• 반도체공학회 논문지
• /
• 제2권1호
• /
• pp.17-20
• /
• 2024

본 논문은 전통적인 자원 집약적인 컴퓨터 비전 모델의 한계를 해결하기 위해 저전력 엣지 장치에 최적화된 새로운 경량 객체 검출 모델을 제안합니다. 제안된 검출기는 Single Shot Detector (SSD)에 기반하여 소형이면서도 견고한 네트워크를 설계하였고, 작은 객체를 효율적으로 감지하는 데 있어 효율성을 크게 향상시키도록 모델을 구성하였다. 이 모델은 주로 두 가지 구성요소로 구성되어 있습니다: Depthwise 와 Pointwise Convolution 레이어를 사용하여 효율적인 특징 추출을 위한 Light_Block, 그리고 작은 객체의 향상된 감지를 위한 Enhancer_Block 으로 나누었다. 우리의 모델은 300x480 의 이미지 크기를 가진 Udacity 주석이 달린 데이터셋에서 처음부터 훈련되었으며, 사전 훈련된 분류 가중치의 필요성을 제거하였다. 약 0.43M 의 파라미터로 5.5MB 만의 무게를 가진 우리의 검출기는 평균 정밀도 (mAP) 27.7%와 140 FPS 의 처리 속도를 달성하여, 정밀도와 효율성 모두에서 기존 모델을 능가하였다. 따라서, 본 논문은 추론의 정확성을 손상시키지 않으면서 엣지 장치를 위한 객체 검출에서의 효과적인 경량화를 보여주고 있다.

• 스마트미디어저널
• /
• 제13권6호
• /
• pp.44-53
• /
• 2024

현대사회의 기업들은 소셜 미디어, 제품 리뷰, 고객 피드백 등 다양한 영역에 걸쳐 소비자 의견을 정확하게 이해하는 것이 경쟁에서 성공하기 위한 주요 과제임을 강조하며 감성 분류를 점점 더 중요한 작업으로 채택하고 있다. 감성 분류는 소비자의 다양한 의견과 감성을 파악하여 제품이나 서비스 개선에 도움을 주는 이유로 많은 연구가 진행중이다. 감성 분류에서는 대규모 데이터 셋과 사전 학습된 언어 모델을 통한 미세 조정이 성능 향상에 중요한 역할을 한다. 최근 인공지능 기술의 발전으로 감성 분류 모델은 높은 성능을 보이고 있으며, 특히 ELECTRA 모델은 효율적인 학습 방법과 적은 컴퓨팅 자원을 통해 뛰어난 결과를 제공한다. 따라서 본 논문에서는 ELECTRA에서 한국어를 학습한 KoELECTRA 모델을 이용하여 다양한 데이터 셋에 대한 효율적인 미세 조정을 통해 감성 분류 성능을 향상하는 방법을 제안한다.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• 제15권10호
• /
• pp.3668-3684
• /
• 2021

Video action recognition is widely used in video surveillance, behavior detection, human-computer interaction, medically assisted diagnosis and motion analysis. However, video action recognition can be disturbed by many factors, such as background, illumination and so on. Two-stream convolutional neural network uses the video spatial and temporal models to train separately, and performs fusion at the output end. The multi segment Two-Stream convolutional neural network model trains temporal and spatial information from the video to extract their feature and fuse them, then determine the category of video action. Google Xception model and the transfer learning is adopted in this paper, and the Xception model which trained on ImageNet is used as the initial weight. It greatly overcomes the problem of model underfitting caused by insufficient video behavior dataset, and it can effectively reduce the influence of various factors in the video. This way also greatly improves the accuracy and reduces the training time. What's more, to make up for the shortage of dataset, the kinetics400 dataset was used for pre-training, which greatly improved the accuracy of the model. In this applied research, through continuous efforts, the expected goal is basically achieved, and according to the study and research, the design of the original dual-flow model is improved.

• 산업경영시스템학회지
• /
• 제46권3호
• /
• pp.123-138
• /
• 2023

The Fourth Industrial Revolution and sensor technology have led to increased utilization of sensor data. In our modern society, data complexity is rising, and the extraction of valuable information has become crucial with the rapid changes in information technology (IT). Recurrent neural networks (RNN) and long short-term memory (LSTM) models have shown remarkable performance in natural language processing (NLP) and time series prediction. Consequently, there is a strong expectation that models excelling in NLP will also excel in time series prediction. However, current research on Transformer models for time series prediction remains limited. Traditional RNN and LSTM models have demonstrated superior performance compared to Transformers in big data analysis. Nevertheless, with continuous advancements in Transformer models, such as GPT-2 (Generative Pre-trained Transformer 2) and ProphetNet, they have gained attention in the field of time series prediction. This study aims to evaluate the classification performance and interval prediction of remaining useful life (RUL) using an advanced Transformer model. The performance of each model will be utilized to establish a health index (HI) for cutting blades, enabling real-time monitoring of machine health. The results are expected to provide valuable insights for machine monitoring, evaluation, and management, confirming the effectiveness of advanced Transformer models in time series analysis when applied in industrial settings.

• 대한임베디드공학회논문지
• /
• 제15권2호
• /
• pp.43-49
• /
• 2020

In this paper, we propose a stage-wise knowledge transfer method that uses block-wise retraining to transfer the useful knowledge of a pre-trained residual network (ResNet) in a teacher-student framework (TSF). First, multiple hint information transfer and block-wise supervised retraining of the information was alternatively performed between teacher and student ResNet models. Next, Softened output information-based knowledge transfer was additionally considered in the TSF. The results experimentally showed that the proposed method using multiple hint-based bottom-up knowledge transfer coupled with incremental block-wise retraining provided the improved student ResNet with higher accuracy than existing KD and hint-based knowledge transfer methods considered in this study.

• 산업기술연구
• /
• 제29권A호
• /
• pp.47-54
• /
• 2009

In this paper, we proposed hand detection and hand gesture recognition from USB camera video. Firstly, we extract hand region extraction using skin color information from a difference images. Background image is initially stored and extracted from the input images in order to reduce problems from complex backgrounds. After that, 16-directional chain code sequence is computed from the tracking of hand motion. These chain code sequences are compared with pre-trained models using DP matching. Our hand gesture recognition system can be used to control PowerPoint slides or applied to multimedia education systems. We got 92% hand region extraction accuracy and 82.5% gesture recognition accuracy, respectively.

• ETRI Journal
• /
• 제43권4호
• /
• pp.684-693
• /
• 2021

This paper presents a low-cost two-stage approximate multiplier for bfloat16 (brain floating-point) data processing. For cost-efficient approximate multiplication, the first stage implements Mitchell's algorithm that performs the approximate multiplication using only two adders. The second stage adopts the exact multiplication to compensate for the error from the first stage by multiplying error terms and adding its truncated result to the final output. In our design, the low-cost multiplications in both stages can reduce hardware costs significantly and provide low relative errors by compensating for the error from the first stage. We apply our approximate multiplier to the convolutional neural network (CNN) inferences, which shows small accuracy drops with well-known pre-trained models for the ImageNet database. Therefore, our design allows low-cost CNN inference systems with high test accuracy.

• International Journal of Computer Science & Network Security
• /
• 제21권7호
• /
• pp.324-330
• /
• 2021

Agriculture is one of the essential needs of human life on planet Earth. It is the source of food and earnings for many individuals around the world. The economy of many countries is associated with the agriculture sector. Lots of diseases exist that attack various fruits and crops. Apple Tree Leaves also suffer different types of pathological conditions that affect their production. These pathological conditions include apple scab, cedar apple rust, or multiple diseases, etc. In this paper, an automatic detection framework based on deep learning is investigated for apple leaves disease classification. Different pre-trained models, VGG16, ResNetV2, InceptionV3, and MobileNetV2, are considered for transfer learning. A combination of parameters like learning rate, batch size, and optimizer is analyzed, and the best combination of ResNetV2 with Adam optimizer provided the best classification accuracy of 94%.

• 한국컴퓨터정보학회논문지
• /
• 제23권11호
• /
• pp.31-41
• /
• 2018

Text classification has been studied for a long time in the Natural Language Processing field. In this paper, we propose an article- and paragraph-level genre classification system using Word2Vec-based LSTM, GRU, and CNN models for large-scale English corpora. Both article- and paragraph-level classification performed best in accuracy with LSTM, which was followed by GRU and CNN in accuracy performance. Thus, it is to be confirmed that in evaluating the classification performance of LSTM, GRU, and CNN, the word sequential information for articles is better than the word feature extraction for paragraphs when the pre-trained Word2Vec-based word embeddings are used in both deep learning-based article- and paragraph-level classification tasks.

• Journal of Information Processing Systems
• /
• 제15권2호
• /
• pp.399-409
• /
• 2019

Three-dimensional (3D) human pose reconstruction from single-view image is a difficult and challenging topic. Existing approaches mostly process frame-by-frame independently while inter-frames are highly correlated in a sequence. In contrast, we introduce a novel spatial-temporal 3D human pose reconstruction framework that leverages both intra and inter-frame relationships in consecutive 2D pose sequences. Orthogonal matching pursuit (OMP) algorithm, pre-trained pose-angle limits and temporal models have been implemented. Several quantitative comparisons between our proposed framework and recent works have been studied on CMU motion capture dataset and Vietnamese traditional dance sequences. Our framework outperforms others by 10% lower of Euclidean reconstruction error and more robust against Gaussian noise. Additionally, it is also important to mention that our reconstructed 3D pose sequences are more natural and smoother than others.