• Proceedings of the Korean Society of Broadcast Engineers Conference
• /
• 2020.07a
• /
• pp.540-543
• /
• 2020

기존의 초해상도 딥러닝 기법은 모델의 깊이가 깊어지면서, 좋은 성능을 내지만 점점 더 복잡해지고 있고, 실제로 사용하는데 있어 많은 시간을 요구한다. 이를 해결하기 위해, 우리는 딥러닝 모델의 가중치를 양자화 하여 추론시간을 줄이고자 한다. 초해상도 모델은 feature extraction, non-linear mapping, reconstruction 세 부분으로 나누어져 있으며, 레이어 사이에 많은 skip-connection 이 존재하는 특징이 있다. 따라서 양자화 시 최종 성능 하락에 미치는 영향력이 레이어 별로 다르며, 이를 감안하여 강화학습으로 레이어 별 최적 bit 를 찾아 성능 하락을 최소화한다. 본 논문에서는 Skip-connection 이 많이 존재하는 MSRN 을 사용하였으며, 결과에서 feature extraction, reconstruction 부분과 블록 내 특정 위치의 레이어가 항상 높은 bit 를 가짐을 알 수 있다. 기존에 영상 분류에 한정되어 사용되었던 혼합 bit 양자화를 사용하여 초해상도 딥러닝 기법의 모델 사이즈를 줄인 최초의 논문이며, 제안 방법은 모바일 등 제한된 환경에 적용 가능할 것으로 생각된다.

• Proceedings of the Korea Information Processing Society Conference
• /
• 2024.05a
• /
• pp.592-595
• /
• 2024

엄청난 성능에도 불구하고, 심층 신경망은 예측결과에 대한 설명이 없는 블랙 박스로 작동한다는 비판을 받고 있다. 이러한 불투명한 표현은 신뢰성을 제한하고 모델의 대한 과학적 이해를 방해한다. 본 연구는 여러 개의 교사 신경망으로부터 설명 중심의 학생 신경망으로 지식 증류를 통해 해석 가능성을 향상시키는 것을 제안한다. 구체적으로, 인간이 정의한 개념 활성화 벡터 (CAV)를 통해 교사 모델의 개념 민감도를 방향성 도함수를 사용하여 계량화한다. 목표 개념에 대한 민감도 점수에 비례하여 교사 지식 융합을 가중치를 부여함으로써 증류된 학생 모델은 양호한 성능을 달성하면서 네트워크 논리를 해석으로 집중시킨다. 실험 결과, ResNet50, DenseNet201 및 EfficientNetV2-S 앙상블을 7 배 작은 아키텍처로 압축하여 정확도가 6% 향상되었다. 이 방법은 모델 용량, 예측 능력 및 해석 가능성 사이의 트레이드오프를 조화하고자 한다. 이는 모바일 플랫폼부터 안정성이 중요한 도메인에 걸쳐 믿을 수 있는 AI 의 미래를 여는 데 도움이 될 것이다.

• Journal of the Korea Concrete Institute
• /
• v.26 no.3
• /
• pp.239-245
• /
• 2014

This study analyzed a comprehensive database including 6331 ready-mixed concrete plant mixtures to quantitatively assess the environmental impact of concrete under mixture proportions variable according to the domestic region and season. The environmental impact indicator includes global warming, photochemical oxidant creation, abiotic resource depletion, acidification, eutrophication and human toxicity, which are determined from categorization, characterization, normalization and weighting process based on Korea lifecycle inventories. The determined environmental impact indicator was also normalized by concrete compressive strength ($f_{ck}$), which is defined as impact index, to calculate the environmental impact per unit strength of 1 MPa. The most common compressive strength of concrete used in the country is estimated to be 24 MPa and 27 MPa. For $f_{ct}$ of 24 MPa, the lowest environmental impact indicator is observed in Ulsan, whereas the highest region is Gwangju and Daegu. This difference according to domestic region is primarily resulted from by the replacement of different supplementary cementitious materials. Furthermore, the impact index of concrete with $f_{ck}$ of 24 MPa is higher by approximately 5% at wintertime than at summertime and standard season. The impact index gradually decreases with the increase of $f_{ck}$ up to 35 MPa, beyond which it remains constant.

• Journal of the Korean Geosynthetics Society
• /
• v.16 no.4
• /
• pp.93-101
• /
• 2017

In this study, a subjective weighting factors were awarded based on some indication of the difficulty of assessing the preconsolidation stress using traditional methods (Casagrande, Onitsuka et al., Silva, Becker et al., Janbu and Karlsrud methods) such as those proposed by Casagrande and Janbu using undisturbed sample obtained from Gwangyang dredged clay with high plasticity located in the southern area of Korean peninsular. These numbers only assess the relative ease of finding preconsolidation stress and say nothing regarding the accuracy of the value. The data were compared with measurements of undrained shear strength using strength incremental ratio, checking where or not the values are in the range of 0.25 to 0.35 (typical values of Korean marine clay) and analyzing standard deviation(degree of variability). The measurements of undrained shear strength were obtained from unconfined compression tests (UCT). When determining preconsolidation stress of Korean marine clay, at first, the work method proposed by Becker et al. and the bilogarithmic method proposed by Onitsuka et al. should be used. In addition, preconsolidation pressure should be estimated using the traditional Casagrande method as a basic of comparison.

• Geotechnical Engineering
• /
• v.3 no.1
• /
• pp.65-76
• /
• 1987

This study has been carried out as a basic course for the analysis of foundation deformations based on the Cap model using the finite element methods. Material parameters should firstly be determined in order to use the Cap model for numerical solution. Associated with the fact described above, a method determining the soil parameters is suggested using algorithm for numerical ana])isis from raw truly triaxial compression laboratory test data of Pueblo.Colorado sand by Zaman, et at. (1982) More specifically, the change of soil parameters Is thoroughly examined by weighting the data obtained from CTC and RTE tests, respectively. The main results obtained are as follows; 1. The obtained values of parameters (E, V and 2) are same irrespective of data obtained from various kind of tests. 2. The values of the other parameters are dependent on data used. 3. The determination of parameters is little affected by the weighting factor.

• 전자공학회논문지 IE
• /
• v.46 no.4
• /
• pp.31-41
• /
• 2009

The Image sensor of digital still camera has a limited dynamic range. In high dynamic range scenes, a picture often turns out to be underexposed or overexposed. Retinex algorithm based on the theory of the human visual perception is known to be effective contrast enhancement technique. However, it happens the unbalanced contrast enhancement which is the global contrast increased, and the local contrast decreased in the high dynamic range scenes. In this paper, to enhance the both global and local contrast, we propose the weight mapping retinex algorithm. Weight map is composed of the edge and exposure data which are extracted in the each retinex image, and merged with the retinex images in the fusion processing. According to the output picture comparing and numerical analysis, the proposed algorithm gives the better output image with the increased global and local contrast.

• Journal of the Korea Institute of Information and Communication Engineering
• /
• v.18 no.3
• /
• pp.727-734
• /
• 2014

Digital processing technologies are being studied in various areas of image compression, recognition and recovery. However, image deterioration still occurs due to the noises in the process of image acquisition, storage and transmission. Generally in the typical noises which are included in the images, there are Gaussian noise and the mixed noise where the Gaussian noise and impulse noise are overlapped and in order to remove these noises, various researches are being executed. In order to preserve the edge and effectively remove mixed noises, image recovery filter algorithm was suggested in this study which sets and processes the adaptive weight using the median values and average values after noise judgment. Additionally, existing methods were compared through simulations and PSNR(peak signal to noise ratio) was used as a judgment standard.

• Journal of the Korea Institute of Information and Communication Engineering
• /
• v.23 no.4
• /
• pp.401-406
• /
• 2019

In this paper, we propose a system for analyzing documents in XML format and in reports. The system extracts the paper or reports of keywords, shows them to the user, and then extracts the paragraphs containing the keywords by inputting the keywords that the user wants to search within the document. The system checks the frequency of keywords entered by the user, calculates weights, and removes paragraphs containing only keywords with the lowest weight. Also, we divide the refined paragraphs into 10 regions, calculate the importance of the paragraphs per region, compare the importance of each region, and inform the user of the main region having the highest importance. With these features, the proposed system can provide the main paragraphs with higher compression ratio than analyzing the papers or reports using the existing document analysis system. This will reduce the time required to understand the document.

• The KIPS Transactions:PartD
• /
• v.12D no.2 s.98
• /
• pp.219-232
• /
• 2005

With the rapid progress of computer technology in recent years, a digital video data has been used in many applications. As a result, various technologies have been introduced to discover knowledge embedded in a video database. However, few researches on data mining for a video database have been performed due to the unclear boundary of the information in a large amount of a video stream. This paper proposes an efficient generation method of a feature profile in a set of similar video data. To extract the information embedded in original video data efficiently, several refinement techniques are also presented. These methods include merging pixels, restricting preferred areas, removing noises under a minimum repeat factor, extracting important key frames, generating derived blocks and applying weights to dynamic and static areas differently. Finally, the performance of these methods is evaluated by comparing a result profile obtained by a data mining process with original video streams.

• The Journal of Korean Institute of Communications and Information Sciences
• /
• v.26 no.12A
• /
• pp.2059-2066
• /
• 2001

In this paper, we proposed a method to reduce the number of input pins in one-dimensional VLSI array for fractal image compression. We use quad-tree partition scheme and can reduce the number of the input pins up to 50% by sharing the domain\`s and the range\`s data input pins in the proposed VLSI array architecture. Also, we can reduce the input pins and simplify the internal operation circuit of the processing elements by eliminating a few number of bits of the least significant bits of the input data. We simulated using the 256$\times$256 and 512$\times$512 Lena images to verify performance of the proposed method. As the result of simulation, we can decompress the original image with about 32dB(PSNR) in spite of elimination of the least significant 2-bit in the original input data, and additionally reduce the number of input pins up to 25% compared to VLSI array sharing input pins of range and domain.