A conventional prototype based unsupervised neural network such as LVQ and Kohonen network is a useful tool for clustering/classification problems with simple structure, simple learning method, and fast convergence. However, it has inherently three well-known problems ; 1) how to place weights (prototypes) into input pattern space, 2) how to find an optimal number of cluster and 3) how to define class discriminant boundaries. In this thesis, we consider the improvement of clustering and classification performance of the neural network to solve the above mentioned problems and propose three methods to overcome the criticisms. First, a new optimal clustering method based on a new probability function with an annealing optimization method is developed. Second, a classification performance index to select an optimal number of cluster for the classification is introduced based on two criteria such as interset distance between clusters and intraset distance between input patterns within a cluster. Finally, a new neuron model based on crystal growth phenomenon is to developed to describe complex class boundaries and an effective learning algorithm is developed.
The developed algorithm is a new learning method based on a new probabilistic distortion function with an annealing method. The so called distortion function is expressed by two terms; an intraset distance and an interset distance. The intraset distance term is expressed by the sum of the Euclidean distances between the weight vector of each class and input patterns belonging to its class. The interset term is expressed by the difference between the desired probability and calculated probability of each class. The learning algorithm is based on minimizing an energy function and an annealing method is also adopted to find optimal solution in the given energy function. Its learning algorithm is based on unsupervised manner and also adopts supervised manner by defining the desired probability supervidely.
Next, we introduce a new classification performance index to find an optimal number of cluster/class. The proposed index can be also expressed by two criteria; intraset distance can be measured by the average distance of input patterns of the class and interset distance can be expressed the dissimilarity between the prototypes. Considering the above two terms together, we evaluate the proposed index to find the optimal number of cluster.
Finally, we suggest a new neuron model to describe complex class boundaries. The proposed neuron model adopts the phenomenon of crystal growth. The prototype in each class is initialized as nuclear in crystal. It gradually grows until it meets its class boundary which are taught in supervised manner. At this moment, neuron is said to be fully grwon to describe precise class boundaries.
A series of numerical simulations and experiments are performed to cluster and classify solder joint images within PCBs. To do this a vision system is utilized to obtain images of the joints under 3-tiered ring illumination. Another application is illustrated for BGA cross-sectional solder joint. In this case, the images are obtained by X-ray tomosynthesis method. Both simulations and experimental results reveal that the proposed neural network show high accuracy and thus practical usefulness in real classification problems.
본 논문에서는 기존의 단층형 신경회로망의 한계점을 극복하고자 보다 복잡한 부분 구간별 선형 경계를 나타낼 수 있는 새로운 형태의 뉴런과 이를 이용한 학습 방법을 제안하고자 한다. 제안된 뉴런은 자연 현상에서의 석영의 결정 성장 모델을 사용하여 복잡한 군집 형태를 학습에 의해 표현하도록 하도록 구성하였다. 새로운 형태의 뉴런을 사용한 신경회로망 분류기는 먼저 확률(statistical)개념을 바탕으로 하여 신경회로망의 새로운 에너지 함수를 정의하고 이를 이용하여 담금질(annealing)방법을 사용하여 전역 해를 찾아 가는 군집화 학습 알고리즘을 개발하여 사용하였으며, 제안된 군집화 알고리즘은 초기 핵의 형태로 출발하는 뉴런을 패턴의 군집 중앙에 위치하기 위한 중요한 역할을 수행한다. 또한, 최적의 뉴런의 수를 구하기 위하여 뉴런간의 상호 배치에 바탕을 둔 새로운 성능 지수를 정의하였다. 마지막으로, 입력 패턴의 분포 형태를 근사화 하기 위하여 결정 성장 모델중의 하나인 level set이론을 응용하여 뉴런의 수학적 모델을 세우고 이에 적합한 학습 알고리즘을 제안 하였다.
제안된 모델을 사용하여 실제 여러 가지 산업용 패턴들에 대한 분류 실험 과정을 통하여 성능을 분석하였으며, 다른 신경회로망 제어기와 비교 연구를 통하여 제안된 모델 우수성을 검정하였다.