实验四、基于三层BP神经网络的人脸识别
一、 实验要求
采用三层前馈BP神经网络实现标准人脸YALE数据库的识别，编程语言为C系列语言。
二、BP神经网络的结构和学习算法
实验中建议采用如下最简单的三层BP神经网络，输入层为 ，有n个神经元节点，输出层具有m个神经元，网络输出为 ,隐含层具有k个神经元，采用BP学习算法训练神经网络。

BP神经网络的结构
BP网络在本质上是一种输入到输出的映射，它能够学习大量的输入与输出之间的映射关系，而不需要任何输入和输出之间的精确的数学表达式，只要用已知的模式对BP网络加以训练，网络就具有输入输出对之间的映射能力。
BP网络执行的是有教师训练，其样本集是由形如（输入向量，期望输出向量）的向量对构成的。在开始训练前，所有的权值和阈值都应该用一些不同的小随机数进行初始化。
…

import java.awt.image.BufferedImage;
import java.io.File;
import java.io.IOException;
import java.util.Scanner;import javax.imageio.ImageIO;public class ANN {static int B = 16;
//	输入层神经元个数static int N = 2*B*B*2;
//	隐含层神经元个数static int L = B*B;
//	输出层神经元个数static int M = 16;
//	进度控制产生static double yiphxir = 0.001;
//	学习率static double arph = 0.2;static double a = -1.0;static double max = 255;
//	输入层到隐含层的权值static double[][] V = new double[N][L];
//	隐含层到输出层的权值static double[][] W = new double[L][M];
//	隐含层的阈值static double[] fai = new double[L];
//	输出层的阈值static double[] sita = new double[M];static double[][] deltaW = new double[L][M];static double[] deltaSita = new double[M];static double[][] deltaV = new double[N][L];static double[] deltaFai = new double[L];static double[] FE = new double[M];static double[] getImagePixel(String image) {File file = new File(image);BufferedImage bi = null;try {bi = ImageIO.read(file);} catch (IOException e) {e.printStackTrace();}int width = bi.getWidth()/2;int height = bi.getHeight()/2;double[] pixel = new double[N];int cnt = 0;for(int y = height-B; y < height+B; y++) {for(int x = width-B; x < width+B; x++) {int t = bi.getRGB(x, y);t = (t & 0xff);pixel[cnt] = t/max;cnt++;}}return pixel;}static double FS(double x){return 1.0/(1.0+Math.exp(a*x)); }static void init(){for(int i=0; i<N; i++){for(int j=0; j<L; j++){V[i][j] = Math.random()*2.0-1.0;}}for(int j=0; j<L; j++){for(int k=1; k<M; k++){W[j][k] = Math.random()*2.0-1.0;}}for(int j=0; j<L; j++){fai[j] = Math.random()*2.0-1.0;}for(int k=1; k<M; k++){sita[k] = Math.random()*2.0-1.0;}}static boolean trainBP(double[] X, int[] D){double[] H = new double[L];for(int j=0; j<L; j++){double derta = 0.0;for(int i=0; i<N; i++){derta += X[i]*V[i][j];}
//			System.out.print("derta"+Double.toString(derta));
//			System.out.println("    derta-fai="+Double.toString(derta-fai[j]));H[j] = FS(derta+fai[j]);
//			System.out.println(Double.toString(H[j])+" ");}double[] Y = new double[M];for(int k=1; k<M; k++){double derta = 0.0;for(int j=0; j<L; j++){derta += H[j]*W[j][k];}
//			System.out.print("derta"+Double.toString(derta));
//			System.out.println("    derta-sita="+Double.toString(derta-sita[k]));Y[k] = FS(derta+sita[k]);
//			System.out.print(Double.toString(Y[k])+" ");}
//		System.out.println();double[] deltaK = new double[M];double E = 0.0;for(int k=1; k<M; k++){E += (D[k] - Y[k])*(D[k] - Y[k]);deltaK[k] = (D[k]-Y[k])*Y[k]*(1-Y[k]);}
//		System.out.print("E=");
//		System.out.println(E/2);if(E/2 < yiphxir) return true;double[] deltaJ = new double[L];for(int j=0; j<L; j++){double beta = 0.0;for(int k=1; k<M; k++){beta += deltaK[k]*W[j][k];}deltaJ[j] = H[j]*(1-H[j])*beta;}for(int j=0; j<L; j++){for(int k=1; k<M; k++){deltaW[j][k] = arph*deltaK[k]*H[j];
//				deltaW[j][k] = (arph/(1+L))*(deltaW[j][k]+1)*deltaK[k]*H[j];W[j][k] += deltaW[j][k];}}for(int k=1; k<M; k++){deltaSita[k] = arph*deltaK[k];
//			deltaSita[k] = (arph/(1+L))*(detaSitla[k]+1)*deltaK[k];sita[k] += deltaSita[k];}for(int i=0; i<N; i++){for(int j=0; j<L; j++){deltaV[i][j] = arph*deltaJ[j]*X[i];
//				deltaV[i][j] = (arph/(1+N))*(deltaV[i][j]+1)*deltaJ[j]*X[i];V[i][j] += deltaV[i][j];}}for(int j=0; j<L; j++){deltaFai[j] = arph*deltaJ[j];
//			deltaFai[j] = (arph/(1+N))*(deltaFai[j]+1)*deltaJ[j];fai[j] += deltaFai[j];}return false;}static int BP(double[] X){int ans = 0;double[] H = new double[L];for(int j=0; j<L; j++){double derta = 0.0;for(int i=0; i<N; i++){derta += X[i]*V[i][j];}H[j] = FS(derta-fai[j]);}double[] Y = new double[M];for(int k=1; k<M; k++){double derta = 0.0;for(int j=0; j<L; j++){derta += H[j]*W[j][k];}Y[k] = FS(derta-sita[k]);}double min = Double.MAX_VALUE;for(int n=1; n<M; n++){double E = 0.0;int[] D = new int[M];D[n] = 1;for(int k=1; k<M; k++){E += Math.abs((D[k] - Y[k])*(D[k] - Y[k]));}FE[n] = 1.0/Math.exp(E/2.0);if(E/2 < min){min = E/2;ans = n;}}return ans;}@SuppressWarnings("resource")public static void main(String[] args) {init();int temp = 0;while(temp < 100){int t = 1;while(t <= 11){int ca = 1;while(ca < M){
//					System.out.println(t);
//					System.out.println(ca);String imagePath;if(ca >= 10){
//						imagePath = "C:\\Users\\DELL\\Downloads\\AAII\\src\\image\\subject01_10.bmp";imagePath = "C:\\Users\\DELL\\Downloads\\AAII\\src\\image\\"+"subject"+Integer.toString(ca)+"_"+Integer.toString(t)+".bmp";}else imagePath = "C:\\Users\\DELL\\Downloads\\AAII\\src\\image\\"+"subject0"+Integer.toString(ca)+"_"+Integer.toString(t)+".bmp";double[] X = getImagePixel(imagePath);int[] D = new int[M];D[ca] = 1;boolean flag = trainBP(X, D);if(flag) break;ca++;}t++;}temp++;}Scanner in = new Scanner(System.in);while(true){System.out.println("请输入subject(t1)和_(t2):");String t1 = in.next();String t2 = in.next();String imagePath = "C:\\Users\\DELL\\Downloads\\AAII\\src\\image\\"+"subject"+t1+"_"+t2+".bmp";int ans = BP(X);System.out.println("(subject01~subject15)的匹配率如下：");for(int i=1; i<M; i++){System.out.println("subject"+Integer.toString(i)+"："+String.format("%g", FE[i]*100)+"%");}System.out.println();System.out.print("综合结果为subject");System.out.println(ans);System.out.println();}}
}

图片地址：https://download.csdn.net/download/qq_43179428/19880190