汉明距离的应用
Google、Baidu 等搜索引擎相继推出了以图搜图的功能,测试了下效果还不错~ 那这种技术的原理是什么呢?计算机怎么知道两张图片相似呢?
根据Neal Krawetz博士的解释,原理非常简单易懂。我们可以用一个快速算法,就达到基本的效果。
这里的关键技术叫做"感知哈希算法"(Perceptual hash algorithm),它的作用是对每张图片生成一个"指纹"(fingerprint)字符串,然后比较不同图片的指纹。结果越接近,就说明图片越相似。
下面是一个最简单的实现:
第一步,缩小尺寸。
将图片缩小到8x8的尺寸,总共64个像素。这一步的作用是去除图片的细节,只保留结构、明暗等基本信息,摒弃不同尺寸、比例带来的图片差异。
第二步,简化色彩。
将缩小后的图片,转为64级灰度。也就是说,所有像素点总共只有64种颜色。
第三步,计算平均值。
计算所有64个像素的灰度平均值。
第四步,比较像素的灰度。
将每个像素的灰度,与平均值进行比较。大于或等于平均值,记为1;小于平均值,记为0。
第五步,计算哈希值。
将上一步的比较结果,组合在一起,就构成了一个64位的整数,这就是这张图片的指纹。组合的次序并不重要,只要保证所有图片都采用同样次序就行了。
得到指纹以后,就可以对比不同的图片,看看64位中有多少位是不一样的。在理论上,这等同于计算"汉明距离"(Hamming distance)。如果不相同的数据位不超过5,就说明两张图片很相似;如果大于10,就说明这是两张不同的图片。
具体的代码实现,可以参见Wote用python语言写的imgHash.py。代码很短,只有53行。使用的时候,第一个参数是基准图片,第二个参数是用来比较的其他图片所在的目录,返回结果是两张图片之间不相同的数据位数量(汉明距离)。
这种算法的优点是简单快速,不受图片大小缩放的影响,缺点是图片的内容不能变更。如果在图片上加几个文字,它就认不出来了。所以,它的最佳用途是根据缩略图,找出原图。
实际应用中,往往采用更强大的pHash算法和SIFT算法,它们能够识别图片的变形。只要变形程度不超过25%,它们就能匹配原图。这些算法虽然更复杂,但是原理与上面的简便算法是一样的,就是先将图片转化成Hash字符串,然后再进行比较。
下面我们来看下上述理论用java来做一个DEMO版的具体实现:
- import java.awt.Graphics2D;
- import java.awt.color.ColorSpace;
- import java.awt.image.BufferedImage;
- import java.awt.image.ColorConvertOp;
- import java.io.File;
- import java.io.FileInputStream;
- import java.io.FileNotFoundException;
- import java.io.InputStream;
- import javax.imageio.ImageIO;
- /*
- * pHash-like image hash.
- * Author: Elliot Shepherd ([email protected]
- * Based On: http://www.hackerfactor.com/blog/index.php?/archives/432-Looks-Like-It.html
- */
- public class ImagePHash {
- private int size = 32;
- private int smallerSize = 8;
- public ImagePHash() {
- initCoefficients();
- }
- public ImagePHash(int size, int smallerSize) {
- this.size = size;
- this.smallerSize = smallerSize;
- initCoefficients();
- }
- public int distance(String s1, String s2) {
- int counter = 0;
- for (int k = 0; k < s1.length();k++) {
- if(s1.charAt(k) != s2.charAt(k)) {
- counter++;
- }
- }
- return counter;
- }
- // Returns a 'binary string' (like. 001010111011100010) which is easy to do a hamming distance on.
- public String getHash(InputStream is) throws Exception {
- BufferedImage img = ImageIO.read(is);
- /* 1. Reduce size.
- * Like Average Hash, pHash starts with a small image.
- * However, the image is larger than 8x8; 32x32 is a good size.
- * This is really done to simplify the DCT computation and not
- * because it is needed to reduce the high frequencies.
- */
- img = resize(img, size, size);
- /* 2. Reduce color.
- * The image is reduced to a grayscale just to further simplify
- * the number of computations.
- */
- img = grayscale(img);
- double[][] vals = new double[size][size];
- for (int x = 0; x < img.getWidth(); x++) {
- for (int y = 0; y < img.getHeight(); y++) {
- vals[x][y] = getBlue(img, x, y);
- }
- }
- /* 3. Compute the DCT.
- * The DCT separates the image into a collection of frequencies
- * and scalars. While JPEG uses an 8x8 DCT, this algorithm uses
- * a 32x32 DCT.
- */
- long start = System.currentTimeMillis();
- double[][] dctVals = applyDCT(vals);
- System.out.println("DCT: " + (System.currentTimeMillis() - start));
- /* 4. Reduce the DCT.
- * This is the magic step. While the DCT is 32x32, just keep the
- * top-left 8x8. Those represent the lowest frequencies in the
- * picture.
- */
- /* 5. Compute the average value.
- * Like the Average Hash, compute the mean DCT value (using only
- * the 8x8 DCT low-frequency values and excluding the first term
- * since the DC coefficient can be significantly different from
- * the other values and will throw off the average).
- */
- double total = 0;
- for (int x = 0; x < smallerSize; x++) {
- for (int y = 0; y < smallerSize; y++) {
- total += dctVals[x][y];
- }
- }
- total -= dctVals[0][0];
- double avg = total / (double) ((smallerSize * smallerSize) - 1);
- /* 6. Further reduce the DCT.
- * This is the magic step. Set the 64 hash bits to 0 or 1
- * depending on whether each of the 64 DCT values is above or
- * below the average value. The result doesn't tell us the
- * actual low frequencies; it just tells us the very-rough
- * relative scale of the frequencies to the mean. The result
- * will not vary as long as the overall structure of the image
- * remains the same; this can survive gamma and color histogram
- * adjustments without a problem.
- */
- String hash = "";
- for (int x = 0; x < smallerSize; x++) {
- for (int y = 0; y < smallerSize; y++) {
- if (x != 0 && y != 0) {
- hash += (dctVals[x][y] > avg?"1":"0");
- }
- }
- }
- return hash;
- }
- private BufferedImage resize(BufferedImage image, int width, int height) {
- BufferedImage resizedImage = new BufferedImage(width, height, BufferedImage.TYPE_INT_ARGB);
- Graphics2D g = resizedImage.createGraphics();
- g.drawImage(image, 0, 0, width, height, null);
- g.dispose();
- return resizedImage;
- }
- private ColorConvertOp colorConvert = new ColorConvertOp(ColorSpace.getInstance(ColorSpace.CS_GRAY), null);
- private BufferedImage grayscale(BufferedImage img) {
- colorConvert.filter(img, img);
- return img;
- }
- private static int getBlue(BufferedImage img, int x, int y) {
- return (img.getRGB(x, y)) & 0xff;
- }
- // DCT function stolen from http://stackoverflow.com/questions/4240490/problems-with-dct-and-idct-algorithm-in-java
- private double[] c;
- private void initCoefficients() {
- c = new double[size];
- for (int i=1;i<size;i++) {
- c[i]=1;
- }
- c[0]=1/Math.sqrt(2.0);
- }
- private double[][] applyDCT(double[][] f) {
- int N = size;
- double[][] F = new double[N][N];
- for (int u=0;u<N;u++) {
- for (int v=0;v<N;v++) {
- double sum = 0.0;
- for (int i=0;i<N;i++) {
- for (int j=0;j<N;j++) {
- sum+=Math.cos(((2*i+1)/(2.0*N))*u*Math.PI)*Math.cos(((2*j+1)/(2.0*N))*v*Math.PI)*(f[i][j]);
- }
- }
- sum*=((c[u]*c[v])/4.0);
- F[u][v] = sum;
- }
- }
- return F;
- }
- public static void main(String[] args) {
- ImagePHash p = new ImagePHash();
- String image1;
- String image2;
- try {
- image1 = p.getHash(new FileInputStream(new File("C:/Users/june/Desktop/1.jpg")));
- image2 = p.getHash(new FileInputStream(new File("C:/Users/june/Desktop/1.jpg")));
- System.out.println("1:1 Score is " + p.distance(image1, image2));
- image1 = p.getHash(new FileInputStream(new File("C:/Users/june/Desktop/1.jpg")));
- image2 = p.getHash(new FileInputStream(new File("C:/Users/june/Desktop/2.jpg")));
- System.out.println("1:2 Score is " + p.distance(image1, image2));
- image1 = p.getHash(new FileInputStream(new File("C:/Users/june/Desktop/1.jpg")));
- image2 = p.getHash(new FileInputStream(new File("C:/Users/june/Desktop/3.jpg")));
- System.out.println("1:3 Score is " + p.distance(image1, image2));
- image1 = p.getHash(new FileInputStream(new File("C:/Users/june/Desktop/2.jpg")));
- image2 = p.getHash(new FileInputStream(new File("C:/Users/june/Desktop/3.jpg")));
- System.out.println("2:3 Score is " + p.distance(image1, image2));
- image1 = p.getHash(new FileInputStream(new File("C:/Users/june/Desktop/4.jpg")));
- image2 = p.getHash(new FileInputStream(new File("C:/Users/june/Desktop/5.jpg")));
- System.out.println("4:5 Score is " + p.distance(image1, image2));
- } catch (FileNotFoundException e) {
- e.printStackTrace();
- } catch (Exception e) {
- e.printStackTrace();
- }
- }
- }
运行结果为:
- DCT: 163
- DCT: 158
- 1:1 Score is 0
- DCT: 168
- DCT: 164
- 1:2 Score is 4
- DCT: 156
- DCT: 156
- 1:3 Score is 3
- DCT: 157
- DCT: 157
- 2:3 Score is 1
- DCT: 157
- DCT: 156
- 4:5 Score is 21
结果说明:汉明距离越大表明图片差异越大,如果不相同的数据位不超过5,就说明两张图片很相似;如果大于10,就说明这是两张不同的图片。从结果可以看到1、2、3是相似图片,4、5差异太大,是两张不同的图片。
附:图1、2、3
图1
图2
图3
参考地址:
代码参考:http://pastebin.com/Pj9d8jt5
原理参考:http://www.ruanyifeng.com/blog/2011/07/principle_of_similar_image_search.html
汉明距离:http://baike.baidu.com/view/725269.htm