• Proceedings of the Korean Society of Broadcast Engineers Conference
• /
• 2020.11a
• /
• pp.25-28
• /
• 2020

Automatic segmentation of brain tissues such as WM, GM, and CSF from brain MRI scans is helpful for the diagnosis of many neurological disorders. Accurate segmentation of these brain structures is a very challenging task due to low tissue contrast, bias filed, and partial volume effects. With the aim to improve brain MRI segmentation accuracy, we propose an end-to-end convolutional based U-SegNet architecture designed with multi-scale kernels, which includes cascaded dilated convolutions for the task of brain MRI segmentation. The multi-scale convolution kernels are designed to extract abundant semantic features and capture context information at different scales. Further, the cascaded dilated convolution scheme helps to alleviate the vanishing gradient problem in the proposed model. Experimental outcomes indicate that the proposed architecture is superior to the traditional deep-learning methods such as Segnet, U-net, and U-Segnet and achieves high performance with an average DSC of 93% and 86% of JI value for brain MRI segmentation.

• Steel and Composite Structures
• /
• v.46 no.6
• /
• pp.823-837
• /
• 2023

This paper presents the seismic performance of seven beam-column joints with an eccentricity between beam depths under cyclic loadings. The failure modes of the panel zone were divided into two types. One was the shear force failure that appeared in the entire panel zone (SFEPZ), the other was the shear force failure that appeared in the partial panel zone (SFPPZ). Seven finite element models were established using multi-scale methods. Compared with the experimental specimens, the hysteretic loops exhibited a similar trend. The multi-scale models could accurately simulate the experimental results. Furthermore, the calculation formulas of yield and plastic shear capacity of unequal-depth joints with outer annular stiffener were proposed.

• Journal of Multimedia Information System
• /
• v.8 no.2
• /
• pp.101-110
• /
• 2021

In computer vision, single-image super resolution has been an area of research for a significant period. Traditional techniques involve interpolation-based methods such as Nearest-neighbor, Bilinear, and Bicubic for image restoration. Although implementations of convolutional neural networks have provided outstanding results in recent years, efficiency and single model multi-scalability have been its challenges. Furthermore, previous works haven't placed enough emphasis on real-number scalability. Interpolation-based techniques, however, have no limit in terms of scalability as they are able to upscale images to any desired size. In this paper, we propose a convolutional neural network possessing the advantages of the interpolation-based techniques, which is also efficient, deeming it suitable in practical implementations. It consists of convolutional layers applied on the low-resolution space, post-up-sampling along the end hidden layers, and additional layers on high-resolution space. Up-sampling is applied on a multiple channeled feature map via bicubic interpolation using a single model. Experiments on architectural structure, layer reduction, and real-number scale training are executed with results proving efficient amongst multi-scale learning (including scale multi-path-learning) based models.

• Journal of Mechanical Science and Technology
• /
• v.20 no.1
• /
• pp.110-124
• /
• 2006

The multi scale wavelet-Galerkin method implemented in an adaptive manner has an advantage of obtaining accurate solutions with a substantially reduced number of interpolation points. The method is becoming popular, but its numerical efficiency still needs improvement. The objectives of this investigation are to present a new numerical scheme to improve the performance of the multi scale adaptive wavelet-Galerkin method and to give detailed implementation procedure. Specifically, the subdomain technique suitable for multiscale methods is developed and implemented. When the standard wavelet-Galerkin method is implemented without domain subdivision, the interaction between very long scale wavelets and very short scale wavelets leads to a poorly-sparse system matrix, which considerably worsens numerical efficiency for large-sized problems. The performance of the developed strategy is checked in terms of numerical costs such as the CPU time and memory size. Since the detailed implementation procedure including preprocessing and stiffness matrix construction is given, researchers having experiences in standard finite element implementation may be able to extend the multi scale method further or utilize some features of the multiscale method in their own applications.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.15 no.12
• /
• pp.4420-4438
• /
• 2021

Image dehazing is an ill-posed problem which is far from being solved. Traditional image dehazing methods often yield mediocre effects and possess substandard processing speed, while modern deep learning methods perform best only in certain datasets. The haze removal effect when processed by said methods is unsatisfactory, meaning the generalization performance fails to meet the requirements. Concurrently, due to the limited processing speed, most dehazing algorithms cannot be employed in the industry. To alleviate said problems, a lightweight fast dehazing network based on a multiple scale-patch framework (MSP) is proposed in the present paper. Firstly, the multi-scale structure is employed as the backbone network and the multi-patch structure as the supplementary network. Dehazing through a single network causes problems, such as loss of object details and color in some image areas, the multi-patch structure was employed for MSP as an information supplement. In the algorithm image processing module, the image is segmented up and down for processed separately. Secondly, MSP generates a clear dehazing effect and significant robustness when targeting real-world homogeneous and nonhomogeneous hazy maps and different datasets. Compared with existing dehazing methods, MSP demonstrated a fast inference speed and the feasibility of real-time processing. The overall size and model parameters of the entire dehazing model are 20.75M and 6.8M, and the processing time for the single image is 0.026s. Experiments on NTIRE 2018 and NTIRE 2020 demonstrate that MSP can achieve superior performance among the state-of-the-art methods, such as PSNR, SSIM, LPIPS, and individual subjective evaluation.

• The Journal of Korean Institute of Communications and Information Sciences
• /
• v.39A no.8
• /
• pp.459-466
• /
• 2014

Fading effect of old pictures and printings is shown up differently according to the ink property, temperature, humidity, illuminants, and so on. Faded image enhancement techniques based on illuminant estimation are proposed such as the gray world algorithm and white patch retinex methods. However, conventional simple operators are not suitable for enhancing faded images because partial fading effect is appeared differently. Thus, this paper presents a color enhancement algorithm based on integrating correction coefficients for faded images. First, the proposed method adopts local process by using multi-scale average mask. The coefficients for each multi-scale average mask are obtained to apply the gray world algorithm. Then, integrating the coefficients with weights is performed to calculate correction ratio for red and blue channels in the gray world assumption. Finally, the enhanced image is obtained by applying the integrated coefficients to the gray world algorithm. In the experimental results, the proposed method reproduces better colors for both wholly and partially faded images compared with the previous methods.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.12 no.10
• /
• pp.4976-4994
• /
• 2018

The diffusion-based salient object detection methods have shown excellent detection results and more efficient computation in recent years. However, the current diffusion-based salient object detection methods still have disadvantage of detecting the object appearing at the image boundaries and different scales. To address the above mentioned issues, this paper proposes a multi-scale diffusion-based salient object detection algorithm with background and objectness seeds. In specific, the image is firstly over-segmented at several scales. Secondly, the background and objectness saliency of each superpixel is then calculated and fused in each scale. Thirdly, manifold ranking method is chosen to propagate the Bayessian fusion of background and objectness saliency to the whole image. Finally, the pixel-level saliency map is constructed by weighted summation of saliency values under different scales. We evaluate our salient object detection algorithm with other 24 state-of-the-art methods on four public benchmark datasets, i.e., ASD, SED1, SED2 and SOD. The results show that the proposed method performs favorably against 24 state-of-the-art salient object detection approaches in term of popular measures of PR curve and F-measure. And the visual comparison results also show that our method highlights the salient objects more effectively.

• Transactions of the Korean Society for Noise and Vibration Engineering
• /
• v.24 no.8
• /
• pp.621-627
• /
• 2014

In this paper, a possibility of controlling the motion of a floating wind turbine with the tuned liquid damper(TLD) is numerically investigated. First, motion of the scale model of a floating wind turbine without the TLD is predicted and its results are compared to the measured data. There are reasonably good agreements between two results, which confirms validity of the present numerical methods. Then, the effect of TLD is quantitatively assessed by comparing the prediction results for the floating wind turbine with and without the TLD. It is shown that the motion of the scale model derived by external forces can be reduced by using the TLD. On a basis of this result, a multi-layer TLD is proposed to generate larger reaction force of the TLD at the fixed target frequency. The motions of the scale model with the multi-layer TLDs are computed and compared with that of the single-layer TLD. It is shown that the multi-layer TLD generate stronger reaction force and thus more reduce the motion of the floating body than the single-layer TLD.

• Journal of Information Processing Systems
• /
• v.16 no.5
• /
• pp.1083-1094
• /
• 2020

Recognizing plant species based on leaf images is challenging because of the large inter-class variation and inter-class similarities among different plant species. The effective extraction of leaf descriptors constitutes the most important problem in plant leaf recognition. In this paper, a multi-scale angular description method is proposed for fast and accurate leaf recognition and retrieval tasks. The proposed method uses a novel scale-generation rule to develop an angular description of leaf contours. It is parameter-free and can capture leaf features from coarse to fine at multiple scales. A fast Fourier transform is used to make the descriptor compact and is effective in matching samples. Both support vector machine and k-nearest neighbors are used to classify leaves. Leaf recognition and retrieval experiments were conducted on three challenging datasets, namely Swedish leaf, Flavia leaf, and ImageCLEF2012 leaf. The results are evaluated with the widely used standard metrics and compared with several state-of-the-art methods. The results and comparisons show that the proposed method not only requires a low computational time, but also achieves good recognition and retrieval accuracies on challenging datasets.

• Journal of the Korea Academia-Industrial cooperation Society
• /
• v.19 no.5
• /
• pp.9-16
• /
• 2018

Images with lower illumination from the light source or with dark regions due to shadows, etc., can improve subjective image quality by using retinex-based image enhancement schemes. The retinex theory is a method that recognizes the relative lightness of a scene, rather than recognizing the brightness of the scene. The way the human visual system recognizes a scene in a specific position can be in one of several methods: single-scale retinex, multi-scale retinex, and multi-scale retinex with color restoration (MSRCR). The proposed method is based on the MSRCR method, which includes a color restoration step, which consists of three phases. In the first phase, the existing MSRCR method is applied. In the second phase, the dynamic range of the MSRCR output is adjusted according to its histogram. In the last phase, the proposed method transforms the retinex output value into the display dynamic range using a logarithm transformation function considering human visual system characteristics. Experimental results show that the proposed algorithm effectively increases the subjective image quality, not only in dark images but also in images including both bright and dark areas. Especially in a low lightness image, the proposed algorithm showed higher performance improvement than the conventional approaches.