant velocity, fuel element geometry, inlet temperature, flow direction and water column above the top of the reactor core. Enough safety margins to ONB, DNB and OFI must-emphasized. A heat transfer package is used for calculating convection heat transfer coefficient in single phase turbulent, transition and laminar regimes. The main objective of this paper is to study the possibility of power upgrading of WWR-S research reactor from 2 to 10 MW_th. This study presents a one-dimensional mathematical model (axial direction) for steady-state thermal-hydraulic design and analysis of the upgraded WWR-S reactor in which two types of plate fuel elements are employed. FOR-CONV computer program is developed for the needs of the power upgrading of WWR-S reactor up to 10 MW_th.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.13 no.6
• /
• pp.3182-3198
• /
• 2019

Vision-based Human Pose Estimation has been considered as one of challenging research subjects due to problems including confounding background clutter, diversity of human appearances and illumination changes in scenes. To tackle these problems, we propose to use a new multi-stage convolution machine for estimating human pose. To provide better heatmap prediction of body joints, the proposed machine repeatedly produces multiple predictions according to stages with receptive field large enough for learning the long-range spatial relationship. And stages are composed of various modules according to their strategic purposes. Pyramid stacking module and dilation module are used to handle problem of human pose at multiple scales. Their multi-scale information from different receptive fields are fused with concatenation, which can catch more contextual information from different features. And spatial and channel information of a given input are converted to gating factors by squeezing the feature maps to a single numeric value based on its importance in order to give each of the network channels different weights. Compared with other ConvNet-based architectures, we demonstrated that our proposed architecture achieved higher accuracy on experiments using standard benchmarks of LSP and MPII pose datasets.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.12 no.11
• /
• pp.5555-5567
• /
• 2018

State-of-art instance segmentation networks are successful at generating 2D segmentation mask for region proposals with highest classification score, yet 3D object segmentation task is limited to geocentric embedding or detector of Sliding Shapes. To this end, we propose an amodal 3D instance segmentation network called A3IS-CNN, which extends the detector of Deep Sliding Shapes to amodal 3D instance segmentation by adding a new branch of 3D ConvNet called A3IS-branch. The A3IS-branch which takes 3D amodal ROI as input and 3D semantic instances as output is a fully convolution network(FCN) sharing convolutional layers with existing 3d RPN which takes 3D scene as input and 3D amodal proposals as output. For two branches share computation with each other, our 3D instance segmentation network adds only a small overhead of 0.25 fps to Deep Sliding Shapes, trading off accurate detection and point-to-point segmentation of instances. Experiments show that our 3D instance segmentation network achieves at least 10% to 50% improvement over the state-of-art network in running time, and outperforms the state-of-art 3D detectors by at least 16.1 AP.