类神经网路与模糊理论在影像判释上之应用.pptVIP

* 報 告 人: Dr. Brandt Tso Image Classification Using Artificial Neural Networks And Fuzzy Theory: An Introduction 使用時間: 四十分鐘 報告大綱 類神經網路(ANN)之結構及應用 Multilayer Perceptron(or called Feed-Forward NN) Kohonen’s Self-Organized feature Map (SOM) Other neural network types 模糊理論(Fuzzy Theory) Why Fuzzy? Fuzzy Set Definition Fuzzy Set Operators Applications: Fuzzy Air-Conditioner Image Classification Using Fuzzy Rules Fuzzy c-means Clustering Mixture Classification Input Layer Hidden Layer Output Layer Weights ANN: Multilayer Perceptron Structure and Training 0.56 0.93 1 0 0 Input Expected Output 0.98 0.03 0.06 Real Error . . . ? a1 a2 a3 am wj1 wj2 wj3 wjm (ai wji xj = aj=f(xj) 1 ?j -?j ) bias Neurons and Weights 1 0 f(x) x T =1 T 1 T 1 1 f(x) 0 ? ?? (a) (b) To avoid trapped to local minimum Weight update Learning Rate Momentum ANN: Multilayer Perceptron How to train a Multilayer Perceptron Learning Rate h Momentum x (For weights update) Normally h : 0.1 or 0.2 x : 0.8 or 0.9 1 Output Coding In 4-class case use 4 output neurons, e.g. [1, 0, 0, 0],[0, 1, 0, 0] [0, 0, 1, 0],[0, 0, 0, 1] rather than just 2 neurons [0, 0], [0, 1] [1, 0], [1, 1] 0 0 0 1 0 ANN: Multilayer Perceptron Training Issue How to train a Multilayer Perceptron Error Iterations Training Error Test Error Stop Training Point ANN: Multilayer Perceptron How big the network we need? XOR Problem Three-Class Problem Network Structure 2 3 3 2 40 60 3 . . . . . . ANN: Multilayer Perceptron How big the network we need? Network Structure 2\60\10\2 2\3\2\2 Intended Training Error Solution: (1) Construct a big network then start cutting (2) Using some kinds of algorithms to disable some weights ANN: Multilayer Perceptron Training samples : Class 1 : Class 2 Applications in image processing ANN: Multilayer Perceptron Supervised Classification Winner takes all 0.66 0.26 0.11 0.23 Classified as class 1 output Others: Line detection Image filter Radar image parameter detec