JAVA实现的Locality Sensitive Hash

我在之前的博客已大致介绍了LSH的原理及其的适用场景，有兴趣的朋友可以移步至
[url]http://grunt1223.iteye.com/blog/937600[/url]

这里我给出它的具体实现及实验效果：

private int dimention; //维度大小，例如对于sift特征来说就是128
private int max; //所需向量中元素可能的上限，譬如对于RGB来说，就是255
private int hashCount; //哈希表的数量，用于更大程度地削减false positive
//LSH随机选取的采样位数，该值越小，则近似查找能力越大，但相应的false positive也越大；若该值等于size，则为由近似查找退化为精确匹配
private int bitCount; 
private int size; //转化为01字符串之后的位数，等于max乘以dimensions
private int[][] hashFamily; //LSH哈希族，保存了随机采样点的INDEX
VectorComparator comparator; 
// private HashMap<String, ArrayList<IdentifiedVector>> map;
private HashMap<String, ArrayList<String>> map;

public SimpleLSH(int dimention, int max, int hashCount, int bitCount) {
    this.dimention = dimention;
    this.max = max;
    this.hashCount = hashCount;
    this.bitCount = bitCount;
    this.size = this.dimention * this.max;
    this.hashFamily = new int[hashCount][bitCount];
    // map = new HashMap<String, ArrayList<IdentifiedVector>>();
    map = new HashMap<String, ArrayList<String>>();
    this.comparator = new VectorComparator(new int[] { 0 });
}

//生成随机的投影点
private void generataHashFamily() {
    Random rd = new Random();
    for (int i = 0; i < hashCount; i++) {
    for (int j = 0; j < bitCount; j++) {
        hashFamily[i][j] = rd.nextInt(size);
    }
}

//将向量转化为二进制字符串，比如元素的最大范围255，则元素65就被转化为65个1以及190个0
private int[] unAray(int[] data) {
    int unArayData[] = new int[size];
    for (int i = 0; i < data.length; i++) {
        for (int j = 0; j < data[i]; j++) {
	    unArayData[i * max + j] = 1;
        }
    }
    return unArayData;
}

//将向量映射为LSH中的key
private String generateHashKey(int[] list, int hashNum) {
    StringBuilder sb = new StringBuilder();
    int[] tempData = unAray(list);
    int[] hashedData = new int[bitCount];
    //首先将向量转为二进制字符串
    for (int i = 0; i < bitCount; i++) {
        hashedData[i] = tempData[hashFamily[hashNum][i]];
	sb.append(hashedData[i]);
	// System.out.print(hashedData[i]);
    }

    //再用常规hash函数比如MD5对key进行压缩
    MessageDigest messageDigest = null;
    try 
    {
        messageDigest = MessageDigest.getInstance("MD5");
    }
    catch (NoSuchAlgorithmException e) {
    }

    byte[] binary = sb.toString().getBytes();
    byte[] hash = messageDigest.digest(binary);
    String hashV = MathUtils.getHexDigest(hash);
    return hashV;
}

//将向量映射为LSH中的key，并保存至map中
private void generateHashMap(String id, int[] vercotr) {
    for (int j = 0; j < hashCount; j++) {
        String key = generateHashKey(vercotr, j);
        ArrayList<String> l;
        if (map.containsKey(key)) {
            l = map.get(key);
        } else {
            l = new ArrayList<String>();
        }
        l.add(id);
        map.put(key, l);
    }
}

// 查询与输入向量最接近（海明空间）的向量
public Set<String> query(int[] data) {
    // Set<IdentifiedVector> result = new HashSet<IdentifiedVector>();
    Set<String> result = new HashSet<String>();
    for (int j = 0; j < hashCount; j++) {
        String key = generateHashKey(data, j);
        if (map.containsKey(key)) {
            result.addAll(map.get(key));
        }
    }
    return result;
}

我利用上面的LSH对图片的边缘直方图特征进行建模，获得了不错的效果，可以用于近似图片的查找，效果如下：

[img]http://dl.iteye.com/upload/attachment/430748/d695bda2-6ad5-3afc-8ca4-fa4e10f08509.jpg[/img]