Earticle

본 연구는 자율주행차의 안전한 운행을 위해 도로 표면 상태 점검의 중요성을 강조한다. 분석 결과, 자율주행에 장애를 줄 수 있는 주요 요인으로는 ‘급커브’와 ‘차선 문제’(예: 차선 마모 및 구조적 손상)가 도출되었다. 그러나 유지보수를 위한 차선과 기타 객체의 구분에 대한 연구는 부족하여 전이학습에 어려움이 존재하였다. 이를 해결하기 위해 새로운 데이터셋과 차 선 식별 알고리즘을 개발하였으며, 해당 알고리즘은 정밀도, 재현율, mAP@0.5 지표에서 91% 에서 96% 사이의 높은 성능을 보였다. 특히 ‘차선’, ‘노면표시’, ‘화살표’, ‘안전지대’ 등 주요 클래스는 92%~97%의 정확도를 달성하였다. 또한 KITTI와 Tusimple 데이터셋을 활용한 검증을 통해 도로 이상 탐지에서 높은 정확도를 확인하였다. 기존 연구와 비교했을 때 본 알고리즘은 안정적인 차선 식별에서 90%의 정밀도를 달성하며 성능 향상을 입증하였다. 이러한 결과는 차선 상태 평가를 위한 핵심 자료로 활용될 수 있으며, YOLOv5xl 알고리즘 기반의 도로 관리 및 자율주행 인지 향상, 자동화된 차선 유지보수를 통한 도로 안전성 제고에 기여할 수 있다.

This study highlights the importance of inspecting road surfaces to ensure safe autonomous vehicle operation. Analysis identified "sharp curves" and "lane issues" (e.g., faded markings and structural damage) as key factors that can disrupt autonomous driving. However, limited research on distinguishing lanes from other objects for maintenance posed challenges for transfer learning. To address this, a new dataset and a lane identification algorithm were developed. The algorithm showed high performance, with precision, recall, and mAP@0.5 metrics ranging from 91% to 96%, and key classes like 'lane,' 'road marking,' 'arrow,' and 'safety zone' achieving 92%-97% accuracy. Validation with KITTI and Tusimple datasets confirmed high accuracy in detecting road issues. Compared to previous studies, this algorithm demonstrated improved performance, reaching 90% precision for stable lane identification. These findings provide essential data for assessing lane conditions, supporting road management and autonomous vehicle perception using the YOLOv5xl algorithm, and enhancing road safety through automated lane maintenance.

의료인공지능의 알고리즘을 적용하는 과정에서 진료행위의 투명성 부족, 의료 알고리즘의 편견 유도, 보안 부족으로 인한 문제 등이 발생한다. 이러한 문제로 인한 위험을 예방하는 과정에서 전통적 규제 결함을 비롯하여 의료인공지능의 민사주체로서 지위의 모호함과 의료 알고리즘 권한에 대한 통제의 불균형, 그리고 의료행위 책임주체의 추적에 대한 어려움 등 현실적으로 해결해야 할 것들 이 나타난다. 이는 의료인공지능에 대한 기술적 도전으로 법적 구제수단을 명확 히 할 필요가 있다. 즉, 의료인공지능에 대해 기존의 의료불법행위에 기초하여 새로운 의료행위의 개념을 세워야 한다. 이에 본 연구는 의료인공지능의 알고리즘에 대해 어떻게 규제할 것인지를 중 점으로 의료인공지능의 발전를 위한 입법적 해결방안을 제시하고자 하였다. 우 선, 알고리즘 규제의 관점에서 의료 알고리즘의 발전 현황을 바탕으로 의료인공 지능 알고리즘을 하나의 제품으로 간주하였다. 즉, 의료인공지능 알고리즘을 인 공지능기술의 응용 발전에 대한 방향성 지침을 명확히 하기 위함이다. 의료인공 지능이 궁극적으로 과학기술의 발전으로 실현될 수 있는지는 여러 요인에 의해 영향을 받는다. 이에 관련법의 역할은 (1) 환자의 권리를 보호하고, (2) 의료사 고를 통제 또는 회피하는데 중점을 두는지, (3) 의료 알고리즘의 활용에 있어 R&D 설계자가 의료데이터의 규범성 및 의료 알고리즘 제품 운영의 안전성을 검토하도록 하고, (4) 의료행위로 인한 침해행위에 대해 책임을 부담하도록 하 는 것이다. 이러한 법적 보장을 통해 의료 알고리즘의 안전성과 신뢰성을 제고 할 수 있다. 본 연구에서는 이러한 점에 기초하여 의료인공지능의 알고리즘 규제에 대한 입법에 있어 환자의 권리보호를 중심으로 검토함으로써 향후 의료인공지능이 의료분야에서 긍정적 발전을 촉진할 수 있는 좋은 자료가 될 것이라 본다.

In the application of medical artificial intelligence algorithm, there are risks such as limited autonomy of diagnosis and treatment activities, lack of transparency in diagnosis and treatment process, biased induction of medical algorithm and lack of security. In the process of risk prevention, there are some practical challenges, such as the defects of traditional supervision, the vague definition of subject qualification, the imbalance of algorithmic power control, and the difficulty in tracing the responsible subject, which urgently need to be relieved by legal means. Based on the integrity of the legal system and the particularity of medical application scenarios. This paper studies the regulation of medical artificial intelligence algorithm from the perspective of algorithm regulation. Based on the development status of artificial intelligence algorithm, it regards medical artificial intelligence algorithm as a product and provides directional guidance for the application and development of medical artificial intelligence. Whether the medical artificial intelligence can finally realize the improvement of science and technology will be affected by many factors. The role of law in it focuses on preventing, controlling or avoiding the algorithm risks to patients' rights. When using the medical artificial intelligence algorithm, the R&D designer needs to review the standardization of medical data and the safety of medical algorithm products and bear the tort liability to ensure the safety and reliability of medical algorithms. For the effective prevention and control of medical artificial intelligence algorithm risk, it is necessary to clarify the dilemma of medical artificial intelligence algorithm risk prevention and control in order to make accurate policies. This paper holds that the legislation of medical artificial intelligence algorithm regulation in China should focus on the protection of patients' rights. Legislation should choose to improve the relevant content of algorithm regulation in the existing legislation related to patient rights protection and algorithm regulation on the basis of maintaining the integrity and systematization of legislation, so as to promote the benign development of artificial intelligence algorithm in the medical field.

医疗人工智能算法应用过程中存在诊疗活动自主性受限、诊疗过程透明性匮乏、 医疗算法偏见性诱导以及安全性缺失风险。在风险防治过程中出现了传统监管缺 陷、主体资格界定模糊、算法权力控制失衡、责任主体追溯难等现实挑战，迫切需 要运用法律手段予以纾解。基于对法律体系完整性和医疗应用场景的特殊性。本文 对医疗人工智能算法的规制研究以算法规制为视角，基于人工智能算法发展现状， 将医疗人工智能算法视为产品，对医疗人工智能的应用发展提供方向性指引。医疗 人工智能最终能否实现科技向善将受到多种因素影响，法律在其中的作用集中于预 防、控制或避免算法风险对患者权利，医疗人工智能算法使用时，研发设计者需要 对医疗数据的规范性和医疗算法产品运行的安全性进行审查并承担侵权责任，以确 保医疗算法的安全可靠。对于医疗人工智能算法风险的有效防治，需要厘清医疗人 工智能算法风险防治的困境，以便精准施策。本文认为，我国医疗人工智能算法规 制立法应当以患者权利保障为核心。立法应当选择在保持立法的完整性和体系性的 基础上，针对算法风险，将算法规制的相关内容在已有的患者权利保障和算法规制 相关的立法当中进行完善，从而促进人工智能算法在医疗领域良性发展。

사람마다 수면장애의 정도가 다르기 때문에 개인별로 각기 다른 맞춤형 수면케어 서비스가 필요하다. 뇌파 데이터는 사람마다 고유한 파형 특성을 보이기 때문에 이 특성을 이용하면 개인을 식별할 수 있다. 개인식별은 맞춤형 서비스를 가능하게 해주는 중요한 역할을 제공한다. 눈을 깜박일 때 전두엽 부위를 측정하면 뇌파특성을 획득할 수 있다. 따라서 본 논문에서는 맞춤형 수면케어 서비스를 위한 눈 깜빡임 EOG(Electrooculography) 기 반의 실시간 개인식별 알고리즘을 제안한다. 평가를 위해 10명을 대상으로 개인식별 정확도 실험을 하였다. 실험결 과 최대 93%의 정확도를 확인하였다. 향후 외부 환경 변화와 같은 특성을 반영하여 알고리즘을 발전시킬 수 있을 것이다.

Customized sleep care service needs to be provided differently for individuals since individual has different degree of sleep disorder. Because the brainwave data shows unique waveform characteristics for each person, this characteristic can be used to identify individuals. Personal identification provides an important role in enabling customized services. When you blink, you can obtain brain wave characteristics by measuring the area of the frontal lobe. Therefore, a real-time personal identification algorithm based on blinking EOG for customized sleep care service is proposed in this paper. For evaluation, 10 individuals were tested for personal identification accuracy. The results of the experiment confirmed that a maximum accuracy of 93% were taken. Algorithms can be developed by reflecting characteristics such as changes in the external environment in the future.

본 논문은 C4.5 알고리즘을 이용한 피로 인식 방법을 제안한다. 피로 평가에 대한 국내외 연구를 바탕으로 중국인의 생활습관 및 문화적 특성과 결합하여 피로 자기평가 척도를 완성하였다. 본문에서 사용한 척도는 58개 하위 항목에 적용되어 있으며 피로의 유형과 정도를 평가하는 데 사용되었다. 이 항목들은 육체적 피로, 정신적 피로, 개인의 습관 및 피로의 결과 등을 측정하는 4가지 분류 항목에 포함된다. 본 연구의 목적은 피로 형성의 주요 원인을 분석하고 그에 따른 피로 정도를 인식함으로써 피로에 대한 주관적 관심을 증가시키고 과도한 피로로 인한 심뇌혈관계 질환의 위험을 감소시키는 데 있다. C4.5 알고리즘을 활용한 피로 인식 시스템의 인식률은 평균 85%로 나타나 본 제안의 유용성을 확인하였다.

This paper proposes a fatigue recognition method using the C4.5 algorithm. Based on domestic and international studies on fatigue evaluation, we have completed the fatigue self - assessment scale in combination with lifestyle and cultural characteristics of Chinese people. The scales used in the text were applied to 58 sub items and were used to assess the type and extent of fatigue. These items fall into four categories that measure physical fatigue, mental fatigue, personal habits, and fatigue outcomes. The purpose of this study is to analyze the leading causes of fatigue formation and to recognize the degree of fatigue, thereby increasing the personal interest in fatigue and reducing the risk of cerebrovascular disease due to excessive fatigue. The recognition rate of the fatigue recognition system using the C4.5 algorithm was 85% on average, confirming the usefulness of this proposal.

본 연구에서는 재범위험성 평가 도구와 관련한 국내 현황과 한계점에 대하여 설명 하고 이를 위한 새로운 통계적 접근 방법을 제시하고자 하였다. 현재 미국에서 활발한 연구가 이루어지고 있는 머신 러닝 기법인 의사결정 나무(decision tree) 분석과 랜덤 포레스트 기법을 도입하여 한국 성범죄자 위험성 평가척도인 K-SORAS의 개정판을 이 용하여 수집한 데이터 샘플에서 성범죄 재범을 일으키는 가장 큰 위험요인을 확인하였다. 그 결과 KSORAS의 문항 중 8번 문항인 ‘감독 기간 내 문제 행동’이 재범과 가장 관련이 큰 것을 확인할 수 있었다. 이를 통해 한국의 재범위험성 평가가 가진 한계점을 보완하는 방법을 고찰해 보았다. 이와 같이 재범위험성 평가에 머신 러닝 기법을 함께 활용하여 분석하는 연구들은 현재 국내에서는 아직 그 흔적을 찾기가 어려운 상황이다. 따라서 본 연구가 그 발판이 되어 머신 러닝 기법 등을 재범위험성 평가 등에 적용하여 분석해보는 활발한 추가 연구가 이루어지게 되기를 바란다.

The purpose of this study is to identify recidivism risk factors of KSORAS (Korean Sex Offender Risk Assessment Scale) by using machine learning method, especially decision tree analysis and random forest algorithm which are currently being studied actively in the US. Also the limitation of existing risk assessments is explained. By using revised version of KSORAS data of the offenders who are under electronic monitoring, it was feasible to figure out the most related factors among 55 other factors to recidivism. The most affective factor to recidivism was the item number 8 which is problematic behavior while supervision period. This statistical approach is expected to be a latest way of making up for any weak points of existing risk assessments.

Recently bank card number recognition plays an important role in improving payment efficiency. In this paper we propose a new bank-card number identification algorithm. The proposed algorithm consists of three modules which include edge detection, candidate region generation, and recognition. The module of ‘edge detection’ is used to obtain the possible digital region. The module of ‘candidate region generation’ has the role to expand the length of the digital region to obtain the candidate card number regions, i.e. to obtain the final bank card number location. And the module of ‘recognition’ has Convolutional deep learning Neural Network (CNN) to identify the final bank card numbers. Experimental results show that the identification rate of the proposed algorithm is 95% for the card numbers, which shows 20% better than that of conventional algorithm or method.

In this paper, a network traffic identification model is established using a multilayer excitation function quantum neural network which is suitable for data classification. Because the conventional quantum neural network has different target function in the training of the weights of the network and the sigmoid function of the neurons in the hidden layer, the coupling effect of the two parameters is not processed. This will result in the middle and later stage of the training iteration process, and it may be possible to reduce the objective function value of a kind of parameter, and make the objective function value of another kind of parameter increase. In order to avoid this situation, using LM algorithm to optimize, using the same objective function not only as the target function of the network weight, but also the function of translational spacing of sigmoid function of neurons in the hidden layer, and the training objective is to minimize the sum of squared error of the neural network output and the desired value. Finally, the recognition performance of the proposed algorithm is compared with that of the conventional quantum neural network and LM-BP neural network. The results show that the convergence rate of the proposed algorithm is the fastest and the convergence accuracy is the highest.

Based on the second generation wavelet and information entropy, in this paper, we recognize the partial discharge pattern using the second generation wavelet (SGWT) and adaptive BP. Firstly, feature extraction of discharge signals are obtained using the SGWT and information entropy. Then, the extracted features are feed into the training BP network. The learning algorithm employed the conjugate gradient methods and the adaptive adjustment to train the error for BP network. Finally, we get the optimum training network, and the simulation results verified the feasibility of the algorithm.

Urban road traffic state identification is a key link to realize the intelligent transportation based on the Internet of Vehicles, and accurately positioning vehicles is the foundation to realize the traffic state identification. Aiming at the problem that GPS has signal blind area in positioning vehicles, a vehicle positioning algorithm based on particle filter was proposed, it could improve the traditional algorithm on degradation and large amount of calculations; Based on vehicle positioning, an urban road traffic state identification algorithm based on fuzzy discrimination was proposed, it could comprehensively consider multiple factors’ influence on traffic state. The experiment results show that the improved particle filter algorithm’s mean squared error has increased about 55.437% compared with GPS method, and the traffic state identification algorithm can accurately identify the traffic state of the study area, it can prove that the urban road traffic state identification algorithm based on particle filter and fuzzy discrimination is feasible and effective.

In this work, we use the formulas of statistic techniques for developing an algorithm based on third order moments and autocorrelation function. This algorithm permits to identify non linear system coefficients for recovering the real information from input-output systems. Simulation examples and comparison with other method in the literature are provided to verify the performance of the developed algorithm. The obtained results demonstrate the efficiency and the accuracy of the developed algorithm for non linear system identification under various values of signal to noise ratio (SNR) and different sample sizes N. To corroborate the theoretical results for a real process, we applied the developed algorithm to search a model able to represent the internet traffic data.

This paper presents a set of programme to extract the application-layer protocol features. Based on frequent itemsets mining, the program automatically extracts four most common features of a protocol: characteristic string, session tag, packet length, and packet order. It is experimentally demonstrated that this progran can significantly improve the efficiency of feature extraction, and can be extended to other areas such as intrusion detection and extracting worm signature.

The challenges faced by the today’s network are significantly increased. They decrease the efficiency of the network. If enough measures are not taken, this will lead to severe problems in future generation networks. Though many parameters are involved in improving the efficiency, this paper focuses on distance and bandwidth. The proposed algorithm identifies the shortest route faster than Dijkstra’s algorithm. It suggests new ideas in constructing source-todestination route by refining the steps in existing algorithm. It also identifies the bandwidtheffective route using proposed algorithm. Ultimately, the outcome of this paper increases the efficiency of the network.

In order to restrain the bad influences of fake and inferior products, an automatic identification authentic work anti-counterfeiting algorithm is proposed in this paper. The authentic work anti-counterfeiting uses a unique random character structure to carry out security. The algorithm is convenient for customers to distinguish the anti-counterfeiting tag by both naked eye and mobile. There is not yet a similar algorithm published. In the algorithm, the feature vectors are extracted using DWT-DCT. Using extracted feature vectors from anti-counterfeiting tag to realize automatic identification. Therefore, this algorithm saves the storage space and improves the identification rate. In addition, the experimental results show this algorithm have strong robustness under common attacks, geometrical attacks.

The PS algorithm divides the tags within the reader’s identification range into smaller groups by increasing the transmission power incrementally and identifies them. This algorithm uses the fixed frame size at every scan. Therefore, it has problems that the performance can be variously shown according to the number of tags, frame size, and power level increase. In this paper, we propose an EPS algorithm that allocates the optimal frame size by estimating the number of tags at each scan. The simulation results showed that the identification delay of EPS algorithm could be improved 70% compared with PS algorithm. It also provided a stable identification delay regardless of power level increase.

Genetic algorithm was used to identify the damage of frame structure. Stiffness coefficient damage factor is selected as design variable, and the weighted array difference value between inherent frequency and vibration mode of structure calculated and measured. According to the difficulty in selecting crossover rate and mutational rate for fundamental algorithm, the process of selection operator, crossover operator and mutation operator was improved. All operators were operated on parent individual. Crossover rate and mutational rate were set for 100%. Punishment function was applied for keeping the difference among individuals. The improved genetic algorithm can conserve the better individual in parent and keep off fall into local optimum. Through a 3-story frame with single variable damage and multiple variables damage study, the results showed that the improved genetic algorithm can identify the damage location and degree.

Recently, the use of RFID(Radio Frequency Identification) for object identification is used more often, but the tag collision problem by the use of a radio frequency still exists. Therefore, in this paper, we analyzed tag identification time to minimize tag collisions and improve tag identification time by applying various tag anti-collision algorithms in the suggested method. As a result, we drew a conclusion about the number of optimum readers used in the RFID system environment. Significantly, we expect that the suggested method will be used more efficiently in the simultaneous multi-tags identification RFID system environment.

Currently, one of the fields leading the 4th industrial revolution is the image recognition field of artificial intelligence, which is showing good results in many fields. In this paper, using is a YOLO V2 model, which is one of the image recognition models, we intend to classify and select into three types according to the characteristics of fruits. To this end, it was designed to proceed the number of iterations of learning 9000 counts based on 640 mandarin image data of 3 classes. For model evaluation, normal, rotten, and unripe mandarin oranges were used based on images. We as a result of the experiment, the accuracy of the learning model was different depending on the number of learning. Normal mandarin oranges showed the highest at 60.5% in 9000 repetition learning, and unripe mandarin oranges also showed the highest at 61.8% in 9000 repetition learning. Lastly, rotten tangerines showed the highest accuracy at 86.0% in 7000 iterations. It will be very helpful if the results of this study are used for fruit farms in rural areas where labor is scarce.