Implementations

Skill sets: Verilog HDL, PyTorch

Summary: In this work, we present the architecture for the hardware implementation of a fully-connected network, one of the deep learning networks, and also verify it with the weights trained with MNIST dataset and quantized by QAT (Quantization Aware Training). Finally, we compare the software inference performance of the CPU and GPU of a fully connected network implemented with the software framework of PyTorch and the inference performance of the AI hardware accelerator implemented using Verilog HDL.

Domestic conference paper by my students: 서정윤, 이종윤, 박성준, "Implementation and Evaluation of fully-connected network on Hardware using Verilog HDL," 한국정보기술학회 하계학술대회, 2024.
International conference paper by my students: J. Y. Lee, J. Y. Seo, S. J. Park, H. Lee, and Y. H. Lee*, "Hardware Accelerator for Real-time Image Classification," The 2024 International Conference on Computational Science and Computational Intelligence (CSCI), Las Vegas, USA, Dec. 11-13, 2024 [Scopus].

Skill sets: PyTorch

Summary: In this study, we analyze a two-stage real-time system for recognizing traffic lights composed of TLD and TLSR modules. We propose a network architecture for the TLSR module that is both lightweight and effective, implement a data preprocessing approach to optimize TLSR performance, and comprehensively evaluate the appropriate system parameters for the TLD module. Furthermore, we showcase a system created through the proposed network structure, trained with established system parameters and ratio-preserving zero padding. Our demonstration illustrates the capacity of our real-time traffic light detection system to accurately detect traffic lights in a reliable manner. The system can continually detect signals and identify their status even when a vehicle is in proximity to the traffic light, thereby proving the system's reliability.

International Journal written by my student: 최지환, "Real-Time Traffic Light Recognition with Lightweight State Recognition and Ratio-Preserving Zero Padding," Electronics, 2024

Skill sets: PyTorch, Robot Operating System (ROS)

Summary: Despite the advance of science technologies, finding people is still performed by human such as experienced specialist, which takes a lot of times (ex. Wanted, Missing child). Therefore, we present the efficient and autonomous solution using object detection and face recognition technologies in the field of deep learning. Finally, we expect that the proposed autonomous system can find a targeted person with a few of human resource.

Domestic conference written by my students: 서정윤, 박성준, "Implementation of Deep-learning based Specific Person Identification System using Facial Similarity," 한국통신학회 동계학술대회, 2024.

Skill sets: PyTorch

Summary: The number of breast cancer patients is consistently increasing, and accordingly, the lymph node metastasis of cancer is also on the rise. Existing methods have a critical problem: a physician manually determines whether breast cancer has spread to lymph nodes based on the analysis of X-rays and pathological slides. To solve the problem, our approach is that a deep learning network to automatically determine breast cancer lymphatic metastasis.

Domestic conference written by my student: 김영섭, "Prediction of Lymph Node Metastasis by Combining Causal Inference and Vision based Networks", 한국통신학회 동계학술대회, 2024.

Skill sets: PyTorch

Summary : 웨이퍼 맵에서 오류가 있는 다이들이 형성하는 결함 패턴별로 특정 공정 과정의 이상 징후와 연관되어 있기 때문에 결함의 패턴을 정확하고 신속하게 검출하는 기술은 오류가 없는 웨이퍼 맵 생산에 도움을 줄 수 있다. 본 연구에서는 웨이퍼 맵에 다양한 결함 패턴들이 혼재하고 있는 상황에서 결함 패턴별로 높은 정확도와 재현율을 보장하는 딥러닝 기반의 다중 결함 검출 기술을 구현한다. 해당 기술을 위해 벡터 표현 기반의 캡슐 네트워크를 사용하여 모든 결함 패턴에 대한 mAP가 특징 맵 기반의 네트워크들에 비해 최대 65% 의 성능 이득이 있음을 보인다.

Domestic Journal written by my students: 김미선, 최지환, "웨이퍼 맵의 다중 패턴 인지를 위한 벡터 표현 기반 캡슐네트워크 구현", 한국통신학회 논문지, 우수 논문 선정, 2023.

Skill sets: PyTorch

대회 개요: 최근 반도체 구조의 폭, 물성 등 정량적으로 Monitoring하는 계측 분야가 반도체 구조가 미세화, 복잡화되면서 더욱 중요해지고 있으며, 이 분야에 AI 알고리즘을 개발하고자 하는 시도가 반도체 제조사에서 다양하게 이루어지고 있습니다. 대표적인 반도체 계측 방식은 상부에서 촬영한 (Top-down) SEM (주사 전자 현미경, Scanning Electron Microscope) 영상을 활용하는 것으로, 구조별 2차원 정보인 폭/두께 계측으로 한정되어 사용되어 있으며, 현재 깊이를 계측하기 위해서 OCD (Optical Critical Dimension), TEM (Transmission Electron Microscope) 영상 등을 활용하고 있습니다.

Skill sets: PyTorch

Summary : 반도체 공정에서 발생하는 각 웨이퍼 결함은 공정 과정 중 특정 이상 징후와 연관이 되어 있다. 따라서 웨이퍼 맵에 있는 결함을 정확하고 신속하게 검출하여 공정 기술자에게 알려주는 것은 에러가 없는 고품질의 웨이퍼 생산에 지대한 영향을 미친다. 또한 공정이 복잡해짐에 따라 웨이퍼 맵에 다양한 결함들이 혼재하고 있는 경우도 발생할 수 있기 때문에 다중 결함 검출로의 확장이 가능한 출력단 설계 역시 중요하다. 따라서 본 연구에서는 기존의 연구들과는 달리 다중 결함 검출 기법 개발로의 확장에 용이한 다중 라벨 인지 손실 함수를 이용하여 출력단 설계를 하고 이를 바탕으로 기존의 다양한 딥 네트워크들을 학습하고 성능을 분석한다. 또한 네트워크가 인식한 결함 패턴의 위치를 파악하기 위해 XAI 기법인 GradCAM을 적용하였으며 추가적으로 네트워크의 패턴 위치 파악의 정밀도를 높이기 위해 GAIN 기법을 학습에 적용하였다.

Skill sets: Python, PyTorch, MATLAB, MATLAB-Python API​

Summary : In the 20th century, large-scale natural disasters have been steadily occurring in Haiti, Japan, and the United States. These large-scale natural disasters inevitably result in a huge number of victims. For the survival of the victims, it is essential to develop a system that can efficiently find victims. Drones are a good candidate for the development of an efficient and general victim search system. However, drones powered by batteries have a limited operating time. As such, with the installation of complex protocols comes the difficulty of greater battery consumption. Therefore, in this paper, to achieve a light and efficient victim search drone, we propose a system that detects the number of asynchronously transmitted signals from victims’ mobile devices. For this system, we propose a machine learning-based asynchronous stream-sensing network (‘ASensNet’), which is developed by combining several neural networks. In addition, to train our proposed network, we propose a framework that connects our MATLAB-based simulator and PyTorch-based network training code. Finally, extensive evaluation has proven that by outperforming existing mathematical methods our proposal performs well enough for use in a disaster situation.

Skill sets: ROS, Python, PyTorch

Summary : In the project,  I proposed a face-anonymizing drone patrol system. In this system, one person’s face in a video is transformed into a different face with facial components maintained. In order to construct our privacy-preserving system, we proposed a face-anonymizing approach and a training architecture by adopting the latest generative adversarial networks (GANs) frameworks. For the purpose of our system, we made a few modifications to the losses of these frameworks. Our face-anonymizing approach is evaluated with various public face-image and video dataset. Moreover, our system is evaluated with a customized drone consisting of a high-resolution camera, a companion computer, and a drone control computer. Finally, we confirm that our system can protect privacy-sensitive information with our face-anonymizing algorithm while preserving the performance of vision-based robot perception, i.e., simultaneous localization and mapping (SLAM). 

Skill sets: ROS, Python, TensorFlow​

Summary : In the project, "Privacy-Preserving Robot Vision with Anonymized Faces by Extreme Low Resolution", I took the role of making a patrol robot with the privacy-preserving robot vision scheme using the deep neural network proposed by my co-worker. I used Turtlebot 3 as the patrol robot and added an Nvidia Xavier and a ZED camera to it.  In addition, I implemented ORB SLAM2 which is one of the visual odometry schemes on the Xavier. I utilized ROS (Robot Operating System) to take images from the ZED camera, to put the images in the proposed neural network, to send the results from the network to the ORB SLAM2 module, and finally to transmit ORB-SLAM2 results to our ground station.

Skill sets: LAA-LTE MAC/PHY, Wi-Fi MAC/PHY, MATLAB​

Summary : This simulator implements the key MAC and PHY functionalities, and the new channel model between LAA-LTE and Wi-Fi, where MAC and PHY parameters are set based on the specifications of LAA-LTE and Wi-Fi.

Skill sets: USRP, Wi-Fi PHY, MATLAB

Summary : In this platform, we utilized two USRP N210s, implementing a Wi-Fi transmitter and an LAA-LTE transmitter with the proposed Wi-Fi frame detection. For comparison, another N210 was used to implement a conventional Wi-Fi transmitter with conventional Wi-Fi frame detection and energy detection. Each USRP was equipped with an XCVR2450 daughterboard and used the same 20 MHz channel in 5 GHz unlicensed bands.

Skill sets: LAA-LTE MAC, Wi-Fi MAC, MATLAB

Skill sets: TensorFlow, Python​

Summary : This project utilized the deep neural network for lung nodule detection. In this project, the proposed network is constructed by utilizing the 3D-CNN which is one of the most popular networks in the field of deep neural networks. 

Skill sets: NS-3 simulator, Wi-Fi MAC, video encoding