• KSII Transactions on Internet and Information Systems (TIIS)
• /
• 제11권2호
• /
• pp.785-804
• /
• 2017

Due to various factors such as postures, facial expressions and illuminations, face recognition by videos often suffer from poor recognition accuracy and generalization ability, since the within-class scatter might even be higher than the between-class one. Herein we address this problem by proposing a hierarchical cascaded classifier for video face recognition, which is a multi-layer algorithm and accounts for the misclassified samples plus their similar samples. Specifically, it can be decomposed into single classifier construction and multi-layer classifier design stages. In single classifier construction stage, classifier is created by clustering and the number of classes is computed by analyzing distance tree. In multi-layer classifier design stage, the next layer is created for the misclassified samples and similar ones, then cascaded to a hierarchical classifier. The experiments on the database collected by ourselves show that the recognition accuracy of the proposed classifier outperforms the compared recognition algorithms, such as neural network and sparse representation.

• Journal of Biosystems Engineering
• /
• 제40권2호
• /
• pp.153-158
• /
• 2015

Purpose: The research presented in this paper applied the chemometric analysis to the near-infrared spectral data from line-scanned hyperspectral images of beef and fish meals in animal feeds. The chemometric statistical models were developed to distinguish beef meals from fish ones. Methods: The meal samples of 40 fish meals and 15 beef meals were line-scanned to obtain hyperspectral images. The spectral data were retrieved from each of 3600 pixels in the Region of Interest (ROI) of every sample image. The wavebands spanning 969 nm to 1551 nm (across 176 spectral bands) were selected for chemometric analysis. The partial least squares regression (PLSR) and the principal component analysis (PCA) methods of the chemometric analysis were applied to the model development. The purpose of the models was to correctly classify as many beef pixels as possible while misclassified fish pixels in an acceptable amount. Results: The results showed that the success classification rates were 97.9% for beef samples and 99.4% for fish samples by the PLSR model, and 85.1% for beef samples and 88.2% for fish samples by the PCA model. Conclusion: The chemometric analysis-based PLSR and PCA models for the hyperspectral image analysis could differentiate beef meals from fish ones in animal feeds.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• 제14권12호
• /
• pp.4664-4681
• /
• 2020

Modulation recognition (MR) plays a key role in cognitive radar, cognitive radio, and some other civilian and military fields. While existing methods can identify the signal modulation type by extracting the signal characteristics, the quality of feature extraction has a serious impact on the recognition results. In this paper, an end-to-end MR method based on long short-term memory (LSTM) and the gated recurrent unit (GRU) is put forward, which can directly predict the modulation type from a sampled signal. Additionally, the sliding window method is applied to fast-changing mixed-modulation signals for which the signal modulation type changes over time. The recognition accuracy on training datasets in different SNR ranges and the proportion of each modulation method in misclassified samples are analyzed, and it is found to be reasonable to select the evenly-distributed and full range of SNR data as the training data. With the improvement of the SNR, the recognition accuracy increases rapidly. When the length of the training dataset increases, the neural network recognition effect is better. The loss function value of the neural network decreases with the increase of the training dataset length, and then tends to be stable. Moreover, when the fast-changing period is less than 20ms, the error rate is as high as 50%. As the fast-changing period is increased to 30ms, the error rates of the GRU and LSTM neural networks are less than 5%.

• 한국임상수의학회지
• /
• 제29권1호
• /
• pp.58-62
• /
• 2012

To protect aquatic animal health of importing countries from the potential risks associated with exotic diseases introduced through international trade of live aquatic animals, inspection of designated commodities at ports of entry is a critical component of the safeguarding system. The only way to be 100% confident that no fishes in a shipment are infected with a specific agent is to test every fish in the commodity imported with a perfect diagnostic test. For the majority of cases, this is unrealistic since the group of interest may very large particularly for aquatic animals, or imperfect tests are often available. It is, therefore, more common to test a fixed proportion of a group by preplanned sampling schemes. However, decision making based on results of testing the sample can provide quite a chance that infected groups may be misclassified as uninfected, depending on sampling strategy employed. The objective of this study was to determine the possibility that one or more fishes in the group imported being infected but tests negative after inspecting samples. This question is critical to government authorities to examine whether sampling plan is sufficient to achieve the purpose intended for. At fixed population size, the maximum number of infected fishes when all tests negative was decreased as the sampling fraction increased. The probability of including at least one undetected but infected fish in a group for negative tests increased with the number of fish tested or true prevalence. The risk was much lesser where high sensitivity test was assumed; when increasing test sensitivity from 0.9 to 0.99, this risk was dramatically reduced to about a tenth or a fourth for prevalence ranges from 2 to 10%, given sample size ranges from 10 to 200. Based on the preliminary analysis, the author concluded that current sampling plan testing 4-8% of the import proposal for human consumption still can yield high false negative results. Therefore, from the quarantine inspection point of view, an enforced commodity-specific sampling design that accounts for the cost of testing with an imperfect test at the specified design prevalence is urgent.

• 지적과 국토정보
• /
• 제49권1호
• /
• pp.103-112
• /
• 2019

위성영상 분류를 위한 관심 지역 추출은 국토 공간을 효율적으로 관리하기 위한 중요한 기술 중 하나이다. 하지만 위성영상 분류에 관한 최근의 연구들은 관심 지역을 선택하는데 있어서 영상 내의 정보에 의존하는 경우가 많다. 본 연구에서는 고해상도 영상으로부터 구축된 공간정보인 연속수치지형도를 활용하여 효과적인 관심 지역 선택 방안을 제시하였다. 본 연구에 사용된 공간정보는 국토지리정보원에서 제공하는 2013년~2017년 연속수치지형도와 환경부에서 제공하는 2015년 세종시 토지피복도를 활용하였다. 공간정보를 통해 추출된 관심 지역의 정확도 검증을 위해 2015년 10월 28일과 2018년 7월 7일 촬영된 KOMPSAT-3A호 위성영상을 사용하였다. 2013년~2015년 동안 연속수치지형도에서 변화하지 않은 영역과 2015년 토지피복지도를 사용하여 2015년 기초샘플을 추출하였다. 또한, 2015년~2017년 동안 연속수치지형도에서 변화하지 않은 영역과 2015년 토지피복지도를 사용하여 2018년 기초샘플을 추출하였다. 연속수치지형도와 토지피복도를 융합할 때 발생하는 중복된 영역은 데이터의 혼동을 방지하기 위해 모두 제거하였다. 최종적으로 관심 지역 내에서 검사점을 생성하고, 2015년, 2018년 K3A 위성영상과 오차행렬을 통해 추출된 관심 지역의 정확도를 나타냈으며 전체 정확도는 각각 약 93%, 72%로 나타났다. 관심 지역의 정확도 검증을 통해 정확하게 분류된 지역은 관심 지역으로써 사용할 수 있고 오분류된 지역은 변화탐지를 위한 참고자료로서 활용할 수 있다고 판단된다.

• 경영정보학연구
• /
• 제24권1호
• /
• pp.39-57
• /
• 2022

본 연구에서는 범주 불균형 문제가 내재된 기업부도 예측 AdaBoost 앙상블 모형의 성과를 개선하기 위하여 GMOPTBoost 알고리즘을 제안한다. AdaBoost 알고리즘은 오분류 표본에 대하여 강건한 학습기회를 제공한다는 장점이 있지만, 산술평균 정확도에 기반하기 때문에 범주 불균형 문제를 효과적으로 해결하지 못한다는 한계점이 존재한다. GMOPTBoost는 가우시안 경사하강법(Gaussian gradient descent)을 적용하여 기하평균 정확도를 최적화하고 범주 불균형 문제를 효과적으로 해결할 수 있다는 장점이 있다. 본 연구에서는 첫째, 범주 불균형 문제가 예측 모형의 성과에 미치는 효과와 GMOPTBoost의 성과 개선 효과를 검증하기 위하여 5개의 범주 불균형 데이터를 구성하였으며, 둘째, 범주 균형 데이터에 대한 GMOPTBoost의 성과 개선 효과를 검증하기 위하여 데이터 샘플링 기법을 통하여 구성된 균형 데이터를 구성하였다. 30회의 교차타당성 분석의 주요 결과는 다음과 같다. 첫째, 범주 불균형 문제는 예측 성과에 부정적인 영향을 미친다. 둘째, GMOPTBoost는 불균형 데이터에 적용된 AdaBoost의 성과를 유의적으로 개선시키는 긍정적인 효과를 제공한다. 셋째, 데이터 샘플링 기법은 성과 개선에 긍정적인 영향을 미친다. 마지막으로 데이터 샘플링 기법을 적용한 범주 균형 데이터에서도 GMOPTBoost는 유의적인 성과 개선에 기여한다.