https://scholar.gist.ac.kr/handle/local/7910 2025-12-08T05:03:40Z https://scholar.gist.ac.kr/handle/local/19891 Title: Wafer-Scale, Quasi-Random Thin-film Filters for Reconstructive Spectrometer Applications Author(s): Serim Kim Abstract: The miniaturization of the spectrometer is essential for commercial applications. The reconstructive spectrometer is less affected by the relationship between size and resolution than conventional spectrometers. The photonic structure (i.e., nanowire, and thin film filter) of reconstructive spectrometer act as response function, and restore the original signal by solving the linear equation. However, it is difficult to commercialize due to complex fabrications and challenges to mass production. In this study, we provide simple wafer-scale fabrication method by adopting multi-stack linear variable deposition.The proposed reconstruction spectrometer restores the original optical signal with relatively good accuracy and a low mean square error ( 5 ×10−4). The reconstructive spectrometer is compatible with CMOS sensors and can be mass-production due to a small number of fabrication processes, so it is expected to be applied to various fields such as medicine, food, and mobile integration. 2022-12-31T15:00:00Z https://scholar.gist.ac.kr/handle/local/19887 Title: Visual Attention Network with 1D Large Kernel for Speaker Verification Author(s): YECHAN YU Abstract: 본 논문은 화자 검증을 위한 1차원 넓은 커널을 갖는 비주얼 주의집중 네트워크 MFAVAN 을 제안한다. 우리는 1차원 넓은 커널을 효율적으로 사용하기 위해서, 이를 3가지의 컴볼루션 모듈로 분해하여 사용한다. 1차원 넓은 커널은 지역성과 전역성 특징을 동시의 추출할 뿐 아니라 채널 별 적응성을 보장하고 계산 복잡도 측면에서 개선을 보였다. 결과적으로, 우리의 제안된 모델은 VoxCeleb1-O 평가 데이터의 대해서 트랜스포터 기반 모델인 MFA-Transformer 보다 0.49% 향상된 EER 성능을 보였다. 또한 3 종류 의 VoxCeleb1 평가 데이터에 대해서 주의 집중 모델중 최첨단 모델인 MFA-Conformer 보다 절반 정도의 모델 파라미터로 거의 동일한 성능을 보였다.|In this paper, we propose a MFA-VAN which convert self-attention to Visual Attention Network (VAN) with 1D Large Kernel Attention (LKA) for Speaker Verification. The proposed model effectively utilizes 1D Large Kernel Attention, which decompose three type of convolution, i.e., depth-wise Convolution, Depth-wise Dilated Convolution, Point-wise Convolution. Large Kernel Attention not only extract local and global feature simultaneously, it but also has a channel adaptability and better computational complexity better than self-attention. As a result, our proposed model show a improvement of performance better by EER of 0.49% than MFA-transformer which transformer-based model in VoxCeleb1- O evaluation set. In addition, Evaluation results for 3 types of evaluation set and a variance on frame lengths show similar performance with about half the parameters of the existing state-of-the-art model, MFA-Conformer. 2021-12-31T15:00:00Z https://scholar.gist.ac.kr/handle/local/19885 Title: Via-hole etching process for back-contact InGaP/GaAs double heterojunction solar cells Author(s): Jiseon Yoo Abstract: High tech industry such as the unmanned aerial vehicles and drones develops, GaAs thin film solar cells that can be attached to streamlined form is studied. Backside contact technology is employed to improve efficiency of the InGaP/GaAs double heterojunction solar cell by making both the anode and cathode electrodes on the backside of the solar cells. It offers the advantages of easy module interconnect, homogeneous surface look, and the higher efficiency. To apply it to the double or triple-junction tandem solar cells, via-hole etching and through via metallization processes are required. A three-step etching process of via-hole for the inverted InGaP/GaAs double heterojunction back-contact solar cells has been developed. Back-contact solar cells are the way to increase the efficiency of the solar cells by eliminating the shading effect caused by the front electrode. In this paper, the etching of via-hole in the InGaP/GaAs double heterostructure has been developed using the three-step etching process based on a combination of two-step dry etching and wet etching realize back-contact solar cell. The via-holes with 3.2 µm depth were fabricated by process combined with dry and wet etching. Two different plasma chemistries such BCl3/Cl2 and CH4/H2 were involved using Inductively Coupled Plasma reactive ion etching (ICP-RIE) and reactive ion etching (RIE), respectively. The pressure and gas ratio were changed to form Sa smooth and tapered sidewall of the via-hole. The remaining layers were removed through wet etching with HCl/H3PO4/H2O2 solution by varying temperatures from 20 ℃ to 40 ℃. After the dielectric insulating layer (SiNx) was deposited on the sidewall of the via-hole. The 3.47-thick Au film was filled in the via-hole to reach the front side n-contact layer. To transfer back-contact solar cell from GaAs substrate to flexible substrate, the ELO process and GaAs wet etching process were studied. The difference between ELO process with stress and ELO process without stress was compared, and GaAs wet etching conditions were optimized. 2021-12-31T15:00:00Z https://scholar.gist.ac.kr/handle/local/19879 Title: Utilization of Deep Learning based feature extraction for Improved Recommendation System Author(s): Khan Zeeshan Abstract: The world of global community groups, social media platforms, and business websites now offers a wealth of information on goods, people, and activities. This is resulting in an abundance of stuff that has to be handled effectively in order to get the needed information. According to the user’s preferred preferences, a recommendation system (RS) makes suggestions for pertinent goods to them. It handles a variety of user and item-related data. However, data sparsity is a problem for RSs. Deep examination of item contents is often done in RSs using deep learning approaches to provide exact suggestions. However, there is still need for future study into how to manage user reviews while concurrently doing item evaluations. In order to address the sparsity issue, a hybrid approach that simultaneously manages user and item information is put forward in this study. 2023-12-31T15:00:00Z