• Proceedings of the Korea Information Processing Society Conference
• /
• 2023.11a
• /
• pp.712-715
• /
• 2023

딥러닝 기술의 발전은 고품질의 대규모 데이터에 크게 의존한다. 그러나, 데이터의 품질과 일관성을 유지하는 것은 상당한 비용과 시간이 소요된다. 이러한 문제를 해결하기 위해 최근 연구에서 최소한의 비용으로 최대의 성능을 추구하는 액티브 러닝(active learning) 기법이 주목받고 있는데, 액티브 러닝은 모델 관점에서 불확실성(uncertainty)이 높은 데이터들을 샘플링 하는데 중점을 둔다. 하지만, 레이블 생성에 있어서 여전히 많은 시간적, 자원적 비용이 불가피한 점을 고려할 때 보완이 불가피 하다. 본 논문에서는 의사-라벨링(pseudo labeling)을 활용한 준지도학습(semi-supervised learning) 방식과 학습 손실을 동시에 사용하여 모델의 불확실성(uncertainty)을 측정하는 방법론을 제안한다. 제안 방식은 레이블의 신뢰도(confidence)와 학습 손실의 최적화를 통해 비용 효율적인 데이터 레이블 생성 방식을 제안한다. 특히, 레이블 데이터의 품질(quality) 및 일관성(consistency) 측면에서 딥러닝 모델의 정확도 성능을 높임과 동시에 적은 데이터만으로도 효과적인 학습이 가능할 수 있는 메커니즘을 제안한다.

• The Journal of Bigdata
• /
• v.8 no.2
• /
• pp.177-187
• /
• 2023

Semi-supervised learning (SSL) is an effective approach to training models using a small amount of labeled data and a larger amount of unlabeled data. However, many papers in the field use a fixed threshold when applying pseudo-labels without considering the feature-wise differences among images of different classes. In this paper, we propose a SSL method for synthetic aperture radar (SAR) image classification that applies different thresholds for each class instead of using a single fixed threshold for all classes. We propose a threshold learning module into the model, considering the differences in feature distributions among classes, to dynamically learn thresholds for each class. We compare the application of a SSL SAR image classification method using different thresholds and examined the advantages of employing class-specific thresholds.

• The Journal of the Acoustical Society of Korea
• /
• v.20 no.3
• /
• pp.70-75
• /
• 2001

This paper proposes a syllabic segmentation method for the korean continuous speech. This method are formed three major steps as follows. (1) labeling the vowel, consonants, silence units and forming the Token the sequence of speech data using the segmental parameter in the time domain, pitch, energy, ZCR and PVR. (2) scanning the Token in the structure of korean syllable using the parser designed by the finite state automata, and (3) re-segmenting the syllable parts witch have two or more syllables using the pseudo-syllable nucleus information. Experimental results for the capability evaluation toward the proposed method regarding to the continuous words and sentence units are 73.5％, 85.9％, respectively.

• The Journal of the Korea Contents Association
• /
• v.6 no.11
• /
• pp.266-275
• /
• 2006

Mostly mechanical products are connected by several components instead of single accessory in product process. Although majority of assembly process is automated, the fault analysis is not automated because it needs expert knowledge in various fields to support inclusive decision-marking. This paper proposes an assembly fault analysis support system that uses image regions which can be easily accessed and understood by experts of various fields. An assembly fault analysis support system helps effective fault analysis from assembly by integrating image regions, product design information, and fault detection information. The proposed method enables fault information access from multimedia information by segmenting product images. After product images are segmented by labeling, design information and fault information are integrated in extended Attributed Relational Graph.

• Journal of the Korean Society of Fisheries and Ocean Technology
• /
• v.36 no.2
• /
• pp.117-125
• /
• 2000

The waveform and spectrum analysis of Tursiops truncatus(bottlenose dolphin) sonar signals were carried out on the basis of data collected during the dolphin show at the aquarium of Cheju Pacificland from October 1998 to February 1999. When greeting to audience, the pulse width, peak frequency and spectrum level from the five dolphins'sonar signals were 3.0ms, 4.54kHz and 125.6dB, respectively. At the time of warm-up just before the show, their figures were 5.0㎳, 5.24kHz and 127.0dB, respectively. During the performance of dolphins, with singing, peak frequency ranged 3.28∼5.78kHz and spectrum level ranged 137.0∼142.0dB. With playing ring, pulse width, peak frequency and spectrum level were 7.0㎳, 2.54kHz and 135.9dB, and when playing the ball, the values were 9.0㎳, 2.78kHz and 135.2dB, respectively. The values determined from the five dolphins during jump-up out of water were : pulse width 2.0㎳, peak frequency 4.50kHz and spectrum level 126.8dB. When they responded to trainer's instructions, the values were 2.25㎳, 248kHz and 148.7dB, respectively, and greeting to audience, the peak frequency and spectrum level were 5.84kHz and 122.5dB. During swimming under water, peak frequency and spectrum level were determined to be 10.10kHz and 126.8dB. It was found that there exited close consistencies in pulse width, frequency distribution and spectrum level between whistle sounds and dolphin's sonar signals. Accordingly, the dolphins can be easily trained by using whistle sound based on the results obtained from the waveform and spectrum of the dolphin's sonar signals.