opencv的features2d包中提供了surf,sift和orb等特征点算法，用于图像查找图像对象，搜索对象，分析对象，识别对象，合成全景等场合。

研究这些算法的原理和实现，是图像识别基础，OPENCV库使用2.413

通过一些代码研究三种特征点算法，我有意把原始图像转为灰度并放置90与照处中人物比较，以研究三种算法对人脸识别的优点和局限。辅助使用了人脸查找获取待查找图像中人脸子矩阵。上代码。

import java.util.ArrayList;
import java.util.LinkedList;
import java.util.List;

import org.opencv.core.Core;
import org.opencv.core.Mat;
import org.opencv.core.MatOfDMatch;
import org.opencv.core.MatOfKeyPoint;
import org.opencv.core.MatOfRect;
import org.opencv.core.Size;
import org.opencv.features2d.DMatch;
import org.opencv.features2d.DescriptorExtractor;
import org.opencv.features2d.DescriptorMatcher;
import org.opencv.features2d.FeatureDetector;
import org.opencv.features2d.Features2d;
import org.opencv.highgui.Highgui;
import org.opencv.imgproc.Imgproc;
import org.opencv.objdetect.CascadeClassifier;

public class ExtractSIFT2 {
	public static void main(String[] args) {
		System.loadLibrary(Core.NATIVE_LIBRARY_NAME);
		Mat src = Highgui.imread("E:/work/qqq/Y9.jpg");
		Mat dst = Highgui.imread("E:/work/qqq/psb.jpg");
		MatOfRect mr = getFace(dst);
		Mat sub = dst.submat(mr.toArray()[0]);

		Highgui.imwrite("E:/work/qqq/Y4.jpg", FeatureSurfBruteforce(src.t(), sub));
		Highgui.imwrite("E:/work/qqq/Y5.jpg", FeatureSiftLannbased(src.t(), sub));
		Highgui.imwrite("E:/work/qqq/Y6.jpg", FeatureOrbLannbased(src.t(), sub));
	}
	
	public static Mat FeatureSurfBruteforce(Mat src, Mat dst){
		FeatureDetector fd = FeatureDetector.create(FeatureDetector.SURF);
		DescriptorExtractor de = DescriptorExtractor.create(DescriptorExtractor.SURF);
		//DescriptorMatcher Matcher = DescriptorMatcher.create(DescriptorMatcher.FLANNBASED);
		DescriptorMatcher Matcher = DescriptorMatcher.create(DescriptorMatcher.BRUTEFORCE_L1);
		
		MatOfKeyPoint mkp = new MatOfKeyPoint();
		fd.detect(src, mkp);
		Mat desc = new Mat();
		de.compute(src, mkp, desc);
		Features2d.drawKeypoints(src, mkp, src);
		
		
		MatOfKeyPoint mkp2 = new MatOfKeyPoint();
		fd.detect(dst, mkp2);
		Mat desc2 = new Mat();
		de.compute(dst, mkp2, desc2);
		Features2d.drawKeypoints(dst, mkp2, dst);
		
		
		// Matching features
		MatOfDMatch Matches = new MatOfDMatch();
		Matcher.match(desc, desc2, Matches);
		
		double maxDist = Double.MIN_VALUE;
		double minDist = Double.MAX_VALUE;

		DMatch[] mats = Matches.toArray();
		for (int i = 0; i < mats.length; i++) {
			double dist = mats[i].distance;
			if (dist < minDist) {
				minDist = dist;
			}
			if (dist > maxDist) {
				maxDist = dist;
			}
		}
		System.out.println("Min Distance:" + minDist);
		System.out.println("Max Distance:" + maxDist);
		List<DMatch> goodMatch = new LinkedList<>();

		for (int i = 0; i < mats.length; i++) {
			double dist = mats[i].distance;
			if (dist < 3 * minDist && dist < 0.2f) {
				goodMatch.add(mats[i]);
			}
		}
		
		Matches.fromList(goodMatch);
		// Show result
		Mat OutImage = new Mat();
		Features2d.drawMatches(src, mkp, dst, mkp2, Matches, OutImage);
		
		return OutImage;
	}
	public static Mat FeatureSiftLannbased(Mat src, Mat dst){
		FeatureDetector fd = FeatureDetector.create(FeatureDetector.SIFT);
		DescriptorExtractor de = DescriptorExtractor.create(DescriptorExtractor.SIFT);
		DescriptorMatcher Matcher = DescriptorMatcher.create(DescriptorMatcher.FLANNBASED);
		
		MatOfKeyPoint mkp = new MatOfKeyPoint();
		fd.detect(src, mkp);
		Mat desc = new Mat();
		de.compute(src, mkp, desc);
		Features2d.drawKeypoints(src, mkp, src);
		
		MatOfKeyPoint mkp2 = new MatOfKeyPoint();
		fd.detect(dst, mkp2);
		Mat desc2 = new Mat();
		de.compute(dst, mkp2, desc2);
		Features2d.drawKeypoints(dst, mkp2, dst);
		
		
		// Matching features
		MatOfDMatch Matches = new MatOfDMatch();
		Matcher.match(desc, desc2, Matches);
		
		List<DMatch> l = Matches.toList();
		List<DMatch> goodMatch = new ArrayList<DMatch>();
		for (int i = 0; i < l.size(); i++) {
			DMatch dmatch = l.get(i);
			if (Math.abs(dmatch.queryIdx - dmatch.trainIdx) < 10f) {
				goodMatch.add(dmatch);
			}
			
		}
		
		Matches.fromList(goodMatch);
		// Show result
		Mat OutImage = new Mat();
		Features2d.drawMatches(src, mkp, dst, mkp2, Matches, OutImage);
		
		return OutImage;
	}
	public static Mat FeatureOrbLannbased(Mat src, Mat dst){
		FeatureDetector fd = FeatureDetector.create(FeatureDetector.ORB);
		DescriptorExtractor de = DescriptorExtractor.create(DescriptorExtractor.ORB);
		DescriptorMatcher Matcher = DescriptorMatcher.create(DescriptorMatcher.BRUTEFORCE_L1);
		
		MatOfKeyPoint mkp = new MatOfKeyPoint();
		fd.detect(src, mkp);
		Mat desc = new Mat();
		de.compute(src, mkp, desc);
		Features2d.drawKeypoints(src, mkp, src);
		
		MatOfKeyPoint mkp2 = new MatOfKeyPoint();
		fd.detect(dst, mkp2);
		Mat desc2 = new Mat();
		de.compute(dst, mkp2, desc2);
		Features2d.drawKeypoints(dst, mkp2, dst);
		
		
		// Matching features
	
		MatOfDMatch Matches = new MatOfDMatch();
		Matcher.match(desc, desc2, Matches);
		
		double maxDist = Double.MIN_VALUE;
		double minDist = Double.MAX_VALUE;

		DMatch[] mats = Matches.toArray();
		for (int i = 0; i < mats.length; i++) {
			double dist = mats[i].distance;
			if (dist < minDist) {
				minDist = dist;
			}
			if (dist > maxDist) {
				maxDist = dist;
			}
		}
		System.out.println("Min Distance:" + minDist);
		System.out.println("Max Distance:" + maxDist);
		List<DMatch> goodMatch = new LinkedList<>();

		for (int i = 0; i < mats.length; i++) {
			double dist = mats[i].distance;
			if (dist < 3 * minDist && dist < 0.2f) {
				goodMatch.add(mats[i]);
			}
		}
		
		Matches.fromList(goodMatch);
		// Show result
		Mat OutImage = new Mat();
		Features2d.drawMatches(src, mkp, dst, mkp2, Matches, OutImage);
		
		//Highgui.imwrite("E:/work/qqq/Y4.jpg", OutImage);
		return OutImage;
	}
	public static MatOfRect getFace(Mat src) {
		Mat result = src.clone();
		if (src.cols() > 1000 || src.rows() > 1000) {
			Imgproc.resize(src, result, new Size(src.cols() / 3, src.rows() / 3));
		}

		CascadeClassifier faceDetector = new CascadeClassifier("./resource/haarcascade_frontalface_alt2.xml");
		MatOfRect objDetections = new MatOfRect();
		faceDetector.detectMultiScale(result, objDetections);
		
		return objDetections;
	}
}

人脸灰度图，待处理的图片和处理后三种方法对比，做了一些简单的取优。结果来看，orb算法似乎优于其他两种。

参考：

http://blog.csdn.net/liufanghuangdi/article/details/52957094?locationNum=2&fps=1

http://blog.csdn.net/shuzhe66/article/details/40824883