• 산업경영시스템학회지
• /
• 제46권1호
• /
• pp.55-67
• /
• 2023

Lightweight face recognition models, as one of the most popular and long-standing topics in the field of computer vision, has achieved vigorous development and has been widely used in many real-world applications due to fewer number of parameters, lower floating-point operations, and smaller model size. However, few surveys reviewed lightweight models and reimplemented these lightweight models by using the same calculating resource and training dataset. In this survey article, we present a comprehensive review about the recent research advances on the end-to-end efficient lightweight face recognition models and reimplement several of the most popular models. To start with, we introduce the overview of face recognition with lightweight models. Then, based on the construction of models, we categorize the lightweight models into: (1) artificially designing lightweight FR models, (2) pruned models to face recognition, (3) efficient automatic neural network architecture design based on neural architecture searching, (4) Knowledge distillation and (5) low-rank decomposition. As an example, we also introduce the SqueezeFaceNet and EfficientFaceNet by pruning SqueezeNet and EfficientNet. Additionally, we reimplement and present a detailed performance comparison of different lightweight models on the nine different test benchmarks. At last, the challenges and future works are provided. There are three main contributions in our survey: firstly, the categorized lightweight models can be conveniently identified so that we can explore new lightweight models for face recognition; secondly, the comprehensive performance comparisons are carried out so that ones can choose models when a state-of-the-art end-to-end face recognition system is deployed on mobile devices; thirdly, the challenges and future trends are stated to inspire our future works.

• 실천공학교육논문지
• /
• 제14권3호
• /
• pp.491-498
• /
• 2022

본 논문에서는 비전공자들을 위한 교양과정으로, 기초 순환신경망 과목 커리큘럼을 설계하는데 필수적으로 요구되는 순환신경망 SW 실습 사례를 개발하였다. 개발된 SW 실습 사례는 순환신경망의 동작원리를 이해시키는 데 초점을 두고, 시각화된 전체 동작 과정을 확인할 수 있도록 스프레드시트를 사용하였다. 개발된 순환신경망 실습 사례는 지도학습 방식의 텍스트완성 훈련데이터 생성, 입력층, 은닉층, 상태층(컨텍스트 노드) 그리고 출력층을 차례대로 구현하고, 텍스트 데이터에 대해 순환신경망의 성능을 테스트하는 것으로 구성되었다. 본 논문에서 개발한 순환신경망 실습사례는 다양한 문자 수를 갖는 단어를 자동 완성한다. 제안한 순환신경망 실습사례를 활용하여, 한글 또는 영어 단어를 구성하는 최대 문자 수를 다양하게 확장하여 자동 완성하는 인공지능 SW 실습 사례를 만들 수 있다. 따라서, 본 순환신경망 기초 실습 사례의 활용도가 높다고 할 수 있다.

• 인터넷정보학회논문지
• /
• 제24권6호
• /
• pp.23-29
• /
• 2023

가상현실 분야에서는 사용자 경험의 몰입도를 극대화하기 위해 햅틱 피드백을 활용하고 발전시키려는 지속적인 노력이 있었다. 그러나 대부분의 햅틱 피드백은 진동 모터 등 경제성을 고려한 액추에이터를 사용하는 문제로 인해 제한적인 촉각 경험만을 사용자에게 제공할 수 있었다. 복싱과 같은 스포츠 시뮬레이션이나 게임에서의 타격 경험의 경우, 실제 물체를 타격하는 감각과 진동 액추에이터로 렌더링되는 감각 사이의 괴리 때문에 한계가 분명하다. 본 연구에서는 이를 주목하여, 사용자가 손으로 가상의 물체를 타격할 때 가상 임팩트를 생성할 수 있는 햅틱 임팩트 시스템을 제안했다. 햅틱 인터페이스는 퀵 리턴 메커니즘을 사용하여 엔드이펙터가 사용자의 주먹에 햅틱 임팩트 피드백을 직접 전달하고 진동 촉감을 통해서 사용자의 손바닥에 가상 임팩트 감각을 전달할 수 있도록 하였다. 제안된 시스템은 인간 대상 실험을 통해 평가하였으며 실험 결과는 햅틱 임팩트 가상 임팩트의 인지 강도와 사실감에 유의한 영향을 미친다는 것을 나타낸다.

• 한국인터넷방송통신학회논문지
• /
• 제23권6호
• /
• pp.1-7
• /
• 2023

본 논문은 인공지능 기술을 활용하여 일반 도로와 침수 도로를 분류하는 학습모델을 개발하였다. 다양한 데이터 증강기법을 사용하여 학습 데이터의 다양성을 확장하며, 여러 환경에서도 좋은 성능을 보이는 모델을 구현하였다. CNN 기반의 Resnet152v2 모델을 사전 학습모델로 활용하여, 전이 학습을 진행하였다. 모델의 학습 과정에서 다양한 파라미터 튜닝 및 최적화 과정을 거쳐 최종 모델의 성능을 향상하였다. 학습은 파이선으로 Google Colab NVIDIA Tesla T4 GPU를 사용하여 구현하였고, 테스트 결과 시험 데이터 세트에서 매우 높은 정확도로 침수상황을 탐지함을 알 수 있었다.

• 한국인터넷방송통신학회논문지
• /
• 제23권5호
• /
• pp.83-88
• /
• 2023

본 논문에서는 홈 트레이닝 인구가 늘어나면서 증가한 운동과 식단 보조 서비스에 대한 수요를 충족시키기 위해 운동 및 식단 보조 애플리케이션을 개발하였다. 애플리케이션은 카메라를 통해 실시간으로 촬영되는 사용자의 운동 자세를 분석하여, 안내선과 음성을 이용해 올바른 자세를 유도하는 기능을 가진다. 또한, 사용자가 사진을 촬영하면 사진에 포함된 음식들을 분류하고 각 음식의 양을 추정하여, 칼로리 등의 영양 정보를 계산하여 제공하는 기능을 가진다. 영양 정보 계산은 외부의 서버에서 수행되도록 구성했다. 서버는 계산된 결과를 애플리케이션으로 전송하고, 애플리케이션은 결과를 받아 시각적으로 출력한다. 추가로, 운동 결과와 영양 정보는 날짜별로 저장하고 확인할 수 있도록 하였다.

• 산업경영시스템학회지
• /
• 제46권3호
• /
• pp.1-6
• /
• 2023

In the realm of dental prosthesis fabrication, obtaining accurate impressions has historically been a challenging and inefficient process, often hindered by hygiene concerns and patient discomfort. Addressing these limitations, Company D recently introduced a cutting-edge solution by harnessing the potential of intraoral scan images to create 3D dental models. However, the complexity of these scan images, encompassing not only teeth and gums but also the palate, tongue, and other structures, posed a new set of challenges. In response, we propose a sophisticated real-time image segmentation algorithm that selectively extracts pertinent data, specifically focusing on teeth and gums, from oral scan images obtained through Company D's oral scanner for 3D model generation. A key challenge we tackled was the detection of the intricate molar regions, common in dental imaging, which we effectively addressed through intelligent data augmentation for enhanced training. By placing significant emphasis on both accuracy and speed, critical factors for real-time intraoral scanning, our proposed algorithm demonstrated exceptional performance, boasting an impressive accuracy rate of 0.91 and an unrivaled FPS of 92.4. Compared to existing algorithms, our solution exhibited superior outcomes when integrated into Company D's oral scanner. This algorithm is scheduled for deployment and commercialization within Company D's intraoral scanner.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• 제17권12호
• /
• pp.3383-3397
• /
• 2023

Scene graphs serve as semantic abstractions of images and play a crucial role in enhancing visual comprehension and reasoning. However, the performance of Scene Graph Generation is often compromised when working with biased data in real-world situations. While many existing systems focus on a single stage of learning for both feature extraction and classification, some employ Class-Balancing strategies, such as Re-weighting, Data Resampling, and Transfer Learning from head to tail. In this paper, we propose a novel approach that decouples the feature extraction and classification phases of the scene graph generation process. For feature extraction, we leverage a transformer-based architecture and design an adaptive calibration function specifically for predicate classification. This function enables us to dynamically adjust the classification scores for each predicate category. Additionally, we introduce a Distribution Alignment technique that effectively balances the class distribution after the feature extraction phase reaches a stable state, thereby facilitating the retraining of the classification head. Importantly, our Distribution Alignment strategy is model-independent and does not require additional supervision, making it applicable to a wide range of SGG models. Using the scene graph diagnostic toolkit on Visual Genome and several popular models, we achieved significant improvements over the previous state-of-the-art methods with our model. Compared to the TDE model, our model improved mR@100 by 70.5% for PredCls, by 84.0% for SGCls, and by 97.6% for SGDet tasks.

• Korean Journal of Radiology
• /
• 제22권6호
• /
• pp.912-921
• /
• 2021

Objective: To compare the performance of the deep learning-based lesion detection algorithm (DLLD) in detecting liver metastasis with that of radiologists. Materials and Methods: This clinical retrospective study used 4386-slice computed tomography (CT) images and labels from a training cohort (502 patients with colorectal cancer [CRC] from November 2005 to December 2010) to train the DLLD for detecting liver metastasis, and used CT images of a validation cohort (40 patients with 99 liver metastatic lesions and 45 patients without liver metastasis from January 2011 to December 2011) for comparing the performance of the DLLD with that of readers (three abdominal radiologists and three radiology residents). For per-lesion binary classification, the sensitivity and false positives per patient were measured. Results: A total of 85 patients with CRC were included in the validation cohort. In the comparison based on per-lesion binary classification, the sensitivity of DLLD (81.82%, [81/99]) was comparable to that of abdominal radiologists (80.81%, p = 0.80) and radiology residents (79.46%, p = 0.57). However, the false positives per patient with DLLD (1.330) was higher than that of abdominal radiologists (0.357, p < 0.001) and radiology residents (0.667, p < 0.001). Conclusion: DLLD showed a sensitivity comparable to that of radiologists when detecting liver metastasis in patients initially diagnosed with CRC. However, the false positives of DLLD were higher than those of radiologists. Therefore, DLLD could serve as an assistant tool for detecting liver metastasis instead of a standalone diagnostic tool.

• 대한통합의학회지
• /
• 제12권1호
• /
• pp.183-190
• /
• 2024

Purpose : This study aims to conduct a comprehensive comparison and analysis of intervention strategies utilized for school-aged children facing difficulties in writing, focusing on evaluating the effectiveness of various intervention approaches both domestically and internationally. The primary focus is on assessing the efficacy of each intervention approach and identifying gaps in the existing literature. Methods : Data for this study were gathered from the domestic database RISS from January 2013 to March 2020, and international databases Pubmed and Google Scholar were utilized. The keywords for domestic literature search included 'occupational therapy', 'handwriting', and 'school-aged', while for international literature search, the keywords were 'occupational therapy', 'handwriting', and 'children'. A total of 4 international and 2 domestic articles were selected for review based on predetermined inclusion and exclusion criteria. Results : The study findings present a thorough comparative analysis of intervention strategies, categorizing them into task-oriented intervention, sensory-motor intervention, and integrated intervention. All intervention methods demonstrated notable improvements in the legibility of handwriting. Comparison between domestic and international literature revealed a predominant use of task-oriented intervention in domestic studies, while international studies showcased a diverse range of intervention methods. Conclusion : Interventions were categorized into computer-based, task-oriented, sensory-motor, and integrated interventions. Task-oriented interventions were the most common in both domestic and international studies, while integrated interventions were the most effective. Based on these findings, it is necessary to increase awareness of the need for handwriting intervention research among occupational therapists in Korea. Additionally, there is a need for well-supported handwriting intervention research with larger sample sizes in both domestic and international occupational therapy. Finally, future research should actively investigate the application of tailored integrated interventions for school-aged children with handwriting difficulties.

• 대한치과기공학회지
• /
• 제46권2호
• /
• pp.36-41
• /
• 2024

Purpose: This study aimed to generate virtual mandibular left first molar teeth using deep convolutional generative adversarial networks (DCGANs) and analyze their matching accuracy with actual tooth morphology to propose a new paradigm for using medical data. Methods: Occlusal surface images of the mandibular left first molar scanned using a dental model scanner were analyzed using DCGANs. Overall, 100 training sets comprising 50 original and 50 background-removed images were created, thus generating 1,000 virtual teeth. These virtual teeth were classified based on the number of cusps and occlusal surface ratio, and subsequently, were analyzed for consistency by expert dental technicians over three rounds of examination. Statistical analysis was conducted using IBM SPSS Statistics ver. 23.0 (IBM), including intraclass correlation coefficient for intrarater reliability, one-way ANOVA, and Tukey's post-hoc analysis. Results: Virtual mandibular left first molars exhibited high consistency in the occlusal surface ratio but varied in other criteria. Moreover, consistency was the highest in the occlusal buccal lingual criteria at 91.9%, whereas discrepancies were observed most in the occusal buccal cusp criteria at 85.5%. Significant differences were observed among all groups (p<0.05). Conclusion: Based on the classification of the virtually generated left mandibular first molar according to several criteria, DCGANs can generate virtual data highly similar to real data. Thus, subsequent research in the dental field, including the development of improved neural network structures, is necessary.