Earticle

본 연구는 노후화된 토목 인프라의 안전점검 정밀도 향상과 효율성을 위해 컴퓨터 비전 기반 구조물 변형 측정 플랫폼(Computer Vision-Based Deformation Measurement Platform)을 개발하였다. 본 플랫폼은 영상 기반 거리·각도 측정 알고리즘을 구현하여 구조물의 변형 상태를 정량화하고, 자동 데이터 저장 및 시각화를 통해 육안 중심 점검의 한계를 보완한다. 실험 결과, 거리 및 각도 측정의 평균 오차율은 1% 미만으로 나타나 정확도와 실무 활용성이 모두 입증되었다. 또한 독립 실행형 형태로 구현되어, 현장 점검의 즉시성과 접근성을 확보하였다. 향후 본 기술은 인공지능 기반 자동 진단과 클라우드 연동형 스마트 안전관리 체계로 확장될 수 있다. 더불어 장기간 축적된 데이터를 기반으로 구조물 열화 패턴을 분석하는 예측 모델 개발에도 활용 가능하다. 이러한 확장성은 노후 인프라 관리의 비용 효율성과 위험도 기반 의사결정 체계 구축에 실질적인 기여를 제공할 것으로 기대된다.

This study developed a Computer Vision-Based Deformation Measurement Platform to enhance the accuracy and efficiency of safety inspections for aging civil infrastructure. The platform implements image-based distance and angle measurement algorithms to quantify structural deformation, while automated data storage and visualization functions address the limitations of traditional visual inspections. Experimental results demonstrated an average error rate of less than 1% for both distance and angle measurements, confirming high accuracy and practical applicability. The system was also implemented as a standalone executable, ensuring improved immediacy and accessibility in field operations. The proposed technology can be extended to AI-driven automated diagnostics and cloud-connected smart safety management frameworks. In addition, long-term accumulated data can be utilized to develop predictive models for analyzing structural deterioration patterns. Such scalability is expected to contribute substantially to cost-efficient infrastructure management and the establishment of risk-based decision-making systems.

The Nepalese Currency Counterfeit Detection System aims to address the growing threat of counterfeit currency in Nepal. Counterfeit banknotes pose a significant risk to the nation's financial stability and erode public trust in the monetary system. This system seeks to differentiate genuine banknotes from forgeries, thereby ensuring the integrity of financial transactions. Counterfeiting activities have a detrimental impact on the livelihoods of individuals and the overall economy. To combat this issue, this project will thoroughly investigate the various security features embedded within Nepali currency. Subsequently, it will leverage advanced image processing and computer vision techniques to develop a software-based solution capable of detecting and validating the authenticity of Nepali banknotes. While specialized machines are available in banks and commercial establishments for currency authentication, these systems are often inaccessible to the general public. This project aims to bridge this gap by developing a user-friendly software application that empowers individuals to independently verify the authenticity of Nepali currency.

The rapid advancement of imaging technology has led to a surge in the volume of image data, making it challenging for image owners to efficiently store and process them. To address these issues, many are turning to cloud service providers (CSP) for their powerful storage and computational resources. Despite this convenience, reliance on cloud servers to enable computationally demanding computer vision applications such as content-based image retrieval (CBIR), poses significant privacy risks. As images may contain personally identifiable information and they may be subjected to copyright. In this regard, a straightforward solution is to encrypt images on the users’ end before sharing them with the third-party owned servers. However, the main challenge is to find a better trade-off between privacy and data usability in a cost-effective manner. Therefore, this paper presents a privacy preserving CBIR scheme that leverages the recent advancements of incorporating sub-block processing in perceptual encryption (PE) for enhanced security. In addition, our image retrieval scheme is histogram based that combines color and edge information with discrete wavelet transform; therefore, it is invariant to the encryption transformation functions. The simulation results show that our privacy preserving CBIR achieves the same retrieval performance as that of the plain images while delivering better security than the conventional privacy preserving CBIR techniques.

현대사회는 모든 정보가 디지털화되고 있으므로 개인 정보의 보호가 무척이나 중요한 이슈다. 특히 미디어 산업의 폭발적 성장으로 인해 비디오에 촬영된 인물의 익명화는 매우 중대한 문제가 되었다. 전통적인 방법은 블러링이나 픽셀화를 사용하고, 최신 기술들 은 생성적 적대 신경망(GAN)을 활용하여 비디오에 촬영된 얼굴을 다시 그리는 방법으로 익명화를 달성한다. 우리는 훨씬 작은 모델 을 활용하여 실시간 연산을 통해 비디오에 촬영된 인물의 신체 전부를 익명화하는 방법을 제안한다. 기존 방법은 인물의 피부색, 의복, 소지품, 체형 등 얼굴 이외의 개인 식별 정보를 제거하는 것이 어려웠으나, 우리의 방법은 영상 내에서 이러한 정보를 모두 지울 수 있다. 또한 자세 인식 알고리즘을 활용하여 인물의 위치, 움직임, 자세 등의 정보는 표현할 수 있다. 이 알고리즘은 다양한 산업 현장에 설치된 CCTV나 IP 카메라에 적용되어 실시간으로 동작할 수 있으므로, 전신 익명화 기술의 보급에 기여할 수 있을 것이다.

In the contemporary digital era, protection of personal information has become a paramount issue. The exponential growth of the media industry has heightened concerns regarding the anonymization of individuals captured in video footage. Traditional methods, such as blurring or pixelation, are commonly employed, while recent advancements have introduced generative adversarial networks (GANs) to redraw faces in videos. In this study, we propose a novel approach that employs a significantly smaller model to achieve real-time full-body anonymization of individuals in videos. Unlike conventional techniques that often fail to effectively remove personal identification information—such as skin color, clothing, accessories, and body shape—our method successfully eradicates all such details. Furthermore, by leveraging pose estimation algorithms, our approach accurately represents information regarding individuals' positions, movements, and postures. This algorithm can be seamlessly integrated into CCTV or IP camera systems installed in various industrial settings, functioning in real-time and thus facilitating the widespread adoption of full-body anonymization technology.

The 2D and 3D pose estimation methods have now improved well in general performance but have not yet been emphasized in terms of speed and efficiency for the infant dataset and the existence of public data on infants is a significant challenge. Furthermore, clinical studies related to the analysis of the pose and movements of infants are attracting considerable attention. That motivated us to collect infant data and develop a lighter model for estimating infant poses that can run on edge devices and CPUs. Most current methods are characterized by complex structures and multiple parallel branches of inference to synthesize pose estimated results. In this project, we aim to refine the architecture of the pose estimation algorithm based on an approach of OpenPose-2016, for use on edge devices and training that model on 2D images. The proposed simplified model features a single-branch structure designed to estimate infant pose with a size of 4.09 million parameters. The model when executed undergo algorithmic complexity of 8.97 giga floating point operations per second (GFLOPS), allowing it to run at approximately 23 frames per second on a Core i5-10400f. The proposed methodology demonstrates compact dimensions while achieving superior performance compared to existing methods on the same self-collected infant dataset. It is hoped that this straightforward and pragmatic approach will establish a robust foundation and provide favorable conditions for future research in the application of pose estimation.

비디오에서 행동의 위치를 추정하기 위해 행동의 시작과 끝에 모두 주석을 다는 것은 많은 비용을 요구한다. 따라서 주석 생성 비용을 줄이기 위해 비디오 단위의 행동 카테고리만 존재하는 약지도 학습이 활발히 진행되고 있지만 약 지도 학습은 행동단위 주석의 부재에 따른 많은 한계가 존재한다. 최근에는 주석의 생성 비용은 줄이면서 약지도 학 습에 비해 훨씬 개선된 성능을 낼 수 있는 포인트 주석 기반 약지도 학습에 대한 연구가 진행되고 있다. 하지만 이는 행동구간 내의 포인트가 존재하는 위치에 많이 의존한다는 한계가 있다. 본 연구진은 이런 점을 보완하고자, 행동구 간을 나타낼 수 있는 확률분포를 설계하고 포인트 주석을 샘플링하여 학습이 진행됨에 따라 점차 포인트 주석을 행 동구간의 가운데 방향으로 움직일 수 있게끔 하는 포인트 업데이트 방식을 제안한다. 실험은 제안하는 방법이 기존 의 포인트 주석 기반 학습에서 치우친 포인트가 주어진 경우 성능의 한계를 개선함을 보여준다.

Labeling the starting and ending points of interesting actions in a video is labor-intensive and expensive. To reduce the labeling costs, many studies have been conducted on weakly supervised learning, where only video-level action labels exist, but it has limitations due to the absence of frame-level annotations. Recently, research has been conducted on point annotation-based learning, which can achieve much improved performance while reducing labeling costs. However, this has a limitation in that the performance depends the location of annotated points in the action interval. To compensate for this, we design a probability distribution that can represent action intervals, which allows us to sample point annotations and update them to gradually move point annotations toward the center of action intervals as learning progresses. Experiments show that the proposed method improves the performance limitation given biased points in a point annotation-based learning scheme.

본 연구에서 객체 인식 과제의 대상으로 선정한 빗속의 사람 그림은 내담자의 스트레스와 스트레스 대처 행동을 반 영하는 검사이다. Lack의 빗속의 사람 그림 채점 기준에 따르면 비, 웅덩이 등 16개 항목은 스트레스 요인으로 분 류되고 우산, 우비 등 19개 항목은 대처 자원으로 분류된다. 본 연구에서는 총 700장의 빗속의 사람 그림 이미지 중 560장을 트레이닝 셋으로, 30장을 밸리데이션 셋으로, 110장을 테스트 셋으로 구성한 데이터셋으로 객체 인식 모델을 시도하였다. 모델은 one-stage detector인 Dynamic Head를 채택하여 대규모 컴퓨터 비전 벤치마크 데 이터셋 (MS COCO 2017)에서 24 에폭으로 사전 훈련을 진행하였고, 빗속의 사람 그림 데이터셋을 36 에폭으로 fine-tune하였다. 모델의 mean average precision은 밸리데이션 셋에서 62.7 이었고 테스트 셋에서는 49.4 이 었다. 본 연구에서 제안한 모델은 다양한 유형의 빗속의 사람 그림 평가에 참고가 되는 비, 사람, 우산 등의 주요 객체들을 인식할 수 있었다.

Person in the rain (PITR) is an art therapy assessment method that reflects both stress and coping behaviors. According to Lack's PITR scoring criteria, 16 items such as rain and puddles are classified as stressors, and 19 items such as umbrellas and raincoats are classified as coping resources. This study used PITR for object detection with a total of 700 images of the dataset consisting of 560 images for training, 30 for validation, and 110 for testing. The object detection model suggested here adopts a one-stage object detection architecture called Dynamic Head (i.e., Dyhead). After pre-training the model on a large computer vision benchmark dataset (MS COCO 2017) for 24 epochs, it was fine-tuned for another 36 epochs using the PITR dataset, with a mean average precision of 62.7 on the validation set and 49.4 on the test set. The proposed model can detect key objects such as rain, person, and umbrella for the PITR test from various image styles.

Often most of the modern human people are suffering from a long time of working or studying on the stationary pose. Subsequently, the health of our life is highly threatened to be exacerbated by chronic orthopedic diseases. In order to solve this social problem, we suggest pose detection that can have the people who have deleterious postures be notified. By using nowadays advanced computer vision techniques, in this paper we suggest the posture recognition module to enhance our quality of life. While most posture recognition recognizes only one person's posture, we made our pipeline to perform posture recognition for multiple people through images obtained through a single camera. One of the big problems in measuring people's postures is that it is necessary to distinguish the various body structures and postures of people. For this, posture images and labeling of various people are required. We created pose images of people of various body types through images of a small number of people through skeleton-based coordinates augmentation. We made a posture classifier using various models and observed the improvement of augmentation performance for each model. Through this, we found that the postures of various people can be measured using a relatively small data set. In particular, for deep-learning models that require a lot of data, generalization performance was greatly improved.

Lossy image compression provides an efficient solution to the exchange and storage of image data for consumer applications. The design of lossy algorithms is based on a principle to discard information that are not perceivable by human visual system (HVS). With the popularity of deep learning models (DL) in computer vision (CV), it is necessary to characterize the loss in image quality with respect to computer vision systems as well. Recent studies have analyzed the image distortions resulted from blur and noise, mainly from an adversarial attack perspective. However, fewer studies have dealt with the lossy nature of the JPEG algorithm. Therefore, the current study presents a quantitative assessment of different types of data loss that occurs due to chroma subsampling, quantization, and rounding functions of the JPEG algorithm. In addition, we have analyzed impact of different interpolation methods that are used for chroma upsampling. The analysis have shown that for compression savings, performing either subsampling or quantization preserved the model accuracy while their combination degraded the accuracy by 6%.

본 논문에서는 안무저작권의 기본적인 요소인 무용 동작 간의 실질적인 유사성을 검증하기 위하여 인공지능 딥러닝 기술을 적용한 무용 동작의 정량화 방법론을 제안한다. 인간 자세 추정을 위한 딥러닝 네트워크 오픈포 즈 라이브러리를 이용해 2D 댄스 영상에서 추출한 이미지 데이터를 분석하여 18개의 신체 키포인트를 검출했 다. 검출된 키포인트는 스켈레톤 이미지를 출력하기 위해 연결되었고, 데이터 분석을 위한 2차원 공간 좌표 형 태로 변환되었다. 수학적 모델을 이용한 데이터 간의 유사성 검증을 통해 무용 동작 유사성 측정의 가능성을 확인하였다. 본 연구에서 제시된 방법론은 안무저작권의 기초 단위인 무용 동작의 정량적 데이터를 제공함으로 써 인공지능 딥러닝 기술을 활용한 무용 동작 분석 기초 연구의 새로운 시각을 제안하였다.

In this paper, I propose a quantification methodology of dance movements applied with artificial intelligence deep learning technology to verify the practical similarity between dance movements, which are basic elements of dance copyright. I analyzed image data extracted from 2D dance videos using the deep learning network OpenPose library for human pose estimation and detected 18 body key points. The detected key point was connected to output a skeleton image and converted into a two-dimensional spatial coordinate form for data analysis. The possibility of measuring the similarity of dance motion was confirmed through the similarity verification between data using a mathematical model. The methodology presented in this study proposed a new perspective of the basic research on dance motion analysis using artificial intelligence deep learning technology by providing quantitative data on dance motion, which is the basic unit of choreography copyright.

The ionosphere is the part of the Earth's atmosphere with a high concentration of free electrons and ions. The ionosphere is characterised by its variability and inhomogeneity. One of the characteristic inhomogeneities is the so-called auroral oval, which determines the range of auroral radiance. Detection of the auroral oval is an important task for forecasting auroral storms, as they affect long-range communication systems, navigation, satellite-to-ground communications, making communications complicated or impossible. Therefore, an auroral oval detection and prediction needs to be performed in order to be informed about the area of their possible influence at certain time intervals. On the basis of the available image dataset from SIMuRG, which is based on GNSS data, it is proposed to use the LSTM model and CNN architecture. The paper reviews existing implementations and proposes a method for predicting auroral oval movements in the images, using the Convolutional LSTM architecture, which combines time series processing and computer vision. The work results in a machine learning model that can make the predictions based on even small sets of data.

수많은 공간지각 연구 결과, Euclidean 3-D 구조는 양안 입체시, 움직임, 입체시와 움직임의 결합, 또는 여러 광학 정보의 결합으로도 복구될 수 없다는 사실이 밝혀졌다. 그러나 인간은 이러한 부정확한 공간지각에도 불구하고 특정 과제를 수행하는 데는 어려움이 전혀 없다. 우리는 인공지능과 컴퓨터 비전에 인간의 기술과 능력을 적용해 왔지 만 이러한 기계들은 여전히 인간의 능력보다 훨씬 뒤떨어져 있다. 따라서 우리는 인간이 공간의 깊이를 어떻게 지각하 는지, 과제를 수행하기 위해 어떠한 정보들을 사용하여 3차원 공간을 정확하게 지각하는지 이해해야 한다. 이 논문의 목적은 미래에 더욱 발전된 인공지능 로봇에 인간의 능력을 적용하기 위해 공간지각 문헌을 검토하는 것이다.

Numerous space perception studies have shown that Euclidean 3-D structure cannot be recovered from binocular stereopsis, motion, combination of stereopsis and motion, or even with combined multiple sources of optical information. Humans, however, have no difficulties to perform the task-specific action despite of poor shape perception. We have applied humans skill and capabilities to artificial intelligence and computer vision but those machines are still far behind from humans abilities. Thus, we need to understand how we perceive depth in space and what information we use to perceive 3-D structure accurately to perform. The purpose of this paper was to review space perception literatures to apply humans abilities to artificial intelligence robots more advanced in future.

The objectives of this study were to evaluate convolutional neural network models and computer vision techniques for the classification of swine posture with high accuracy and to use the derived result in the investigation of the effect of dietary fiber level on the behavioral characteristics of the pregnant sow under low and high ambient temperatures during the last stage of gestation. A total of 27 crossbred sows (Yorkshire × Landrace; average body weight, 192.2 ± 4.8 kg) were assigned to three treatments in a randomized complete block design during the last stage of gestation (days 90 to 114). The sows in group 1 were fed a 3% fiber diet under neutral ambient temperature; the sows in group 2 were fed a diet with 3% fiber under high ambient temperature (HT); the sows in group 3 were fed a 6% fiber diet under HT. Eight popular deep learning-based feature extraction frameworks (DenseNet121, DenseNet201, InceptionResNetV2, InceptionV3, MobileNet, VGG16, VGG19, and Xception) used for automatic swine posture classification were selected and compared using the swine posture image dataset that was constructed under real swine farm conditions. The neural network models showed excellent performance on previously unseen data (ability to generalize). The DenseNet121 feature extractor achieved the best performance with 99.83% accuracy, and both DenseNet201 and MobileNet showed an accuracy of 99.77% for the classification of the image dataset. The behavior of sows classified by the DenseNet121 feature extractor showed that the HT in our study reduced (p < 0.05) the standing behavior of sows and also has a tendency to increase (p = 0.082) lying behavior. High dietary fiber treatment tended to increase (p = 0.064) lying and decrease (p < 0.05) the standing behavior of sows, but there was no change in sitting under HT conditions.

Pig farming, a vital industry, necessitates proactive measures for early disease detection and crush symptom monitoring to ensure optimum pig health and safety. This review explores advanced thermal sensing technologies and computer vision-based thermal imaging techniques employed for pig disease and piglet crush symptom monitoring on pig farms. Infrared thermography (IRT) is a non-invasive and efficient technology for measuring pig body temperature, providing advantages such as non-destructive, long-distance, and high-sensitivity measurements. Unlike traditional methods, IRT offers a quick and labor-saving approach to acquiring physiological data impacted by environmental temperature, crucial for understanding pig body physiology and metabolism. IRT aids in early disease detection, respiratory health monitoring, and evaluating vaccination effectiveness. Challenges include body surface emissivity variations affecting measurement accuracy. Thermal imaging and deep learning algorithms are used for pig behavior recognition, with the dorsal plane effective for stress detection. Remote health monitoring through thermal imaging, deep learning, and wearable devices facilitates non-invasive assessment of pig health, minimizing medication use. Integration of advanced sensors, thermal imaging, and deep learning shows potential for disease detection and improvement in pig farming, but challenges and ethical considerations must be addressed for successful implementation. This review summarizes the state-of-the-art technologies used in the pig farming industry, including computer vision algorithms such as object detection, image segmentation, and deep learning techniques. It also discusses the benefits and limitations of IRT technology, providing an overview of the current research field. This study provides valuable insights for researchers and farmers regarding IRT application in pig production, highlighting notable approaches and the latest research findings in this field.

Rumination in cattle is closely related to their health, which makes the automatic monitoring of rumination an important part of smart pasture operations. However, manual monitoring of cattle rumination is laborious and wearable sensors are often harmful to animals. Thus, we propose a computer vision-based method to automatically identify multi-object cattle rumination, and to calculate the rumination time and number of chews for each cow. The heads of the cattle in the video were initially tracked with a multi-object tracking algorithm, which combined the You Only Look Once (YOLO) algorithm with the kernelized correlation filter (KCF). Images of the head of each cow were saved at a fixed size, and numbered. Then, a rumination recognition algorithm was constructed with parameters obtained using the frame difference method, and rumination time and number of chews were calculated. The rumination recognition algorithm was used to analyze the head image of each cow to automatically detect multi-object cattle rumination. To verify the feasibility of this method, the algorithm was tested on multi-object cattle rumination videos, and the results were compared with the results produced by human observation. The experimental results showed that the average error in rumination time was 5.902% and the average error in the number of chews was 8.126%. The rumination identification and calculation of rumination information only need to be performed by computers automatically with no manual intervention. It could provide a new contactless rumination identification method for multi-cattle, which provided technical support for smart pasture.

There have been exciting advancements in the fields of computer vision with the development of data analysis techniques and computational efficiency. In this project, we have built a user-friendly image processing tools using the tkinter libraries in Python by defining functions for each computer vision method and implementing pre-processing, filtering, feature extraction, and clustering on images collected from various fields such as environmental sciences, medical sciences, and mobilities. We have sought to assess the utility of automated image processing software to improve image classifiers even by nonprofessionals with no coding and no machine learning expertise. Thus, this program can be easily utilized by researchers without programming experience. This program can be used for various purposes by scholars in various fields such as education, environmental sciences and medicine.

Due to high ergonomic risks, forestry considered a high-priority issue in forest workers' health. Timber production workers usually work in open environments, where harsh conditions such as rough terrain and extreme weather. Lack of experience, improper equipment and other related factors can negatively impact occupational health and safety (OHS). In this study, work posture analysis was performed by using computer vision technique. Real-time joint angles were calculated using a MediaPipe-supported system based on machine learning (ML) technique. Real-time joint angle data extracted from video frames were used to compute Rapid Entire Body Assessment (REBA) scores instantly. Besides the entire body analysis, each limb could be assessed separately. The initial user impressions regarding the system were assessed. This method enables the collection of detailed data, which can regularly accumulate into a largescale dataset appropriate for big data applications. CV-based work posture analysis is promising technique as a method open to development as an adaptive system. The risk of biased evaluations in expert observations during body posture analysis can be minimized.

본 연구는 컴퓨터 비전의 폭력장면 자동검출 기술을 활용하여 미디어 폭력성의 새로운 연구방법을 제시한다. 커뮤니케이션학과 컴퓨터 공학의 접점에서 미디어의 폭력성을 재개념화하고 효율적인 특징추출작업, 장면분류작업이 가능한 모바일넷 버전2(MobileNetV2) CNN 알고리즘을 활용하여 한국 드라마의 특성을 반영한 자동검출 프로그램을 개발하였다. 드라마 <펜트하우스> 두 편에 나타난 폭력장면을 자동검출 프로그램으로 분석해 성능 확인 후, 인간 연구자들이 물리적 폭력이라고 내용분석한 결과와 비교분석했다. 그 결과, 폭력 장면의 검출을 기준으로 정확도 81.8%, 정밀도 9.4%, 재현율 22.45%로 나타났다. 자동검출 프로그램의 폭력장면 미검출된 위음성 비율은 4.81%, 과도하게 검출된 위양성 비율은 13.39%이었다. 과도한 검출의 특징은 부정적 표정 빠른 움직임 변화, 다수 출연자 등 미디어 폭력의 특성과 관련된 요소들로, 각 특징 검출에 알맞은 알고리즘 추가를 보완책으로 제시하였다. 본 연구는 미디어 영상에 나타난 폭력장면의 검출방식이 기존 인간 연구자의 직접 분석에서 자동검출 프로그램 이용으로 변화하는 시점에서 자동검출 프로그램을 개발, 실제 방영된 드라마를 분석해 유효성 검증을 거치고, 인간 연구자가 폭력으로 판단한 장면과 비교분석했다는 의의가 있다.