/// <summary>
/// 获取优秀的匹配点
/// <para>
/// 使用 "比较最近邻距离与次近邻距离的SIFT匹配方式" 和 "随机抽样一致(RANSAC)算法" 来进一步获取优秀匹配点
/// </para>
/// </summary>
/// <param name="matches"></param>
/// <param name="argument"></param>
/// <param name="sourceKeyPoints"></param>
/// <param name="searchKeyPoints"></param>
/// <returns></returns>
private IList <DMatch> SelectGoodMatches(IEnumerable <DMatch[]> matches, FeatureMatchArgument argument,
IList <KeyPoint> sourceKeyPoints, IList <KeyPoint> searchKeyPoints)
{
var sourcePoints = new List <Point2d>();
var searchPoints = new List <Point2d>();
var goodMatches = new List <DMatch>();
//比较最近邻距离与次近邻距离的SIFT匹配方式
foreach (var items in matches)
{
if (items.Length == 0)
{
continue;
}
if (argument.MatchPoints > 1 &&
(items.Length < 2 || items[0].Distance > argument.Ratio * items[1].Distance))
{
continue;
}
//此处直接选择欧氏距离小的匹配关键点
var goodMatch = items[0];
goodMatches.Add(goodMatch);
sourcePoints.Add(Point2FToPoint2D(sourceKeyPoints[goodMatch.QueryIdx].Pt));
searchPoints.Add(Point2FToPoint2D(searchKeyPoints[goodMatch.TrainIdx].Pt));
}
argument.OutputDebugMessage($"[FeatureMatch] [SIFT] The number of good matching points is ({goodMatches.Count}).");
//随机抽样一致(RANSAC)算法 (如果原始的匹配结果为空, 则跳过过滤步骤）
if (sourcePoints.Count >= MinCorrespondingPointCount && searchPoints.Count >= MinCorrespondingPointCount)
{
var inliersMask = new Mat();
Cv2.FindHomography(sourcePoints, searchPoints, HomographyMethods.Ransac, argument.RansacThreshold, inliersMask);
// 如果通过RANSAC处理后的匹配点大于10个,才应用过滤. 否则使用原始的匹配点结果(匹配点过少的时候通过RANSAC处理后,可能会得到0个匹配点的结果).
if (inliersMask.Rows > 10)
{
inliersMask.GetArray(out byte[] maskBytes);
var inliers = new List <DMatch>();
for (var i = 0; i < maskBytes.Length; i++)
{
if (maskBytes[i] != 0)
{
inliers.Add(goodMatches[i]);
}
}
argument.OutputDebugMessage($"[FeatureMatch] [SIFT] The number of good matching RANSAC points is ({inliers.Count}).");
return(inliers);
}
}
return(goodMatches);
}
public override FeatureMatchResult Match(Mat sourceMat, Mat searchMat, FeatureMatchArgument argument)
{
//创建SIFT
using var sift = SIFT.Create();
//创建特征点描述对象，为下边的特征点匹配做准备
using var sourceDescriptors = new Mat();
using var searchDescriptors = new Mat();
//提取特征点，并进行特征点描述
sift.DetectAndCompute(sourceMat, null, out var sourceKeyPoints, sourceDescriptors);
sift.DetectAndCompute(searchMat, null, out var searchKeyPoints, searchDescriptors);
//创建Brute-force descriptor matcher
using var bfMatcher = new BFMatcher();
//对原图特征点描述加入训练
bfMatcher.Add(new List <Mat>()
{
sourceDescriptors
});
bfMatcher.Train();
//获得匹配特征点，并提取最优配对
var matches = bfMatcher.KnnMatch(sourceDescriptors, searchDescriptors, (int)argument.MatchPoints);
argument.OutputDebugMessage($"[FeatureMatch] [SIFT] The number of matching points is ({matches.Length}).");
//即使使用SIFT算法，但此时没有经过点筛选的匹配效果同样糟糕，所进一步获取优秀匹配点
var goodMatches = SelectGoodMatches(matches, argument, sourceKeyPoints, searchKeyPoints);
//获取匹配结果
var matchResult = GetMatchResult(goodMatches, sourceKeyPoints, searchKeyPoints);
argument.OutputDebugMessage($"[FeatureMatch] [SIFT] The result of the match is ({matchResult.Success}) ({matchResult.MatchItems.Count}).");
if (matchResult.Success)
{
var bestMatch = matchResult.MatchItems[0];
argument.OutputDebugMessage($"[FeatureMatch] [SIFT] The center point of the best match is ({bestMatch.Point}), and the rect is {bestMatch.Rectangle}.");
}
argument.PreviewDebugFeatureMatchResult(matchResult, sourceMat, searchMat, sourceKeyPoints, searchKeyPoints, goodMatches);
return(matchResult);
}