• 한국지능시스템학회논문지
• /
• 제25권1호
• /
• pp.15-21
• /
• 2015

커널에 대한 모수의 조절은 서포트 벡터 기계의 일반화 능력에 영향을 준다. 이와 같이 모수들의 적절한 값을 결정하는 것은 종종 어려운 작업이 된다. 서포트 벡터 회귀에서 이와 같은 모수들의 값을 결정하기 위한 부담은 앙상블 학습을 사용함으로써 감소시킬 수 있다. 그러나 대용량의 자료에 대한 문제에 직접적으로 적용하기에는 일반적으로 시간 소모적인 방법이다. 본 논문에서 서포트 벡터 회귀의 모수 조절에 대한 부담을 감소하기 위하여 원래 자료집합을 유한개의 부분집합으로 분해하는 방법을 제안하였다. 제안하는 방법은 대용량의 자료들인 경우와 특히 불균등 자료 집합에서 효율적임을 보일 것이다.

• 한국멀티미디어학회논문지
• /
• 제21권2호
• /
• pp.98-107
• /
• 2018

In this paper, we propose a method to classify the skeletal maturity with a small amount of hand wrist X-ray image using deep learning-based meta-learning. General deep-learning techniques require large amounts of data, but in many cases, these data sets are not available for practical application. Lack of learning data is usually solved through transfer learning using pre-trained models with large data sets. However, transfer learning performance may be degraded due to over fitting for unknown new task with small data, which results in poor generalization capability. In addition, medical images require high cost resources such as a professional manpower and mcuh time to obtain labeled data. Therefore, in this paper, we use meta-learning that can classify using only a small amount of new data by pre-trained models trained with various learning tasks. First, we train the meta-model by using a separate data set composed of various learning tasks. The network learns to classify the bone maturity using the bone maturity data composed of the radiographs of the wrist. Then, we compare the results of the classification using the conventional learning algorithm with the results of the meta learning by the same number of learning data sets.

• 방송공학회논문지
• /
• 제27권2호
• /
• pp.185-197
• /
• 2022

최근 객체 인식에 높은 성능을 가진 딥러닝 네트워크가 나오고 있다. 딥러닝을 이용한 객체 인식의 경우 성능 향상을 위해 학습 데이터 셋 구축이 중요하다. 데이터 셋을 구축하기 위해서는 이미지를 수집하고 라벨링 해야 한다. 이 과정은 많은 시간과 인력이 필요하다. 때문에 오픈 데이터 셋을 사용한다. 그러나 방대한 오픈 데이터 셋을 가지고 있지 않는 객체도 존재한다. 그 중 하나가 번호판 검출과 인식에 필요한 데이터이다. 이에 본 논문에서는 이미지를 최소화 하여 대용량 데이터 셋을 만들 수 있는 인조 번호판 생성기 시스템을 제안한다. 또한 인조 번호판 배치구조에 따른 검출률을 분석했다. 분석결과 가장 좋은 배치구조는 FVC_III, B이며 가장 적합한 네트워크는 D2Det이었다. 인조 데이터셋 성능은 실제 데이터셋의 성능보다 2~3%가 낮았지만, 인조 데이터를 구축하는 시간이 실제 데이터셋을 구축하는 시간보다 약 11배 빨라 시간적으로 효율적인 데이터 셋 구축 시스템임을 증명하였다.

• 대한방사선기술학회지:방사선기술과학
• /
• 제47권1호
• /
• pp.29-37
• /
• 2024

This study aims to evaluate the performance of the U-Net based learning model that may vary depending on the histogram equalization algorithm. The subject of the experiment were 17 radiology students of this college, and 1,727 data sets in which the region of interest was set in the thyroid after acquiring ultrasound image data were used. The training set consisted of 1,383 images, the validation set consisted of 172 and the test data set consisted of 172. The equalization algorithm was divided into Histogram Equalization(HE) and Contrast Limited Adaptive Histogram Equalization(CLAHE), and according to the clip limit, it was divided into CLAHE8-1, CLAHE8-2. CLAHE8-3. Deep Learning was learned through size control, histogram equalization, Z-score normalization, and data augmentation. As a result of the experiment, the Attention U-Net showed the highest performance from CLAHE8-2 to 0.8355, and the U-Net and BSU-Net showed the highest performance from CLAHE8-3 to 0.8303 and 0.8277. In the case of mIoU, the Attention U-Net was 0.7175 in CLAHE8-2, the U-Net was 0.7098 and the BSU-Net was 0.7060 in CLAHE8-3. This study attempted to confirm the effects of U-Net, Attention U-Net, and BSU-Net models when histogram equalization is performed on ultrasound images. The increase in Clip Limit can be expected to increase the ROI match with the prediction mask by clarifying the boundaries, which affects the improvement of the contrast of the thyroid area in deep learning model learning, and consequently affects the performance improvement.

• 한국인터넷방송통신학회논문지
• /
• 제18권5호
• /
• pp.171-177
• /
• 2018

최근 많은 이미지 데이터 셋들은 일반적인 특성을 추출하기 위한 다양한 데이터 클래스와 특징을 가지고 있다. 하지만 이러한 다양한 데이터 클래스와 특징으로 인해 해당 데이터 셋으로 훈련된 물체 검출 딥러닝 모델은 데이터 특성이 다른 환경에서 좋은 성능을 내지 못하는 단점을 보인다. 이 논문에서는 하위 카테고리 기반 물체 검출 방법과 오픈셋 물체 검출 방법을 이용하여 이를 극복하고, 강인한 물체 검출 딥러닝 모델을 훈련하기 위해 능동 준지도 학습 (Active Semi-Supervised Learning)을 이용한 다중 분기 트리 구조를 제안한다. 우리는 이 구조를 이용함으로써 데이터 특성이 다른 환경에서 적응할 수 있는 모델을 가질 수 있고, 나아가 이 모델을 이용하여 이전의 모델보다 높은 성능을 확보 할 수 있다.

• 융합신호처리학회논문지
• /
• 제20권3호
• /
• pp.132-137
• /
• 2019

의료영상으로 생성된 데이터의 양은 전문적인 시각적 분석 한계를 점점 초과하여, 자동화된 의료영상 분석의 필요성이 증가되고 있는 실정이다. 이러한 이유 등으로 인하여 본 논문에서는 정상소견과 종양소견을 보이는 각각의 뇌 실질 MRI 의료영상을 이용하여 Inception V3 딥러닝 모델을 이용한 종양 유무에 따른 분류 및 정확도를 평가하였다. 연구 결과, 딥러닝 모델의 정확도 평가는 학습 데이터 세트의 경우 90%, 검증 데이터 세트의 경우 86%의 정확도를 나타내었다. 손실률 평가에서는 학습 데이터 세트의 경우 0.56, 검증 데이터 세트의 경우 1.28의 손실률을 나타내었다. 향 후 연구에서는 딥러닝 모델의 성능 향상 및 평가의 신뢰성 확보를 위하여 공개된 의료영상의 데이터를 충분히 확보하고, 라벨링 분류 작업을 통한 라벨링의 정확도를 개선하여 모델링을 구현해 볼 필요가 있다고 사료된다.

• Journal of Radiation Protection and Research
• /
• 제46권4호
• /
• pp.204-212
• /
• 2021

Background: Identification of radioisotopes for plastic scintillation detectors is challenging because their spectra have poor energy resolutions and lack photo peaks. To overcome this weakness, many researchers have conducted radioisotope identification studies using machine learning algorithms; however, the effect of data normalization on radioisotope identification has not been addressed yet. Furthermore, studies on machine learning-based radioisotope identifiers for plastic scintillation detectors are limited. Materials and Methods: In this study, machine learning-based radioisotope identifiers were implemented, and their performances according to data normalization methods were compared. Eight classes of radioisotopes consisting of combinations of ²²Na, ⁶⁰Co, and ¹³⁷Cs, and the background, were defined. The training set was generated by the random sampling technique based on probabilistic density functions acquired by experiments and simulations, and test set was acquired by experiments. Support vector machine (SVM), artificial neural network (ANN), and convolutional neural network (CNN) were implemented as radioisotope identifiers with six data normalization methods, and trained using the generated training set. Results and Discussion: The implemented identifiers were evaluated by test sets acquired by experiments with and without gain shifts to confirm the robustness of the identifiers against the gain shift effect. Among the three machine learning-based radioisotope identifiers, prediction accuracy followed the order SVM > ANN > CNN, while the training time followed the order SVM > ANN > CNN. Conclusion: The prediction accuracy for the combined test sets was highest with the SVM. The CNN exhibited a minimum variation in prediction accuracy for each class, even though it had the lowest prediction accuracy for the combined test sets among three identifiers. The SVM exhibited the highest prediction accuracy for the combined test sets, and its training time was the shortest among three identifiers.

• International journal of advanced smart convergence
• /
• 제8권3호
• /
• pp.46-53
• /
• 2019

When a partitioned structure is derived from a data set using a clustering algorithm, it is not unusual to have a different set of outcomes when it runs with a different order of data. This problem is known as the order bias problem. Many algorithms in machine learning fields try to achieve optimized result from available training and test data. Optimization is determined by an evaluation function which has also a tendency toward a certain goal. It is inevitable to have a tendency in the evaluation function both for efficiency and for consistency in the result. But its preference for a specific goal in the evaluation function may sometimes lead to unfavorable consequences in the final result of the clustering. To overcome this bias problems, the first clustering process proceeds to construct an initial partition. The initial partition is expected to imply the possible range in the number of final clusters. We apply the data centric sorting to the data objects in the clusters of the partition to rearrange them in a new order. The same clustering procedure is reapplied to the newly arranged data set to build a new partition. We have developed an algorithm that reduces bias effect resulting from how data is fed into the algorithm. Experiment results have been presented to show that the algorithm helps minimize the order bias effects. We have also shown that the current evaluation measure used for the clustering algorithm is biased toward favoring a smaller number of clusters and a larger size of clusters as a result.

• 한국지능시스템학회논문지
• /
• 제11권7호
• /
• pp.666-674
• /
• 2001

Classical methods to find a minimal set of rules based on the rough set theory are known to be ineffective in dealing with new instances added to the universe. This paper introduces an inductive learning algorithm for incrementally retrieving a minimal set of rules from a given decision table. Then, the algorithm is validated via simulations with two sets of data, in comparison with a classical non-incremental algorithm. The simulation results show that the proposed algorithm is effective in dealing with new instances, especially in practical use.

• 한국건축시공학회:학술대회논문집
• /
• 한국건축시공학회 2023년도 봄 학술논문 발표대회
• /
• pp.251-252
• /
• 2023

As technology improves with the 4th industrial revolution, smart construction is becoming a key part of safety management in the architecture and civil engineering. By using object detection technology with CCTV data, construction sites can be managed efficiently. In this study, super resolution technology based on deep learning is proposed to improve the accuracy of object detection in construction sites. As the resolution of a train set data and test set data get higher, the accuracy of object detection model gets better. Therefore, according to the scale of construction sites, different object detection models can be considered.