一：基本原理

NCC是一種基於統計學計算兩組樣本資料相關性的演算法，其取值範圍為[-1, 1]之間，而對影象來說，每個畫素點都可以看出是RGB數值，這樣整幅影象就可以看成是一個樣本資料的集合，如果它有一個子集與另外一個樣本資料相互匹配則它的ncc值為1，表示相關性很高，如果是-1則表示完全不相關，基於這個原理，實現影象基於模板匹配識別演算法，其中第一步就是要歸一化資料，數學公式如下：

二：實現步驟

（1）      獲取模板畫素並計算均值與標準方差、畫素與均值diff資料樣本

（2）      根據模板大小，在目標影象上從左到右，從上到下移動視窗，計

算每移動一個畫素之後視窗內畫素與模板畫素的ncc值，與閾值比較，大於

閾值則記錄位置

（3）      根據得到位置資訊，使用紅色矩形標記出模板匹配識別結果。

（4）      UI顯示結果

三：程式設計實現

基於JAVA語言完成了整個演算法程式設計實現與演示，其中第一步的程式碼如下：

		int tw = template.getWidth();
		int th = template.getHeight();
		int[] tpixels = new int[tw * th];
		getRGB(template, 0, 0, tw, th, tpixels);
		for(int i=0; i<tpixels.length; i++)
		{
			tpixels[i] = (tpixels[i] >> 16) & 0xff;
		}
		double[] meansdev = getPixelsMeansAndDev(tpixels);
		double[] tDiff = calculateDiff(tpixels, meansdev[0]);
		int raidus_width = tw / 2;
		int raidus_height = th / 2;
 

		int[] windowPixels = new int[tw * th];
		Arrays.fill(windowPixels, 0);
		for (int row = 0; row < height; row++) {
			for (int col = 0; col < width; col++) {
				// calculate the means and dev for each window
				if(row <  raidus_height || (row + raidus_height) >= height)
					continue;
				if(col < raidus_width || (col + raidus_width) >= width) 
					continue;
				int wrow = 0;
				Arrays.fill(windowPixels, 0);
				for(int subrow = -raidus_height; subrow <= raidus_height; subrow++ )
				{
					int wcol = 0;
					for(int subcol = -raidus_width; subcol <= raidus_width; subcol++ )
					{
						if(wrow >= th || wcol >= tw)
						{
							continue;
						}
						windowPixels[wrow * tw + wcol] = getPixelValue(width, col + subcol, row + subrow, inPixels);
						wcol++;
					}
					wrow++;
				}
				// calculate the ncc
				double[] _meansDev = getPixelsMeansAndDev(windowPixels);
				double[] diff = calculateDiff(windowPixels, _meansDev[0]);
				double ncc = calculateNcc(tDiff, diff, _meansDev[1], meansdev[1]);
				if(ncc > threhold) {
					Point mpoint = new Point();
					mpoint.x = col;
					mpoint.y  = row;
					points.add(mpoint);
				}
			}
		}
 

		// draw matched template on target image according position
		setRGB( dest, 0, 0, width, height, inPixels );
		Graphics2D g2d = dest.createGraphics();
		g2d.setPaint(Color.RED);
		g2d.setStroke(new BasicStroke(4));
		for(Point p : points)
		{
			g2d.drawRect(p.x - raidus_width, p.y - raidus_height, tw, th);
		}

	private double calculateNcc(double[] tDiff, double[] diff, double dev1, double dev2) {
		// TODO Auto-generated method stub
		double sum = 0.0d;
		double count = diff.length;
		for(int i=0; i<diff.length; i++)
		{
			sum += ((tDiff[i] * diff[i])/(dev1 * dev2));
		}
		return (sum / count);
	}

package com.gloomyfish.image.templae.match;

import java.awt.BorderLayout;
import java.awt.FlowLayout;
import java.awt.Graphics;
import java.awt.Graphics2D;
import java.awt.event.ActionEvent;
import java.awt.event.ActionListener;
import java.awt.image.BufferedImage;
import java.io.IOException;

import javax.imageio.ImageIO;
import javax.swing.JButton;
import javax.swing.JComponent;
import javax.swing.JFrame;
import javax.swing.JPanel;

public class DemoUI extends JComponent {
	
	/**
	 * 
	 */
	private static final long serialVersionUID = 1L;
	private BufferedImage targetImage;
	private BufferedImage template;
	
	public DemoUI()
	{
		super();
		java.net.URL imageURL = this.getClass().getResource("words.png");
		java.net.URL templateURL = this.getClass().getResource("template.png");
		
		try {
			template = ImageIO.read(templateURL);
			targetImage = ImageIO.read(imageURL);
		} catch (IOException e) {
			e.printStackTrace();
		}
	}
	
	public void setTarget(BufferedImage target) {
		this.targetImage = target;
	}

	@Override
	protected void paintComponent(Graphics g) {
		Graphics2D g2 = (Graphics2D) g;
		if(targetImage != null) {
			g2.drawImage(targetImage, 10, 10, targetImage.getWidth(), targetImage.getHeight(), null);
		}
		if(template != null) {
			g2.drawImage(template, 20+targetImage.getWidth(), 10, template.getWidth(), template.getHeight(), null);
		}
	}
	
	public static void main(String[] args) {
		JFrame f = new JFrame("模板匹配與識別");
		JButton okBtn = new JButton("匹配");
		final DemoUI ui = new DemoUI();
		okBtn.addActionListener(new ActionListener() {

			@Override
			public void actionPerformed(ActionEvent e) {
				
				ui.process();
			}
		});
		
		JPanel btnPanel = new JPanel();
		btnPanel.setLayout(new FlowLayout(FlowLayout.RIGHT));
		btnPanel.add(okBtn);
		
		f.getContentPane().add(btnPanel, BorderLayout.SOUTH);
		f.getContentPane().add(ui, BorderLayout.CENTER);
		f.setSize(500, 500);
		f.setDefaultCloseOperation(JFrame.EXIT_ON_CLOSE);
		f.setVisible(true);
	}

	protected void process() {
		NccTemplateMatchAlg algo = new NccTemplateMatchAlg(template);
		targetImage = algo.filter(targetImage, null);
		this.repaint();
	}

}

其中左邊是目標影象、右邊為模板影象

PS:部落格從10月份開始每月都有多篇相關影象處理文章更新

歡迎大家繼續關注