얼굴 인식을 위한 전처리 기법 기반의 퍼지 신경회로망 분류기 설계 : Fuzzy transform 방법을 중심으로

김선환 수원대학교 2017 국내석사

In this thesis, RBF Neural Networks combined with Fuzzy Transform for face recognition are proposed. The extended Radial Basis Function Neural Networks based on fuzzy clustering are used to identify face images. Two-dimensional images based on Fuzzy Transform are partitioned into subsets by using fuzzy membership functions and represented by features that correspond to each membership function. From the preprocessing procedure of Fuzzy Transform, the dimensional reduction of images is carried out like the conventional dimension reduction techniques such as PCA and (2D)2PCA. Particle swarm optimization(PSO) is used to tune the parameters related to interval range and overlapped range between membership functions. Membership functions are partitioned uniformly or nonuniformly in the fuzzy space. Through Fuzzy Transform optimized with the aid of PSO, it is possible to reconstruct the image with nonuniform function that can describe more accurately compared to that with uniform function. The size of F matrix used in Fuzzy Transform is fixed for image dataset before running PSO. After the preprocessing of Fuzzy Transform, the face recognition is carried out by using FCM-based RBF Neural Network. The proposed RBF Neural Networks classifier is more compact than original RBF Neural Networks classifier from the structural viewpoint and consists of three modules such as condition, conclusion, and inference phase functionally. Condition and conclusion phases are related to the formation as well as analysis of fuzzy rules. Inference phase is concerned with a fuzzy inference. Fuzzy clustering-based radial basis function is considered as activation function in the hidden layer of Neural Networks and used to form information granules. In the conclusion phase, connection weights are used as diverse types of polynomials such as linear, quadratic, etc. In this study, the entire experiments of face recognition are carried out by using four databases such as Yale, ORL, Aberdeen, and GT DB. Two main experiments for each database are performed as follows : a) In the first experiment of preprocessing techniques, performance comparison and analysis between Fuzzy Transform and PCA / (2D)2PCA. b) In the second experiment of classifier techniques, Performance comparison and analysis between the proposed RBF Neural Networks and other methods realized with the aid of Weka tool. Here, Weka tool stands for a collection of machine learning algorithms for data mining tasks. As the result of the first experiment, the performance of the proposed method using F-Transform is demonstrated from the viewpoint of feasibility and applicability. In the second experimental result, the proposed RBF Neural Networks classifier shows more preferred performance compared to Weka's machine learning algorithms.

소량의 고차원 데이터 분류를 위한 생성 네트워크의 도움으로 구축된 패턴 분류기 설계 및 분석

정병국 수원대학교 2023 국내석사

Recently, a lot of studies have been carried out to deal with various problems occurring in prediction or classification area using artificial intelligence networks. Data information processing is very significant to effectively cope with the design and realization of artificial intelligence networks architectures that handle small amounts of high-dimensional data sets. Among these problems, One is the problem of being sensitive to noise and outliers due to the small amount of data for learning, and the other is the overfitting problem in which the parameters are fitted on learning data. In the previous studies, Generally, data information processing is done through dimensionality reduction algorithms such as principal component analysis(PCA) or linear discriminant analysis(LDA). In this study, we propose a design methodology related to classifiers through data generated by generative networks to cope with small amounts of high-dimensional data classification problems. The proposed methodology is to design convolutional variational autoencoder (CVAE) and deep convolutional generative adversarial networks (DCGAN) structures to newly generate 1-dimensional and 2-dimensional fake data sets from real data sets. And then fuzzy clustering methods-based radial base function neural network (FCM-based RBFNN) as well as convolutional neural networks (CNN) structures are exploited as classifiers for the processing of small amounts of high-dimensionality data. 1-dimensional and 2-dimensional fake data sets are generated with the aid of individual generative networks as follows: The fake data sets generated by CVAE as well as DCGAN showed similar distributions to actual data in some classes from the viewpoint of data characteristics and patterns. The experimental results for comparison between the performance of the proposed pattern classifiers and the existing pattern classifiers are as follows: The entire classification performance of FCM-based RBFNN pattern classifiers is decreased when using generated fake data sets but, the performance is partially enhanced in some classes. The classification performance of CNN pattern classifiers is decreased partially in some classes but, has been significantly improved from the point of overall classification performance. The total experimental results demonstrate that the generative network could partially enhance pattern classifier learning by generating fake data sets containing the characteristic patterns of real data sets.

최적화된 RBFNNs 기반 얼굴 인식 시스템 설계 및 구현 : 하이브리드 2차원 및 3차원 전처리 기법을 중심으로

장병희 수원대학교 2014 국내석사

본 연구에서는 주야간 및 3차원 얼굴인식을 위해 최적화된 RBFNNs 기반 다차원 얼굴 인식 시스템을 구현한다. 기존 얼굴 인식 시스템의 성능은 외부환경의 변화, 이미지의 크기, 그리고 얼굴의 포즈 등으로 인해 결정된다. 이러한 단점을 보완하기 위한 얼굴 영역 및 특징 추출 기법을 사용한다. 제안된 하이브리드 전처리 기법은 2차원 영상에서 Harr-like feature 및 Ada-Boost 기법을 이용하여 얼굴과 배경을 구분한 후 Active Shape Model(ASM) 알고리즘으로 얼굴 영역만을 찾아낸다. 전처리 과정에서는 CCD 카메라와 나이트비전 카메라로부터 2차원 영상을 입력 받고 Haar-like feature 와 Ada-boost 알고리즘을 통해 전체 이미지에서 얼굴을 분리하여 얼굴영역만을 획득한다. 그리고 ASM 알고리즘을 이용하여 얼굴 윤곽선 및 형상을 추출한다. 그 다음 특징 추출기법인 주성분분석법(PCA), 선형파별분석법(LDA), PCA-LDA 그리고 2-Directional and 2-Dimensional PCA((2D)2 PCA) 알고리즘을 통해 고차원의 이미지를 저차원 이미지로 축소한다. 또한 2차원 주야간 얼굴인식의 한계를 극복하기 위하여 3차원 스캐너를 이용한 인식 시스템을 제안한다. 3차원 얼굴인식 시스템은 얼굴형상 획득, 포즈보상, 특징 추출 및 인식 단계로 구성된다. 3차원 얼굴인식의 전처리 과정은 회전된 얼굴의 포즈를 정면으로의 보상하며 그리고 다중 영역 포인트 특징 기법(Multi area-point signature technique)을 사용한 얼굴의 깊이 정보 추출을 통해 수행한다. 인식단계에서는 얼굴 인식을 수행하기 위해 다항식 기반 RBFNNs 분류기를 제안한다. 제안된 다항식 기반 RBFNNs의 네트워크 구조는 조건부, 결론부, 추론부 세 가지의 기능적 모듈로 나누어진다. 조건부의 입력 공간은 기존의 가우시안 함수 대신 Fuzzy C-means(FCM) 클러스터링에 의하여 분할되며, 결론부의 연결가중치는 상수항을 확장한 다항식 함수로 표현된다. 다항식의 연결가중치 계수는 가중 최소 자승법(Weighted Least Square Estimation: WLSE)에 의하여 추정된다. 추론부의 최종출력은 퍼지 추론식을 통하여 얻는다. 또한 다항식의 형태, 규칙의 개수, 퍼지화 계수, 데이터의 차원수를 최적화를 위해 차분진화 알고리즘(Differential Evolution: DE)과 입자 군집 최적화(Particle Swarm Optimization: PSO)를 사용한다. 본 연구의 주야간 및 3차원 얼굴인식 시스템의 성능을 평가를 위해 주야간 IC&CI Lab 데이터를 사용한다. In this study, the proposed face recognition system is constructed with the aid of the design methodology of optimized RBFNNs-based 3-dimensional as well as 2-dimensional day & night face recognition. The performance of face recognition systems reported in literatures is affected by the chance of external environment, image scale, and face pose. Face area extraction as well as feature extraction techniques are used in order to compensate for these drawbacks mentioned above. The proposed hybrid pre-processing technique detect only face area by using ASM algorithm after distinguishing between face and background through Harr-like feature, and Ada-Boost algorithm on 2-dimensional image. In the pre-processing stage, CCD camera and night vision camera are used to obtain 2-dimensional image and then face area is merely selected by separating face image from entire image through Haar-like feature and Ada-boost Algorithm. Then, facial contour and shape are extracted by using ASM algorithm. High-dimensional images are reduced to low-dimensional images by using diverse preprocessing algorithms such as PCA, LDA, PCA-LDA, and (2D)2 PCA known as feature extraction techniques. Moreover 3-dimensional recognition system is developed by using 3D scanner to overcome the limitation of 2-dimensional day/night face recognition. 3-dimensional face recognition system is realized by some stages such as face shape acquisition, pose compensation, feature extraction, and recognition. In the pre-processing stage of 3-dimensional face recognition system, the procedure is carried out by using both compensation for rotated face pose and the depth information of face through multi area-point signature technique. In the recognition stage, polynomial-based RBFNNs classifiers is used in order to perform face recognition. The structure of the proposed polynomial-based RBFNNs is divided into three modules such as condition, conclusion, and inference phase. The input spaces of the condition phase are divided by Fuzzy C-means(FCM) instead of Gaussian function, and the connection weights of the conclusion phase are represented as polynomial function extended from constant terms. The coefficients of connection weights are estimated by Weighted Least Square Estimation(LSE) method. The final output of inference phase is obtained through fuzzy inference equation. The essential parameters of proposed classifier such as the order of polynomial, the number of rules, fuzzification coefficient, and the number of dimensions are optimized by means of Differential Evolution(DE) and Particle Swarm Optimization(PSO) respectively. We take into consideration the day & night dataset of IC&CI Lab in order to evaluate the output performance of 2-dimensional day & night as well as 3-dimensional face recognition system.

FT-IR 분광법을 이용한 플라스틱 재질에 대한 최적화된 뉴로-퍼지 패턴 분류기 설계

송찬석 수원대학교 2016 국내석사

In this thesis, the new method is proposed for classifying plastic materials based on Neuro-Fuzzy algorithm. In the automatic separation process of plastics, some types of plastics such as PP, PS and PET are classified by using near infrared spectroscopy. However, it is difficult to separate black plastics due to its properties of absorbing Near Infrared Rays(NIR). In this study, in order to overcome the drawback of the existing method to classify black plastics, FT-IR with ATR is used. Transmittance spectrum data is obtained through Fourier transform infrared spectroscopy. The obtained transmittance spectrum data is preprocessed by the proposed two methods and then the preprocessed data is used as inputs to Neuro-fuzzy networks classifier. In the preprocessing step for classifying black plastics for some materials, the characteristic of materials is analysed. In this study, two preprocessing methods are considered. The first method is to extract characteristic peak points from spectrum data. The second one is to extract region based on each characteristic point. The data extracted by the preprocessing methods is used as the input values of RBFNN pattern classifier. RBFNNs consist of three modules such as condition, consequence and aggregation phase. FCM clustering is used as the activation function of hidden layer in the condition phase. In the consequence phase, the coefficients of polynomial function are estimated by using least square estimation. In the aggregation phase, the final output of RBFNN classifier is calculated by using fuzzy inference. The optimal parameters of RBFNN classifier are obtained by using differential evolution. The differential evolution is well known as a kind of various genetic algorithms. Finally, the experimental results of the preprocessing methods are compared and analyzed from the viewpoint of classification performance.

다층 자기구성 네트워크 구조 안정화 를위한 Type-2 퍼지 C 평균 기반 다항식 신경망의 설계

Wang Zheng 수원대학교 2021 국내박사

In this thesis, a design methodology based on type-2 fuzzy c means-based polynomial neural networks for stabilized self-organizing networks structure is introduced to cope with over-fitting as well as multi-collinearity problems which generally appear in conventional fuzzy neural networks. The design method of the proposed self-organizing networks structure provides an efficient solution to construct the type-2 fuzzy c means-based polynomial neural networks(T2FPNNs) through a synergy of multiple techniques such as L2-norm regularization, probability theory, and multi-optimization, in order to generate the structure of the multi-layered self-organizing networks designed with the aid of the learning as well as novel structural design. Overall networks structure is realized with the aid of parallel networks structure with newly added inputs as well as effective neuron selection method through the exponential-based roulette selection technique for each layer in T2FPNNs, and the least square error estimation (LSE)-based learning method with L2-norm regularization is used for constructing the stabilized networks architecture, and their ensuring design methodologies result in alleviating the overfitting phenomenon and also enhancing the generalization ability. For the performance enhancement of T2FPNNs directly affected by some parameters such as the number of input variables, fuzzification coefficient, the number of clusters per each variable, and the order of polynomial in the consequent parts of the fuzzy rules, multi-particle swarm optimization (MPSO) is exploited for the effectively structural as well as parametric optimization of the proposed networks. That is, the multi-optimization helps achieve compromise between the better generation performance and the alleviated over-fitting in order to lead to the stabilization of the proposed multi-layered self-organizing networks structure realized with the aid of synergistic multi-techniques such as a) L2-norm regularization-based LSE learning, b) probability theory for effective neuron selection through the exponential-based roulette selection technique, and c) novel parallel networks structure including newly added inputs and neuron selection method. The performance of the proposed networks structure is quantified by comprehensive experiments and comparative analysis.

부분방전 패턴 분류기를 위한 퍼지 신경회로망의 설계 및 학습방법에 관한 연구

정병진 수원대학교 대학원 2018 국내석사

In this thesis, Fuzzy Neural Networks pattern classifier structure and learning method were proposed. The purpose of this study was focused to redesign the structure of the neural networks and to develop a design methodology to improve the learning method of the neural network based on K-means Clustering. The key features of the proposed Fuzzy Neural Networks pattern classifier were listed as follows. 1) To acquire the partial discharge data, the original data was obtained using the Epoxy Mica Coupling(EMC) sensor equipped with the Phase Resolved Partial Discharge Analysis(PRPDA) method and processed through the Motor Insulation Monitoring System(MIMS) program. The processed data reduced the input variable of the high dimensional input to low dimension through the Independent Component Analysis preprocessing algorithm. And the processing speed became faster. At this time, the data was processed with characteristics of maximum value, average value, and median. 2) The cost function used the cross entropy error function instead of the sum of error squares and applied the softmax function to the number of nodes in the output layer to normalize the output value to a probability value between 0 and 1. 3) In order to adjust the connection weights of the hidden layer, K-means Clustering and Gaussian function were used. The center values generated by the K-means Clustering was applied to the Gaussian function and thus the membership values were generated. The connection weights of the output layer were controlled by the nonlinear least squares method using the Newton’s method. 4) L2 regularization is considered to prevent overfitting and improve generalization capability. By adding the L2 penalty term to the cross entropy error function, the performance of the proposed Fuzzy neural networks pattern classifier is superior to that of the previous classifier reported in the literature.

Fine-grained Object Detection Deep Learning Models Designed for Medical Imaging Recognition

Huang, Hao 수원대학교 2024 국내박사

최적 pRBF 신경회로망 기반 기상 시스템 설계 : 강수 예측 및 레이더 자료 해석 기법을 중심으로

김현명 수원대학교 2014 국내석사

In this study, meteorological system is realized with the aid of optimized RBFNNs focused on precipitation forecast & radar data analysis techniques. The meteorological system is identified with precipitation forecasting system and radar echo classifier system. Precipitation forecasting system is focused on the development of the very short-term precipitation forecasting model as well as classifier based on polynomial radial basis function neural networks by using AWS(Automatic Weather Station) and KLAPS(Korea Local Analysis and Prediction System) meteorological data. The prediction ability of the existing precipitation forecasting systems is usually affected by the diverse processing techniques related to meteorological data. In order to improve these drawbacks of conventional system, the precipitation forecast methods are developed with the aid of some types of preprocessing techniques concerning meteorological data. By using the cumulative precipitation amount accumulated for six hours, the proposed system forecasts in advance before one or two hours and offers the related information to issue a heavy rain warning. Secondly the radar echo classifier system is developed based on preprocessing techniques and fuzzy-neural networks by using weather radar data being exploited for weather forecasting. The design procedure of radar echo classifier system is given as follows. First, the characteristic analysis of precipitation/non-precipitation echo as well as structural analysis of complicated radar data is carried out for detailed design. And then input variables for the proposed classifier are determined through the extraction of radar data as well as the operation of membership variable. RBFNNs classifier combined with logical echo judgement modules are designed to obtain superb performance as precipitation/non-precipitation echo classifier. The polynomial-based radial basis function neural networks(RBFNNs) is designed to realize meteorological system which consists of precipitation prediction and echo classification. The structure of the proposed RBFNNs consists of three modules such as condition, conclusion, and inference phase. The input space of the condition phase is divided by using Fuzzy C-means(FCM) and the conclusion phase is represented as four types of polynomial functions. The coefficients of connection weights are estimated by weighted least square estimation(WLSE) for modeling as well as least square estimation(LSE) for classifier. The final output of the inference phase is obtained through fuzzy inference method. The essential parameters of the proposed model such as the number of input variable, polynomial order, the number of rules, and fuzzification coefficient are optimized by means of Differential Evolution(DE) as well as Particle Swarm Optimization(PSO). The performance of the proposed precipitation forecasting system is evaluated by using KLAPS meteorological data, and also the performance of the proposed echo classifier system is evaluated and quantified by using Oseongsan Radar data. 본 연구에서는 강수 예측 및 레이더 자료 해석 기법을 중심으로 최적화된 RBFNNs 기반 기상 시스템을 구현한다. 기상 시스템은 강수 예측 시스템 및 레이더 에코 분류 시스템으로 구분된다. 강수 예측 시스템은 AWS(Automatic Weather Station)와 KLAPS(Korea Local Analysis and Prediction System)의 기상데이터를 이용하여 다항식 기저함수 신경회로망 기반 초단기 호우 예측 모델 및 분류기의 개발에 중점을 둔다. 기존의 호우예측 시스템들의 예측능력은 일반적으로 다양한 기상데이터의 가공 기법의 영향을 받는다. 기존 시스템의 결점을 개선하기 위하여, 기상 예측 방법은 기상데이터에 관한 몇 가지 전처리 기법의 도움으로 개발된다. 제안된 시스템은, 향후 t(t=1,2) 시간 후 6시간 동안 누적강수량에 대해 예측하고 호우특보를 발령하기 위한 정보를 제공한다. 두 번째로 강수 예측을 위해 이용되는 기상 레이더 데이터를 사용하여 전처리 기법 및 퍼지-신경회로망 기반의 레이더 에코 분류 시스템이 개발된다. 레이더 에코 분류 시스템의 설계 절차는 다음과 같이 주어진다. 우선 복잡한 레이더 자료의 구조 해석과 더불어 강수/비강수에코의 특성 분석 수행에 의해 상세히 설계된다. 그리고 레이더 자료의 추출 및 소속변수의 연산을 통해 제안된 분류기의 입력변수가 결정된다. 이 후 RBFNNs 분류기에 논리적 에코 판단기준 모듈을 결합하여 보다 뛰어난 성능을 얻을 수 있도록 강수/비강수 에코 분류기가 설계된다. 강수 예측과 에코 분류로 구성된 기상 시스템을 구축하기 위하여 다항식 기반 방사형 기저함수 신경회로망(Polynomial-based Radial Basis Function Neural Networks; RBFNNs)을 설계한다. 제안된 RBFNNs의 네트워크 구조는 조건부, 결론부, 추론부 세 가지의 기능적 모듈로 나누어진다. 조건부의 입력공간은 Fuzzy C-means(FCM) 클러스터링에 의하여 분할되며, 결론부는 다항식 함수로 표현된다. 다항식의 연결가중치 계수는 최소자승법(Least Square Estimation; LSE)과 가중최소자승법(Weighted Least Square Estimation; WLSE)에 의하여 추정된다. 추론부의 최종출력은 퍼지 추론식을 통하여 얻는다. 또한 입력변수의 수, 다항식의 형태, 규칙의 개수, 퍼지화 계수와 같은 제안된 분류기의 중요 파라미터는 최적화 기법인 차분진화 알고리즘(Differential Evolution; DE)과 입자 군집 최적화(Particle Swarm Optimization; PSO)를 이용하여 최적화한다. 제안된 강수예측 시스템의 성능평가를 위해 KLAPS 기상데이터를 사용하였고, 제안된 에코 분류 시스템의 성능은 오성산 레이더 데이터를 사용하여 평가된고 정량화 된다.

뉴로퍼지 네트워크 기반 부분방전 열화진단모델 및 패턴분류기 설계

박제현 수원대학교 대학원 2020 국내석사

In this study, the algorithm for deterioration diagnosis model design and partial discharge pattern classification was applied. In the first experiment, fuzzy rules are made for the degradation model to create a health index and noise is added to the RBFNN model. Through this, we would like to confirm the possibility of application in the field. We compared the performance of various machine learning algorithms with Weka Software. The results show that RBFNN is more robust to noise and is likely to be applied in the field. In the second experiment, the partial discharge data obtained through the UHF sensor is preprocessed by PCA (Principal Component Analysis) using the Phase Resolved Partial Discharge Analysis (PRPDA) method. In addition, five statistical parameters (Skewness, Kurtosis, Standard Derivation, Variance, Average) were used to analyze the waveform of partial discharge. As a result, it is confirmed that the performance is better when applied than when not applied, which may help to understand the characteristics of the waveform.

레이저 분광 기술을 이용하여 플라스틱을 재질에 따라 분류하는 방법에 관한 연구

유병건 수원대학교 2021 국내석사

Among the plastics that make our lives convenient, the amount of waste plastics used and disposed of is on a continuous increase. Recognition and sorting by material is very important for recycling of waste plastics that are discarded after use. However, ABS, PS, and PP, which are representative plastics generated from used household appliances, and Eng. PP, PC, and PC/PS, which are representative plastics from waste automobiles, are mostly black. This plastic cannot be recognized and sorted by material using the conventional plastic material sorting technology. In the case of conventional plastic material recognition and sorting technologies, Laser Induced Breakdown Spectroscopy(LIBS) is proposed as an alternative to supplementing the limitations of the recognition technology as it is not possible to identify the materials of black plastic due to the limitations of the plastic-specific recognition technology. In this study, characteristic data on the material of ABS, PS, and PP generated from used household appliances and Eng. PP, PC and PC/PS generated from waste automobile crushed material were obtained using LIBS, and acquired data were treated as artificial intelligence pretreatment algorithms. We designed a classifier(Radial Basis Function Neural Networks, RBFNN) applying various types of artificial intelligence preprocessing algorithms and evaluated the classification performance. In addition, an automatic sorting system for each material of waste plastics was established, the features of the data were extracted using a preprocessing algorithm for the plasma spectral signals and data obtained by each plastic material, and a classifier and Dynamic-Link Library(DLL) were designed and applied using the extracted data. The recognition rate and sorting tests for plastic materials were performed using the automatic sorting system. In the future, we plan to analyze various preprocessing algorithms not used in this study to find out the difference in feature extraction and to compare classification performance. In addition to the RBFNN classifier proposed in this study, the classification performance comparison with other artificial intelligence algorithm classifiers such as fuzzy set-based neural networks will be additionally performed. And we intend to acquire additional characteristic data on plastic materials, convert them into big data, and expand research on processing technology using them.