• International Journal of Internet, Broadcasting and Communication
• /
• 제12권3호
• /
• pp.220-225
• /
• 2020

This paper investigates the non-intrusive speech intelligibility estimation method in noise environments when the bottleneck feature of autoencoder is used as an input to a neural network. The bottleneck feature-based method has the problem of severe performance degradation when the noise environment is changed. In order to overcome this problem, we propose a novel non-intrusive speech intelligibility estimation method that adds the noise environment information along with bottleneck feature to the input of long short-term memory (LSTM) neural network whose output is a short-time objective intelligence (STOI) score that is a standard tool for measuring intrusive speech intelligibility with reference speech signals. From the experiments in various noise environments, the proposed method showed improved performance when the noise environment is same. In particular, the performance was significant improved compared to that of the conventional methods in different environments. Therefore, we can conclude that the method proposed in this paper can be successfully used for estimating non-intrusive speech intelligibility in various noise environments.

• Journal of Information Processing Systems
• /
• 제13권6호
• /
• pp.1467-1486
• /
• 2017

Fuzzy co-clustering is sensitive to noise data. To overcome this noise sensitivity defect, possibilistic clustering relaxes the constraints in FCM-type fuzzy (co-)clustering. In this paper, we introduce a new possibilistic fuzzy co-clustering algorithm based on information bottleneck (ibPFCC). This algorithm combines fuzzy co-clustering and possibilistic clustering, and formulates an objective function which includes a distance function that employs information bottleneck theory to measure the distance between feature data point and feature cluster centroid. Many experiments were conducted on three datasets and one artificial dataset. Experimental results show that ibPFCC is better than such prominent fuzzy (co-)clustering algorithms as FCM, FCCM, RFCC and FCCI, in terms of accuracy and robustness.

• 한국멀티미디어학회논문지
• /
• 제25권1호
• /
• pp.72-81
• /
• 2022

Visual Object Tracking is known as the most fundamental problem in the field of computer vision. Object tracking localize the region of target object with bounding box in the video. In this paper, a custom CNN is created to extract object feature that has strong and various information. This network was constructed as a Siamese network for use as a feature extractor. The input images are passed convolution block composed of a bottleneck layers, and features are emphasized. The feature map of the target object and the search area, extracted from the Siamese network, was input as a local proposal network. Estimate the object area using the feature map. The performance of the tracking algorithm was evaluated using the OTB2013 dataset. Success Plot and Precision Plot were used as evaluation matrix. As a result of the experiment, 0.611 in Success Plot and 0.831 in Precision Plot were achieved.

• Journal of Information Processing Systems
• /
• 제19권6호
• /
• pp.778-790
• /
• 2023

Although density peak clustering can often easily yield excellent results, there is still room for improvement when dealing with complex, high-dimensional datasets. One of the main limitations of this algorithm is its reliance on geometric distance as the sole similarity measurement. To address this limitation, we draw inspiration from the information bottleneck theory, and propose a novel density peak clustering algorithm that incorporates this theory as a similarity measure. Specifically, our algorithm utilizes the joint probability distribution between data objects and feature information, and employs the loss of mutual information as the measurement standard. This approach not only eliminates the potential for subjective error in selecting similarity method, but also enhances performance on datasets with multiple centers and high dimensionality. To evaluate the effectiveness of our algorithm, we conducted experiments using ten carefully selected datasets and compared the results with three other algorithms. The experimental results demonstrate that our information bottleneck-based density peaks clustering (IBDPC) algorithm consistently achieves high levels of accuracy, highlighting its potential as a valuable tool for data clustering tasks.

• 한국정보통신학회논문지
• /
• 제25권12호
• /
• pp.1835-1845
• /
• 2021

본 논문에서는 블랙 아이스를 정확하게 인식하고 도로 노면 정보를 운전자에게 미리 알려줘서 속도를 제어하고 예방 조치를 취할 수 있도록 하기 위해 열화 도로 영상을 기반으로 블랙 아이스 검출하기 위해 lightweight 네트워크를 제안한다. 전이학습을 이용하여 블랙 아이스 인식 실험을 하였고, 블랙 아이스 인식의 정확도 향상을 위해 MobileNetV2 기반의 개선된 lightweight 네트워크를 개발하였다. 계산량을 줄이기 위해 Linear Bottleneck 및 Inverted Residuals를 활용하여 4개의 Bottleneck 그룹을 사용하고 모델의 인식률 향상을 위해 각 Bottleneck 그룹에 3×3 컨볼루션 레이어를 연결하여 지역적 특징 추출을 강화하고 특징 맵의 수를 늘렸다. 마지막으로 구축된 블랙 아이스 데이터 세트 대상으로 블랙 아이스 인식 실험을 진행하였으며, 제안된 모델은 블랙 아이스에 대해 99.07%의 정확한 인식률을 나타내었다.

• 전기전자학회논문지
• /
• 제25권3호
• /
• pp.493-500
• /
• 2021

본 논문에서는 다각형 용기의 품질 향상을 위한 딥러닝 구조 개발을 제안한다. 딥러닝 구조는 convolution 층, bottleneck 층, fully connect 층, softmax 층 등으로 구성된다. Convolution 층은 입력 이미지 또는 이전 층의 특징 이미지를 여러 특징 필터와 convolution 3x3 연산하여 특징 이미지를 얻어 내는 층이다. Bottleneck 층은 convolution 층을 통해 추출된 특징 이미지상의 특징들 중에서 최적의 특징들만 선별하여 convolution 1x1 ReLU로 채널을 감소시키고convolution 3x3 ReLU를 실시한다. Bottleneck 층을 거친 후에 수행되는 global average pooling 연산과정은 convolution 층을 통해 추출된 특징 이미지의 특징들 중에서 최적의 특징들만 선별하여 특징 이미지의 크기를 감소시킨다. Fully connect 층은 6개의 fully connect layer를 거쳐 출력 데이터가 산출된다. Softmax 층은 입력층 노드의 값과 연산을 진행하려는 목표 노드 사이의 가중치와 곱을 하여 합하고 활성화 함수를 통해 0~1 사이의 값으로 변환한다. 학습이 완료된 후에 인식 과정에서는 학습 과정과 마찬가지로 카메라를 이용한 이미지 획득, 측정 위치 검출, 딥러닝을 활용한 비원형 유리병 분류 등을 수행하여 비원형 유리병을 분류한다. 제안된 다각형 용기의 품질 향상을 위한 딥러닝 구조의 성능을 평가하기 위하여 공인시험기관에서 실험한 결과, 양품/불량 판별 정확도 99%로 세계최고 수준과 동일한 수준으로 산출되었다. 검사 소요 시간은 평균 1.7초로 비원형 머신비전 시스템을 사용하는 생산 공정의 가동 시간 기준 내로 산출되었다. 따라서 본 본문에서 제안한 다각형 용기의 품질 향상을 위한 딥러닝 구조의 성능의 그 효용성이 입증되었다.

• 한국지능시스템학회:학술대회논문집
• /
• 한국퍼지및지능시스템학회 1996년도 추계학술대회 학술발표 논문집
• /
• pp.315-318
• /
• 1996

The machine-part cell formation means the grouping of similar parts and similar machines into families in order to minimize bottleneck machines, bottleneck parts, and inter-cell part movements in cellular manufacturing systems and flexible manufacturing systems. The cell formation problem is knows as a kind of NP complete problems. This paper briefly introduces the cell-formation problem and proposes a cell formation method based on the Kohonen's self-organizing feature map which is a neural network model. It also shows some experiment results using the proposed method. The proposed method can be easily applied to the cell formation problem compared to other meta-heuristic based methods. In addition, it can be used to solve large-scale cell formation problems.

• 대한전자공학회:학술대회논문집
• /
• 대한전자공학회 2002년도 ITC-CSCC -2
• /
• pp.798-801
• /
• 2002

This paper proposes unsupervised classification of multiple attributes via five-layer autoassociative neural network with bottleneck layer. In the conventional methods, high dimensional data are compressed into low dimensional data at bottleneck layer and then feature extraction is performed (Fig.1). In contrast, in the proposed method, analog data is compressed into digital data. Furthermore bottleneck layer is divided into two segments so that each attribute, which is a discrete value, is extracted in corresponding segment (Fig.2).

• 한국CDE학회논문집
• /
• 제12권3호
• /
• pp.182-192
• /
• 2007

The reduction of product development time is exposed to the competitive pressure due to shortened product-and production technology lifecycles as well as increasingly dynamic markets. Specially in automobile companies, that is of major importance for designing die because it is a bottleneck process in the development of a new car. To improve this conventional design process, this paper describes how to design it fast and flexibly. This was done by a modular method using standard template and a feature and knowledge based design method along the design process.

• 말소리와 음성과학
• /
• 제7권4호
• /
• pp.9-16
• /
• 2015

In this paper, several experiments are performed according to deep neural network (DNN) based features for the performance comparison of speaker verification (SV) systems. To this end, input features for a DNN, such as mel-frequency cepstral coefficient (MFCC), linear-frequency cepstral coefficient (LFCC), and perceptual linear prediction (PLP), are first compared in a view of the SV performance. After that, the effect of a DNN training method and a structure of hidden layers of DNNs on the SV performance is investigated depending on the type of features. The performance of an SV system is then evaluated on the basis of I-vector or probabilistic linear discriminant analysis (PLDA) scoring method. It is shown from SV experiments that a tandem feature of DNN bottleneck feature and MFCC feature gives the best performance when DNNs are configured using a rectangular type of hidden layers and trained with a supervised training method.