/*
* Copyright 2013 Alibaba.com All right reserved. This software is the
* confidential and proprietary information of Alibaba.com ("Confidential
* Information"). You shall not disclose such Confidential Information and shall
* use it only in accordance with the terms of the license agreement you entered
* into with Alibaba.com.
*/
package com.alibaba.simpleimage.analyze.sift;
import java.util.ArrayList;
import java.util.List;
import com.alibaba.simpleimage.analyze.sift.scale.FeaturePoint;
import com.alibaba.simpleimage.analyze.sift.scale.GaussianArray;
import com.alibaba.simpleimage.analyze.sift.scale.KDFeaturePoint;
import com.alibaba.simpleimage.analyze.sift.scale.OctaveSpace;
import com.alibaba.simpleimage.analyze.sift.scale.Pyramid;
import com.alibaba.simpleimage.analyze.sift.scale.ScalePeak;
/**
* 类SIFT.java的实现描述：面向用户的主操作接口
*
* @author axman 2013-6-28 上午10:15:45
*/
public class SIFT {
// 以下常数均是论文中推荐的参数值
float preprocSigma = 1.5f; // 用于处理被double后的图像预处理的模糊因子
float octaveSigma = 1.6f; // 用于处理每个8度空间图像的模糊因子
int minimumRequiredPixelsize = 32; // 高斯金字塔中缩小时最小的尺寸
int scaleSpaceLevels = 3; // 每个8度空间需要获取极值点的差分图层，加上用于比较的上下层至少要有5个差分图像，所以至少要有6个高斯模糊图象。
float dogThresh = 0.0075f; // 在差分图象上的极致点值最小值，防止大片的模糊后的点被选中，这个值越小选中的点越多。
float dValueLowThresh = 0.008f; // 和周围点比较的差值，这个差值是经过导数运算的差值，不是直接比较的。论文中建议为0.03（page
// 11），但获取的点数太少，这里修改为0.008
float maximumEdgeRatio = 20.0f; // 非角点的过虑比
float scaleAdjustThresh = 0.50f; // 尺度空间的精确点和真实图象上的离散点在投谢时需要调整，这个是最大调整范围，超这个值就可能是下一个点。
float peakRelThresh = 0.8f; //
int relocationMaximum = 4;
public int detectFeatures(ImagePixelArray img) {
return (detectFeaturesDownscaled(img, -1, 1.0f));
}
/**
* @param img
* @param preProcessMark 图象预处理的标记，小于0，img需要double,大于0时，说明图象的长和宽要half到这个尺寸以下，等于0则不预处理
* @param startScale
* @return
*/
public int detectFeaturesDownscaled(ImagePixelArray img, int preProcessMark, float startScale) {
if (preProcessMark < 0) {
img = img.doubled();
startScale *= 0.5;
} else if (preProcessMark > 0) {
while (img.width > preProcessMark || img.height > preProcessMark) {
img = img.halved();
startScale *= 2.0;
}
}
if (preprocSigma > 0.0) {
GaussianArray gaussianPre = new GaussianArray(preprocSigma);
img = gaussianPre.convolve(img);
}
Pyramid pyr = new Pyramid();
pyr.buildOctaves(img, startScale, scaleSpaceLevels, octaveSigma, minimumRequiredPixelsize);
globalFeaturePoints = new ArrayList<FeaturePoint>();
// Generate featurePoints from each scalespace.
for (int on = 0; on < pyr.octaves.size(); ++on) {
OctaveSpace osp = pyr.octaves.get(on);
ArrayList<ScalePeak> peaks = osp.findPeaks(dogThresh);// 寻找图片中的极值点
ArrayList<ScalePeak> peaksFilted = osp.filterAndLocalizePeaks(peaks, maximumEdgeRatio, dValueLowThresh,
scaleAdjustThresh, relocationMaximum);
// 先将要处理的图层上所有象素的梯度大小和方向计算出来
osp.pretreatMagnitudeAndDirectionImgs();
ArrayList<FeaturePoint> faturePoints = osp.makeFeaturePoints(peaksFilted, peakRelThresh, scaleSpaceLevels,
octaveSigma);
osp.clear();
globalFeaturePoints.addAll(faturePoints);
}
return (globalFeaturePoints.size());
}
private List<FeaturePoint> globalFeaturePoints;
private List<KDFeaturePoint> globalKDFeaturePoints;
public List<KDFeaturePoint> getGlobalKDFeaturePoints() {
if (globalKDFeaturePoints != null) return (globalKDFeaturePoints);
if (globalFeaturePoints == null) throw (new IllegalArgumentException("No featurePoints generated yet."));
globalKDFeaturePoints = new ArrayList<KDFeaturePoint>();
for (FeaturePoint fp : globalFeaturePoints) {
globalKDFeaturePoints.add(new KDFeaturePoint(fp));
}
return globalKDFeaturePoints;
}
}