• Journal of Broadcast Engineering
• /
• v.26 no.1
• /
• pp.70-78
• /
• 2021

Estimating 3D information from a single image is one of the essential problems in numerous applications. Since a 2D image inherently might originate from an infinite number of different 3D scenes, thus 3D reconstruction from a single image is notoriously challenging. This challenge has been overcame by the advent of recent deep convolutional neural networks (CNNs), by modeling the mapping function between 2D image and 3D information. However, to train such deep CNNs, a massive training data is demanded, but such data is difficult to achieve or even impossible to build. Recent trends thus aim to present deep learning techniques that can be trained in a weakly-supervised manner, with a meta-data without relying on the ground-truth depth data. In this article, we introduce recent developments of weakly-supervised deep learning technique, especially categorized as scene 3D reconstruction and object 3D reconstruction, and discuss limitations and further directions.

• Journal of Digital Policy
• /
• v.1 no.2
• /
• pp.53-59
• /
• 2022

In this paper, we proposed a method using sub-edge information extracted through a hierarchical structure in the process of generating super resolution based on a single image. In order to improve the quality of super resolution, it is necessary to clearly distinguish the shape of each area while clearly expressing the boundary area in the image. The proposed method assists edge information of the image in deep learning based super resolution method to create an improved super resolution result while maintaining the structural shape of the boundary region, which is an important factor determining the quality in the super resolution process. In addition to the group convolution structure for performing deep learning based super resolution, a separate hierarchical edge accumulation extraction process based on high-frequency band information for sub-edge extraction is proposed, and a method of using it as an auxiliary feature is proposed. Experimental results showed about 1% performance improvement in PSNR and SSIM compared to the existing super resolution.

• Journal of Digital Policy
• /
• v.2 no.3
• /
• pp.1-7
• /
• 2023

In this paper, I proposes a method for performing single image super resolution by separating texture-spatial domains and then classifying features based on detailed information. In CNN (Convolutional Neural Network) based super resolution, the complex procedures and generation of redundant feature information in feature estimation process for enhancing details can lead to quality degradation in super resolution. The proposed method reduced procedural complexity and minimizes generation of redundant feature information by splitting input image into two channels: texture and spatial. In texture channel, a feature refinement process with step-wise skip connections is applied for detail restoration, while in spatial channel, a method is introduced to preserve the structural features of the image. Experimental results using proposed method demonstrate improved performance in terms of PSNR and SSIM evaluations compared to existing super resolution methods, confirmed the enhancement in quality.

• Journal of Digital Policy
• /
• v.3 no.3
• /
• pp.9-14
• /
• 2024

In this paper, proposes a method using a multi block structure composed of residual blocks with adaptive weights to improve the quality of results in single image super resolution. In the process of generating super resolution images using deep learning, the most critical factor for enhancing quality is feature extraction and application. While extracting various features is essential for restoring fine details that have been lost due to low resolution, issues such as increased network depth and complexity pose challenges in practical implementation. Therefore, the feature extraction process was structured efficiently, and the application process was improved to enhance quality. To achieve this, a multi block structure was designed after the initial feature extraction, with nested residual blocks inside each block, where adaptive weights were applied. Additionally, for final high resolution reconstruction, a multi kernel image reconstruction process was employed, further improving the quality of the results. The performance of the proposed method was evaluated by calculating PSNR and SSIM values compared to the original image, and its superiority was demonstrated through comparisons with existing algorithms.

• Proceedings of the Korea Contents Association Conference
• /
• 2003.05a
• /
• pp.416-420
• /
• 2003

In this paper, we implemented a dual-resolution image database system for the fast naked-eye retrieval using interpolation. This system can solve two conventional problems : a blocking noise at zoom-out image in single high resolution method and a big storage to store in simple dual-resolution image database system. The proposed method makes a subsampled image by subsampling a original image, and then a interpolated image of it using interpolation. After that, a hybrid dual-resolution image database is composed based on the differential image between the interpolated image and the original image. Experimental results of simulating through 60 sample images shows that the proposed method is 0.011 second faster than simple high-resolution method in the retrieval time - one is 0.003 second, the other is 0.014 second, respectively. Also, that improves 14.7% more than simple dual-resolution method in the stored size - one is 19,821 byte, the other is 16,910 byte, respectively.

• Journal of the Korean Society of Surveying, Geodesy, Photogrammetry and Cartography
• /
• v.22 no.4
• /
• pp.359-365
• /
• 2004

Data integration techniques are becoming increasing1y important for conquering a limitation with a single data. Image fusion which improves the spatial and spectral resolution from a set of images with difffrent spatial and spectral resolutions, and image registration which matches two images so that corresponding coordinate points in the two images correspond to the same physical region of the scene being imaged have been researched. In this paper, we compared with six image fusion methods(Brovey, IHS, PCA, HPF, CN, and MWD) with panchromatic and multispectral images of IKONOS and developed the registration method for applying to SPOT-5 satellite image and RADARSAT SAR satellite image. As the result of tests on image fusion and image registration, we could find that MWD and HPF methods showed the good result in term of visual comparison analysis and statistical analysis. And we could extract patches which depict detailed topographic information from SPOT-5 and RADARSAT and obtain encouraging results in image registration.

• Korean Journal of Remote Sensing
• /
• v.20 no.2
• /
• pp.125-137
• /
• 2004

In case of using image sequence taken from a moving camera along a road in an urban area, general video mosaicing technique based on a single baseline cannot create 2-D image mosaics. To solve the drawback, this paper proposed a new image mosaicing technique through 3-D multi-baselines that can create image mosaics in 3-D space. The core of the proposed method is that each image frame has a dependent baseline, an equation of first order, calculated by using ground control point (GCP) of optical flows. The proposed algorithm consists of 4 steps: calculation of optical flows using hierarchical strategy, calculation of camera exterior orientation, determination of multi-baselines, and seamless image mosaics. This paper realized and showed the proposed algorithm that can create efficient image mosaics in 3-D space from real image sequence.

• Journal of Broadcast Engineering
• /
• v.27 no.3
• /
• pp.369-377
• /
• 2022

Recent deep learning-based face synthesis research shows the result of generating a realistic face including overall style or elements such as hair, glasses, and makeup. However, previous methods cannot create a face at a very detailed level, such as the microstructure of the skin. In this paper, to overcome this limitation, we propose a technique for synthesizing a more realistic facial image from a single face label image by controlling the types and intensity of skin microelements. The proposed technique uses Pix2PixHD, an Image-to-Image Translation method, to convert a label image showing the facial region and skin elements such as wrinkles, pores, and redness to create a facial image with added microelements. Experimental results show that it is possible to create various realistic face images reflecting fine skin elements corresponding to this by generating various label images with adjusted skin element regions.

• The KIPS Transactions:PartB
• /
• v.17B no.5
• /
• pp.363-374
• /
• 2010

In this paper, we propose the method for extracting Homogeneity Threshold($H_T$) and for segmenting homogeneous regions by USRG(Unseeded Region Growing) with $H_T$. The $H_T$ is a criterion to distinguish homogeneity in neighbor pixels and is computed automatically from the original image by proposed method. Theoretical background for proposed method is based on the Otsu's single level threshold method. The method is used to divide a small local part of original image int o two classes and the sum($\sigma_c$) of standard deviations for the classes to satisfy special conditions for distinguishing as different regions from each other is used to compute $H_T$. To find validity for proposed method, we compare the original image with the image that is regenerated with only the segmented homogeneous regions and show up the fact that the difference between two images is not exist visually and also present the steps to regenerate the image in order the size of segmented homogeneous regions and in order the intensity that includes pixels. Also, we show up the validity of proposed method with various results that is segmented using the homogeneity thresholds($H^*_T$) that is added a coefficient ${\alpha}$ for adjusting scope of $H_T$. We expect that the proposed method can be applied in various fields such as visualization and animation of natural image, anatomy and biology and so on.

• Journal of the Korean Society of Radiology
• /
• v.14 no.7
• /
• pp.1003-1011
• /
• 2020

This study aims to compare and evaluate the image differences between single and dual sources in applying a technique to reduce metal artifacts using dual energy CT. Discovery CT 256 (GE, USA) as a single source device and Somatom Definition Flash (Siemens Health Care, Forchheim, Germany) as a dual source device. The self-made phantom (pigs with medical titanium screws inserted) was quantitative and qualitatively evaluated under the same conditions by varying the dose under the same conditions using a dual energy CT. The evaluation method was compared by measuring SNR for metal artifacts (scattering, stripe) generated by metal inserts, divided around bones and around tissues. There was a difference in images in the method of reducing metal artifacts between single-source and dual-source devices. In a single source device, the linearized prosthesis by metal implantation showed a greater decrease than the image obtained from a double source device, and the surrounding tissue was well observed without interference from the artifact. In dual-source devices, scattering and stripe artifacts caused by metal inserts decreased more than on a single source device, and signals from adjacent tissues surrounding the metal implant were well observed without diminishing. If the examination is conducted separately between single source and dual source devices depending on whether the area to which the patient is intended to be viewed during the examination is adjacent to the metal insert or the total tissue surrounding the metal insert, it is believed that diagnostic helpful images can be obtained.