Earticle

이미지 검색은 주어진 이미지에 대해 검색 목적에 맞는 이미지를 찾는 문제이다. 최근 연구는 주로 시각적으로 유사 한 이미지를 찾는 content-based image retrieval(CBIR) 문제를 해결하기 위해 이미지 분류를 위해 대용량 데 이터셋에 대해 사전 학습된 딥러닝 모델의 특징맵(feature map)을 활용한다. 본 논문은 주어진 이미지와 같은 객 체를 포함하고 있는 이미지를 찾는 object-based image retrieval(OBIR) 문제를 제안한다. 기존의 이미지 검색 기법은 global feature나 local feature를 사용해 유사한 이미지를 검색하는데 이때 주어진 이미지가 복잡한 배경 을 가지고 있거나 이미지의 객체가 작은 경우 객체 정보를 얻는 데 어려움이 발생한다. 이를 해결하기 위해 배경에 구애받지 않고 객체의 정보를 사용하는 최신 객체 인식 모델인 Detection Transformer(DETR)와 Bag of Visual Words(BoVW)를 활용한 객체 임베딩(object embedding) 및 이미지 검색 기법을 제안한다. 본 논문은 객체 인식 문제에 주로 쓰이는 COCO 데이터와 기존 이미지 검색 기법에서 사용하는 AID, Oxford & Paris 데이 터를 사용해 기존 이미지 검색 기법들과 비교 실험을 수행하였고, DETR 기반 BoVW가 OBIR 문제에서 기존 방법 대비 우수한 성능을 내는 것을 확인하였다.

Image retrieval is a problem of finding an image suitable for the purpose of search for a given image. Recent work utilizes a feature map of pre-trained deep learning models from large dataset for image classification to solve the content-based image retrival(CBIR) problem, mainly finding visually similar images. This paper proposes an object-based image retrival(OBIR) problem of finding an image containing an object same as a given image. Traditional image retrieval techniques use global or local features to search for similar images, which makes it difficult to obtain object information if a given image has a complex background or if the object of the image is small. To address this, we propose object embedding and image retrieval techniques using the latest object recognition models, DETR (Detection Transformer) and Bag of Visual Words (BoVW), which can represent object information regardless of background. This paper conduct a comparative experiment with existing image retrieval techniques using COCO data mainly used in object recognition problems and AID, Oxford & Paris data used in existing image retrieval techniques, and show that DETR-based BoVW outperforms existing methods in OBIR problems.

Support vector machines (SVMs), a machine learning technique, has been applied to not only binary classification problems such as bankruptcy prediction but also multi-class problems such as corporate credit ratings. However, in general, the performance of SVMs can be easily worse than the best alternative model to SVMs according to the selection of predictors, even though SVMs has the distinguishing feature of successfully classifying and predicting in a lot of dichotomous or multi-class problems. For overcoming the weakness of SVMs, this study has proposed an approach for selecting features for multi-class SVMs that utilize the impurity measures of classification trees. For the selection of the input features, we employed the C4.5 and CART algorithms, including the stepwise method of discriminant analysis, which is a well-known method for selecting features. We have built a multi-class SVMs model for credit rating using the above method and presented experimental results with data regarding S&P 500 companies.

플라스틱 사출 제품은 다양한 가전제품과 하이테크 제품에 널리 사용되고 있다. 그러나 현재의 치열한 경쟁적 비즈니스 환경에서 플라스틱 사출 제품 제조업자들은 고객을 만족시키면서 경쟁력을 얻기 위하여 다른 경쟁자들보다 먼저 새로운 제품을 시장에 출시하고 신제품의 개발기간을 줄이기 위한 노력을 할 여유가 부족하다. 따라서 무한 경쟁의 시장에서 살아남기 위해서는 제조업자들은 시장 마켓 점유를 빠르게 올리는 것과 동시에 제품의 가격 경쟁력을 가져야 한다. 특징기반 모델의 구조는 현재 연구에서 3D 제작 도구로서 일반적으로 적용되고 있으며 신제품 개발 엔지니어들이 새로운 제품의 개념을 개발하는 데에도 널리 사용되고 있다. 본 연구에서는 특징기반 플라스틱 사출제품을 위한 유전자 알고리즘과 Support Vector Regression (SVR) 기반의 새로운 하이브리드 비용 평가 모델을 제안한다. 제안하는 하이브리드 모델은 기존의 플라스틱 사출제품의 비용평가절차와 계산을 위해 필요로 하는 변수들을 극적으로 간단하게 하고 줄일 수 있다. 사례연구에서는 제안하는 하이브리드 모델과 기존의 multilayer perceptron networks (MLP) 및 pure SVR과의 비교분석을 통하여 제안모델이 플라스틱 사출 제품의 개발단계에서의 비용평가문제를 해결하는데 효율성과 효과성이 있음을 입증한다.

본 논문에서는 유도 전동기의 결함을 조기에 검출하고 진단하기 위해 이산 코사인 변환과 특이치 분해를 이용한 결함 특징 벡터 추출 방법을 제안하였고, 추출한 특징 벡터를 다층 서포트 벡터 머신의 입력으로 이용하여 유도 전동기의 결함 을 유형별로 분류하는 기법을 기술하였다. 모의실험 결과, 제안한 방법은 5가지의 결함 데이터에 대해 거의 100%의 분류 성능을 보였다.

This paper proposes a method of feature vector extraction using discrete cosine transform(DCT) and singular value decomposition(SVD) for early detection of faults in an induction motor. This, the extracted feature vector is used as an input of multi-layer support vector machine(MLSVM) to classify each fault of the induction motor. Experimental results show that the proposed method achieves 100% accuracy to classify among 5 different faults.

최근 인공지능 기술의 발달과 더불어 감정인식은 중요한 연구 분야로 대두되고 있다. 기존의 음성, 얼굴을 기반으로 한 감정연구에서 최근에는 보다 객관적인 뇌파 등과 같은 생체신호를 활용하는 연구로 확대되고 있다. 본 연구에서 는 뇌파 데이터를 활용하여 긍/부정에 대한 뇌파에 특징요소를 추출하고, 대표적인 기계학습 분류기인 SVM(Support Vector Machine) 분류기를 활용하여 감정분류를 진행하였다. 뇌파 신호에 대한 전처리를 통하여 몸과 눈의 움직임(eye-blick)과 같은 방해파를 제거하였다. 특징요소 추출을 위하여 전처리된 뇌파 데이터를 10ms 당 데이터의 평균값을 생성하고, 이를 50%로 중첩하는 방법으로 1ms 단위의 이미지를 형성하였다. 추출된 특징요 소는 대표적인 기계학습 분류기인 SVM을 사용하여 감정분류를 진행하였다. 그 결과, 3가지 감정에 대하여 평균 94.6%의 결과를 도출하였다. 본 연구의 결과는 감정분류에 있어서 시계열 데이터를 이미지화하는 방법으로 특징을 추출하였다. 향후 시계열 데이터 처리에 있어서 새로운 특징요소 추출방법으로 활용 가능하며, 또한, 감정을 세분화 하여 다양한 기계학습 알고리즘에 적용할 수 있는 새로운 접근법을 제안한 것에 의의를 둔다.

According to the development of artificial intelligence, emotion recognition has emerged as an important element from psychology to engineering. In the previous research, there are various of materials such as voice, motion, and behaviors. However, these responses have a weakness to identify real emotional states because of making the faking voice or hiding facial expression unlike their emotion as intended by oneself. Some studies related to emotion recognition have extended to study emotion recognition using physiological signals such as electroencephalogram (EEG), electrocardiogram (ECG), skin conductance (SKT) and respiration. In this study, we performed the emotion recognition using SVM (Support Vector Machine) with public open database. The EEG signals conducted the preprocessing for removal to artifacts such as muscle and eye blicks. In the feature extraction, the window size is 10 seconds and 50% overlap and the output image data is produced per 1ms. As a result, we achieved the classification average accuracy of 94.6% among three emotion, positive, negative and calm. In conclusion, this proposed method is a novel approach for deal with time-series data such as voice and physiological signals. Also, we proposed that feature extraction relates to emotion recognition for machine learning.

본 논문에서는 풍력 예측의 정확도를 높이기 위해 풍향 특성을 고려한 풍력 예측 방법을 제안한다. 제안 된 방법은풍력 특성을 추출하고 발전을 예측하는 것을 포함한다. 바람의 특성을 추출하기 위해 발전량, 풍향 및 풍속의 상관분석이 수행된다. 풍향과 풍속의 상관 관계를 바탕으로 K-means 법을 이용한 클러스터링을 통해 특징 벡터를 추출한다. 예측 부분에서는 임의의 실수 값을 예측할 수 있도록 SVM (Support Vector Machine)을 일반화 한 SVR (Support Vector Regression)을 사용하여 기계 학습을 수행한다. 제안 된 방법의 정확성과 타당성을 검증하기위해 제주도 풍력 발전 단지의 3 곳에서 수집 한 자료를 이용하였다. 실험 결과는 제안 된 방법의 오차가 기존 풍력발전 방법의 오차보다 우수하다는 것을 보여준다.

In this paper, we propose a wind power forecasting method that takes into consideration wind characteristics to improve the accuracy of wind power prediction. The proposed method involves extracting wind characteristics and predicting power generation. Correlation analysis of power generation amount, wind direction, and wind speed is performed for extracting wind characteristics. Based on the correlation between the wind direction and the wind speed, the feature vector is extracted by clustering using the K-means method. In the prediction part, machine learning is performed using the SVR (support vector regression) that generalizes the SVM (support vector machine) so that an arbitrary real value can be predicted. To verify the accuracy and feasibility of the proposed method, we used the data collected from three different locations of the Jeju Island wind farm. Experimental results show that the error of the proposed method is better than that of existing wind power generation methods.

In order to restrain the bad influences of fake and inferior products, an automatic identification authentic work anti-counterfeiting algorithm is proposed in this paper. The authentic work anti-counterfeiting uses a unique random character structure to carry out security. The algorithm is convenient for customers to distinguish the anti-counterfeiting tag by both naked eye and mobile. There is not yet a similar algorithm published. In the algorithm, the feature vectors are extracted using DWT-DCT. Using extracted feature vectors from anti-counterfeiting tag to realize automatic identification. Therefore, this algorithm saves the storage space and improves the identification rate. In addition, the experimental results show this algorithm have strong robustness under common attacks, geometrical attacks.

This paper proposes a new feature vector selection method for voice pattern recognition tasks, especially for speaker or emotion recognition. During the model training phase, robust speaker or emotion models are constructed by using meaningful feature vectors while discarding confusing vectors that may induce recognition error. To select meaningful feature vectors, the proposed method classifies feature vectors into overlapped and non-overlapped sets using log-likelihood ratio. Speaker- and emotion-recognition experiments confirmed that these robust models significantly reduce recognition errors.

In the area of crowd abnormal detection, the parameter of population density, is seldom used to the global crowd behavior detection. Some of the references simply use the LBP or spatial-temporal LBP features to fulfill the abnormal detection. They don’t make full use of the crowd density characteristics and dynamic characteristics. This paper proposes a novel method by increasing the dimension of feature vector to increase the information content so as to improve the recognition accuracy. That is to say the crowd dynamic information and crowd density information will be combined together to form a higher dimension of feature vectors, which is named as the crowd behavior feature vector in this paper to improve the robustness of the algorithm. Finally, Support Vector Machines (SVM) is adopted to detect the abnormal events using the crowd behavior feature vector. This work utilizes the Local Binary Pattern Co-Occurrence Matrix (LBPCM) for crowd density estimation to ensure the excellent accuracy. At the same time, it adopts high accuracy optical flow histograms of the orientation of interaction force to extract the crowd dynamic information (HOIF). After verification, we discovered this algorithm not only can get the good discrimination on the benchmark dataset UMN, but also can achieve the pretty high recognition rate about the web dataset.

With the very fast development in today`s digital world, Information Retrieval on Internet is gaining importance, day by day. The web comprises of huge amount of data and search engine provides an efficient way to navigate the web and get the relevant information. The search engine has proven to be less efficient in providing relevant information from a query processed by a user. Fors olving this problem and getting accurate results there is need to categorize these web pages. Many optimizations have also done to speedup the classification process as it is required to be fast while maintaining the efficiency. To maintain the accuracy with the lesser time requirement, researchers have developed a SVM based Layered approach with the help of firefly feature selection method.

It is evident from various researches that disease diagnosis using machine learning methods has been increasing rapidly. In this research work, feature selection based Least Square Twin Support Vector Machine (LSTSVM), which is a machine learning method, is used for diagnosis of heart diseases. In this approach F-score is used to calculate the weight of each feature and then features are selected according to their weight. The higher weight is assigned to the feature having high F-score. Grid search approach is also utilized to select the best value of classifier's parameters in order to enhance its performance. The heart-statlog disease dataset is used in this study, which is taken from the UCI repository. The performance of proposed model with different feature sets has been evaluated for different training-test datasets. The results indicate that LSTSVM model with 11 features has achieved highest accuracy. The results are very promising as compared to the other approaches proposed earlier.

This paper describes a process of developing commercial real time image recognition system with company. In this paper we will make a system that is combining an empirical kernel map method and conjugate least squares support vector machine in order to represent images in a low-dimensional subspace for real time image recognition. In the traditional approach calculating these eigenspace models, known as traditional PCA method, model must capture all the images needed to build the internal representation. Updating of the existing eigenspace is only possible when all the images must be kept in order to update the eigenspace, requiring a lot of storage capability. Proposed method allows discarding the acquired images immediately after the update. By experimental results we can show that empirical kernel map has similar accuracy compare to traditional batch way eigenspace method and more efficient in memory requirement than traditional one. This experimental result shows that proposed model is suitable for commercial real time image recognition system.

A feature parameter modification algorithm is proposed to increase the accuracy of voice activity detection that based on support vector machine and energy acceleration parameters. The three energy acceleration parameters should have the equal importance for voice activity detection, in that all the three parameters can suitably characterize the classification features of speech and non-speech frames. For the three energy acceleration parameters, its minimum values vary a little; but its maximum values are greatly different. When radial basis function is chosen as the kernel function for voice activity detection, the coordinate values in Hilbert space is dominative determined by the energy values over a sub-region spectrum, the other two parameters only have a few contributions to it. The proposed algorithm extends the three energy acceleration parameters into the nearby or same order of magnitude. It make the three energy acceleration parameters at the same importance level in the calculation of coordinate values in Hilbert space, and can finally increase the accuracy of voice activity detection. The experimental results show that this algorithm can increase the accuracy of voice activity detection in the absence of noise and noise conditions.

Muilti-dimensional feature vector matching algorithm uses multiple features such as intensity, gradient, variance, first or second derivative of a pixel to find correspondence pixels in stereo images. In this paper, we proposed a new method for adjusting automatically the weight of feature in multi-dimensional feature vector matching considering sharpeness of a pixel in feature vector distance curve. The sharpeness consists of minimum and maximum vector distances of a small window mask. In the experiment we used IKONOS satellite stereo imagery and obtained accurate matching results comparable to the manual weight-adjusting method.

최근 컴퓨터 기술이 발전하고, 컴퓨터의 형태가 다양해지면서 여러 wearable device들이 생겨났다. 이에 따라 휴먼 인터페이스 기술에서 사람의 감정정보가 중요해졌고, 감정인식에 대한 연구들이 많이 진행 되어 왔다. 본 논문에서는 감정분석에 적합한 특징벡터를 제시하고자 한다. 이를 위해 사람의 감정을 보통, 기쁨, 슬픔, 화남 4가지로 분류하고 방송매체를 통하여 잡음 없이 녹음하였다. 특징벡터는 MFCC, LPC, LPCC 3가지를 추출하였고 Bhattacharyya거리 측정을 통하여 분리도를 비교하였다.

The high-dimensionality and insufficiency of gene expression profiles and proteomic profiles makes feature selection become a critical step in efficiently building accurate models for cancer problems based on such data sets. In this paper, we use a method, called Discrete Function Learning algorithm, to find discriminatory feature vectors based on information theory. The target feature vectors contain all or most information (in terms of entropy) of the class attribute. Two data sets are selected to validate our approach, one leukemia subtype gene expression data set and one ovarian cancer proteomic data set. The experimental results show that the our method generalizes well when applied to these insufficient and high-dimensional data sets. Furthermore, the obtained classifiers are highly understandable and accurate.

In this paper, we propose a two-step noise compensation algorithm in feature extraction for achieving robust speech recognition. The proposed method frees us from requiring a priori information on noisy environments and is simple to implement. First, in frequency domain, the Harmonics-based Spectral Subtraction (HSS) is applied so that it reduces the additive background noise and makes the shape of harmonics in speech spectrum more pronounced. We then apply a judiciously weighted variance Feature Vector Normalization (FVN) to compensate for both the channel distortion and additive noise. The weighted variance FVN compensates for the variance mismatch in both the speech and the non-speech regions respectively. Representative performance evaluation using Aurora 2 database shows that the proposed method yields 27.18% relative improvement in accuracy under a multi-noise training task and 57.94% relative improvement under a clean training task.

In this paper, we propose an efficient method for personal identification by analyzing iris patterns that have a high level of stability and distinctiveness. To improve the efficiency and accuracy of the proposed system, we present a new approach to making a feature vector compact and efficient by using wavelet transform, and two straightforward but efficient mechanisms for a competitive learning method such as a weight vector initialization and the winner selection. With all of these novel mechanisms, the experimental results showed that the proposed system could be used for personal identification in an efficient and effective manner.

영아는 비언어적 의사 소통 방식인 울음이라는 수단을 사용한다[1]. 하지만 영아의 울음소리를 파악하는 것에는 어려움이 따른다. 영아의 울음소리를 해석하기 위해 많은 연구가 진행되었다[2,3]. 이에 본 논문에서는 다양한 음성 데이터 증강을 통한 3D 특징 벡터를 이용한 영아의 울음소리 분류를 제안한다. 연구에서는 총 5개의 클래스 복통, 하품, 불편함, 배고픔, 피곤함(belly pain, burping, discomfort, hungry, tired)로 분류된 데이터 세트를 사용한다. 데이터들은 5가지 기법(Pitch, Tempo, Shift, Mixup-noise, CutMix)을 사용하여 증강한다. 증강 기법 중에서 Tempo, Shift, CutMix 기법을 적용하였을 때 성능의 향상을 보여주었다. 최종적으로 우수한 데이터 증강 기법들을 동시 적용한 결과 단일 특징 벡터와 오리지널 데이터를 사용한 모델보다 17.75%의 성능 향상을 도출하였다.

Infants utilize crying as a non-verbal means of communication [1]. However, deciphering infant cries presents challenges. Extensive research has been conducted to interpret infant cry audios [2,3]. This paper proposes the classification of infant cries using 3D feature vectors augmented with various audio data techniques. A total of 5 classes (belly pain, burping, discomfort, hungry, tired) are employed in the study dataset. The data is augmented using 5 techniques (Pitch, Tempo, Shift, Mixup-noise, CutMix). Tempo, Shift, and CutMix augmentation techniques demonstrated improved performance. Ultimately, applying effective data augmentation techniques simultaneously resulted in a 17.75% performance enhancement compared to models using single feature vectors and original data.