• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.11 no.11
• /
• pp.5419-5435
• /
• 2017

The construction of panoramic images on smartphones and low-powered devices is a challenging task. In this paper, we propose a new approach for smoothly stitching images on mobile phones in the presence of moving objects in the scene. Our main contributions include handling moving object problems, reducing processing time, and generating rectangular panoramic images. First, unique and robust feature points are extracted using fast ORB method and a feature matching technique is applied to match the extracted feature points. After obtaining good matched feature points, we employ the non-deterministic RANSAC algorithm to discard wrong matches, and the hommography transformation matrix parameters are estimated with the algorithm. Afterward, we determine precise overlap regions of neighboring images and calculate their absolute differences. Then, thresholding operation and noise removal filtering are applied to create a mask of possible moving object regions. Sequentially, an optimal seam is estimated using dynamic programming algorithm, and a combination of linear blending with the mask information is applied to avoid seam transition and ghosting artifacts. Finally, image-cropping operation is utilized to obtain a rectangular boundary image from the stitched image. Experiments demonstrate that our method is able to produce panoramic images quickly despite the existence of moving objects.

• Journal of Information Processing Systems
• /
• v.19 no.5
• /
• pp.673-687
• /
• 2023

Convolutional neural network-based deep learning technology is the most commonly used in image identification, but it requires large-scale data for training. Therefore, application in specific fields in which data acquisition is limited, such as in the military, may be challenging. In particular, the identification of ground weapon systems is a very important mission, and high identification accuracy is required. Accordingly, various studies have been conducted to achieve high performance using small-scale data. Among them, the ensemble method, which achieves excellent performance through the prediction average of the pre-trained models, is the most representative method; however, it requires considerable time and effort to find the optimal combination of ensemble models. In addition, there is a performance limitation in the prediction results obtained by using an ensemble method. Furthermore, it is difficult to obtain the ensemble effect using models with imbalanced classification accuracies. In this paper, we propose a transfer learning-based feature fusion technique for heterogeneous models that extracts and fuses features of pre-trained heterogeneous models and finally, fine-tunes hyperparameters of the fully connected layer to improve the classification accuracy. The experimental results of this study indicate that it is possible to overcome the limitations of the existing ensemble methods by improving the classification accuracy through feature fusion between heterogeneous models based on transfer learning.

• Nuclear Engineering and Technology
• /
• v.56 no.6
• /
• pp.2266-2273
• /
• 2024

A machine learning platform is proposed for the diagnosis of a severe accident progression in a nuclear power plant. To predict the key parameters for accident management including lost signals, a long short term memory (LSTM) network is proposed, where multiple accident scenarios are used for training. Training and test data were produced by MELCOR simulation of the Fukushima Daiichi Nuclear Power Plant (FDNPP) accident at unit 3. Feature variables were selected among plant parameters, where the importance ranking was determined by a recursive feature elimination technique using RandomForestRegressor. To answer the question of whether a reduced order ML model could predict the complex transient response, we performed a systematic sensitivity study for the choices of target variables, the combination of training and test data, the number of feature variables, and the number of neurons to evaluate the performance of the proposed ML platform. The number of sensitivity cases was chosen to guarantee a 95 % tolerance limit with a 95 % confidence level based on Wilks' formula to quantify the uncertainty of predictions. The results of investigations indicate that the proposed ML platform consistently predicts the target variable. The median and mean predictions were close to the true value.

• Journal of the Institute of Electronics Engineers of Korea SP
• /
• v.44 no.2 s.314
• /
• pp.82-92
• /
• 2007

This paper describes a novel global shape (GS) feature based on radial basis function network (RBFN) and the extraction method of the proposed feature for 3D face recognition. RBFN is the weighted sum of RBfs, it well present the non-linearity of a facial shape using the linear combination of RBFs. It is the proposed facial feature that the weights of RBFN learned by the horizontal profiles of a face. RBFN based feature expresses the locality of the facial shape even if it is GS feature, and it reduces the feature complexity like existing global methods. And it also get the smoothing effect of the facial shape. Through the experiments, we get 94.7% using the proposed feature and hidden markov model (HMM) to match the features for 100 gallery set with those for 300 test set.

• Proceedings of the Korean Society of Broadcast Engineers Conference
• /
• 2009.01a
• /
• pp.547-550
• /
• 2009

The object recognition mechanism of human being is not well understood yet. On research of animal experiment using an ape, however, neurons that respond to simple shape (e.g. circle, triangle, square and so on) were found. And Hypothesis has been set up as human being may recognize object as combination of such simple shapes. That mechanism is called Figure Alphabet Hypothesis, and those simple shapes are called Figure Alphabet. As one way to research object recognition algorithm, we focused attention to this Figure Alphabet Hypothesis. Getting idea from it, we proposed the feature extraction algorithm for object recognition. In this paper, we described recognition of binarized images of multifont alphabet characters by the recognition model which combined three-layered neural network in the feature extraction algorithm. First of all, we calculated the difference between the learning image data set and the template by the feature extraction algorithm. The computed finite difference is a feature quantity of the feature extraction algorithm. We had it input the feature quantity to the neural network model and learn by backpropagation (BP method). We had the recognition model recognize the unknown image data set and found the correct answer rate. To estimate the performance of the contriving recognition model, we had the unknown image data set recognized by a conventional neural network. As a result, the contriving recognition model showed a higher correct answer rate than a conventional neural network model. Therefore the validity of the contriving recognition model could be proved. We'll plan the research a recognition of natural image by the contriving recognition model in the future.

• The Journal of the Acoustical Society of Korea
• /
• v.20 no.2
• /
• pp.10-15
• /
• 2001

In this paper, we describe the comparison between the combination of features using a speech and music discrimination, which is classifying between speech and music on audio signals. Audio signals are classified into 3classes (speech, music, speech and music) and 2classes (speech, music). Experiments carried out on three types of feature, Mel-cepstrum, energy, zero-crossings, and try to find a best combination between features to speech and music discrimination. We using a Gaussian Mixture Model (GMM) for discrimination algorithm and combine different features into a single vector prior to modeling the data with a GMM. In 3classes, the best result is achieved using Mel-cepstrum, energy and zero-crossings in a single feature vector (speech: 95.1％, music: 61.9％, speech & music: 55.5％). In 2classes, the best result is achieved using Mel-cepstrum, energy and Mel-cepstrum, energy, zero-crossings in a single feature vector (speech: 98.9％, music: 100％).

• Korean Journal of Remote Sensing
• /
• v.21 no.2
• /
• pp.145-162
• /
• 2005

To improve the accuracy of land-cover discrimination in SAB data classification, this paper presents a methodology that includes feature extraction and fusion steps with multi-temporal SAR data. Three features including average backscattering coefficient, temporal variability and coherence are extracted from multi-temporal SAR data by considering the temporal behaviors of backscattering characteristics of SAR sensors. Dempster-Shafer theory of evidence(D-S theory) and fuzzy logic are applied to effectively integrate those features. Especially, a feature-driven heuristic approach to mass function assignment in D-S theory is applied and various fuzzy combination operators are tested in fuzzy logic fusion. As experimental results on a multi-temporal Radarsat-1 data set, the features considered in this paper could provide complementary information and thus effectively discriminated water, paddy and urban areas. However, it was difficult to discriminate forest and dry fields. From an information fusion methodological point of view, the D-S theory and fuzzy combination operators except the fuzzy Max and Algebraic Sum operators showed similar land-cover accuracy statistics.

• Journal of Broadcast Engineering
• /
• v.17 no.3
• /
• pp.447-458
• /
• 2012

We present a method to recognize hand gestures using skeletal joint data obtained from Microsoft's Kinect sensor. We propose a combination feature of multi-angle histograms robust to orientation variations to represent the observation sequence of skeletons. The proposed feature efficiently represents the orientation variations of gestures that can be occurred according to person or environment by combining the multiple angle histograms with various angular-quantization levels. The gesture represented as combination of multi-angle histograms and random decision forest classifier improve the recognition performance. We conduct the experiments in hand gesture dataset obtained from a kinect sensor and show that our method outperforms the other methods by comparing the recognition performance.

• Korean Journal of Cognitive Science
• /
• v.7 no.1
• /
• pp.37-56
• /
• 1996

In this paper,we design a verifier for unconstrained handwritten numerals using structural feature analysis,and use it as a comnination algorithm for multiple recognizers.The existing combination algorithms mainly use learnings,statistical methods,or probabilistic methods without considering structural features of numerals.That is why they cannot recognize some numerals which human can identify clearly.To overcome the shortcomings,we design one-to-one verifiers which compare and analyze the relative structural features between frequently confused numeral pairs,and apply them to combine multiple recongnizers.Structural features for verification consist of contour,direction al chain code,polygonal approximation,and zero crossing number of horizontal/vertical projections. We gained a 97.95% reliability with CENPARMI numeral data,and showed that some misconceived factors generated from typical combination algorithms can be removed.

• The Journal of the Institute of Internet, Broadcasting and Communication
• /
• v.18 no.1
• /
• pp.211-216
• /
• 2018

In this paper, we studied the way that classify whether unknown PE file is malware or not. In the classification problem of malware detection domain, feature extraction and classifier are important. For that purpose, we studied what the feature is good for classifier and the which classifier is good for the selected feature. So, we try to find the good combination of feature and classifier for detecting malware. For it, we did experiments at two step. In step one, we compared the accuracy of features using Opcode only, Win. API only, the one with both. We founded that the feature, Opcode and Win. API, is better than others. In step two, we compared AUC value of classifiers, Bernoulli Naïve Bayes, K-nearest neighbor, Support Vector Machine and Decision Tree. We founded that Decision Tree is better than others.