mirror of
https://github.com/DrizzleTime/Foxel.git
synced 2026-05-13 06:29:54 +08:00
449 lines
16 KiB
C#
449 lines
16 KiB
C#
using Foxel.Models.DataBase;
|
||
using Microsoft.EntityFrameworkCore;
|
||
|
||
namespace Foxel.Services.AI;
|
||
|
||
public class FaceClusteringService(
|
||
IDbContextFactory<MyDbContext> contextFactory,
|
||
ILogger<FaceClusteringService> logger) : IFaceClusteringService
|
||
{
|
||
private const double BASE_SIMILARITY_THRESHOLD = 0.3;
|
||
private const double HIGH_CONFIDENCE_THRESHOLD = 0.5;
|
||
private const int MAX_COMPARISON_FACES = 10;
|
||
|
||
public async Task<List<FaceCluster>> ClusterFacesAsync()
|
||
{
|
||
await using var dbContext = await contextFactory.CreateDbContextAsync();
|
||
|
||
var unclusteredFaces = await dbContext.Faces
|
||
.Where(f => f.Embedding != null && f.ClusterId == null)
|
||
.Include(f => f.Picture)
|
||
.ToListAsync();
|
||
|
||
var clusters = new List<FaceCluster>();
|
||
|
||
foreach (var face in unclusteredFaces)
|
||
{
|
||
var assignedCluster = await FindBestClusterAsync(face, clusters, dbContext);
|
||
|
||
if (assignedCluster != null)
|
||
{
|
||
// 分配到现有聚类
|
||
face.ClusterId = assignedCluster.Id;
|
||
}
|
||
else
|
||
{
|
||
// 创建新聚类
|
||
var newCluster = new FaceCluster
|
||
{
|
||
Name = $"未知人物 {clusters.Count + 1}",
|
||
CreatedAt = DateTime.UtcNow
|
||
};
|
||
|
||
dbContext.FaceClusters.Add(newCluster);
|
||
await dbContext.SaveChangesAsync();
|
||
|
||
face.ClusterId = newCluster.Id;
|
||
clusters.Add(newCluster);
|
||
}
|
||
}
|
||
|
||
await dbContext.SaveChangesAsync();
|
||
logger.LogInformation("人脸聚类完成，共处理 {FaceCount} 个人脸，生成 {ClusterCount} 个聚类",
|
||
unclusteredFaces.Count, clusters.Count);
|
||
|
||
return clusters;
|
||
}
|
||
|
||
public async Task<FaceCluster?> AssignFaceToClusterAsync(int faceId)
|
||
{
|
||
await using var dbContext = await contextFactory.CreateDbContextAsync();
|
||
|
||
var face = await dbContext.Faces
|
||
.Include(f => f.Picture)
|
||
.FirstOrDefaultAsync(f => f.Id == faceId);
|
||
|
||
if (face?.Embedding == null) return null;
|
||
|
||
// 获取所有现有聚类的代表人脸
|
||
var existingClusters = await dbContext.FaceClusters
|
||
.Include(c => c.Faces.Take(1))
|
||
.ToListAsync();
|
||
|
||
foreach (var cluster in existingClusters)
|
||
{
|
||
if (cluster.Faces?.Any() == true)
|
||
{
|
||
var representativeFace = cluster.Faces.First();
|
||
if (representativeFace.Embedding != null)
|
||
{
|
||
var similarity = CalculateSimilarity(face.Embedding, representativeFace.Embedding);
|
||
if (similarity >= BASE_SIMILARITY_THRESHOLD)
|
||
{
|
||
face.ClusterId = cluster.Id;
|
||
await dbContext.SaveChangesAsync();
|
||
return cluster;
|
||
}
|
||
}
|
||
}
|
||
}
|
||
|
||
// 创建新聚类
|
||
var newCluster = new FaceCluster
|
||
{
|
||
Name = $"未知人物 {DateTime.Now:yyyyMMddHHmmss}",
|
||
CreatedAt = DateTime.UtcNow
|
||
};
|
||
|
||
dbContext.FaceClusters.Add(newCluster);
|
||
await dbContext.SaveChangesAsync();
|
||
|
||
face.ClusterId = newCluster.Id;
|
||
await dbContext.SaveChangesAsync();
|
||
|
||
return newCluster;
|
||
}
|
||
|
||
public double CalculateSimilarity(float[] embedding1, float[] embedding2)
|
||
{
|
||
if (embedding1.Length != embedding2.Length) return 0;
|
||
|
||
// 1. 余弦相似度
|
||
double cosineSim = CalculateCosineSimilarity(embedding1, embedding2);
|
||
|
||
// 2. 欧几里得距离转换为相似度
|
||
double euclideanSim = CalculateEuclideanSimilarity(embedding1, embedding2);
|
||
|
||
// 3. 曼哈顿距离转换为相似度
|
||
double manhattanSim = CalculateManhattanSimilarity(embedding1, embedding2);
|
||
|
||
// 加权组合多个相似度指标
|
||
double weightedSimilarity = cosineSim * 0.6 + euclideanSim * 0.3 + manhattanSim * 0.1;
|
||
|
||
return weightedSimilarity;
|
||
}
|
||
|
||
private double CalculateCosineSimilarity(float[] embedding1, float[] embedding2)
|
||
{
|
||
double dot = 0, norm1 = 0, norm2 = 0;
|
||
|
||
for (int i = 0; i < embedding1.Length; i++)
|
||
{
|
||
dot += embedding1[i] * embedding2[i];
|
||
norm1 += embedding1[i] * embedding1[i];
|
||
norm2 += embedding2[i] * embedding2[i];
|
||
}
|
||
|
||
if (norm1 == 0 || norm2 == 0) return 0;
|
||
return dot / (Math.Sqrt(norm1) * Math.Sqrt(norm2));
|
||
}
|
||
|
||
private double CalculateEuclideanSimilarity(float[] embedding1, float[] embedding2)
|
||
{
|
||
double sumSquareDiff = 0;
|
||
for (int i = 0; i < embedding1.Length; i++)
|
||
{
|
||
double diff = embedding1[i] - embedding2[i];
|
||
sumSquareDiff += diff * diff;
|
||
}
|
||
|
||
double distance = Math.Sqrt(sumSquareDiff);
|
||
// 转换为相似度：距离越小，相似度越高
|
||
return 1.0 / (1.0 + distance);
|
||
}
|
||
|
||
private double CalculateManhattanSimilarity(float[] embedding1, float[] embedding2)
|
||
{
|
||
double sumAbsDiff = 0;
|
||
for (int i = 0; i < embedding1.Length; i++)
|
||
{
|
||
sumAbsDiff += Math.Abs(embedding1[i] - embedding2[i]);
|
||
}
|
||
|
||
// 转换为相似度
|
||
return 1.0 / (1.0 + sumAbsDiff / embedding1.Length);
|
||
}
|
||
|
||
private async Task<FaceCluster?> FindBestClusterAsync(Face face, List<FaceCluster> newClusters, MyDbContext dbContext)
|
||
{
|
||
if (face.Embedding == null) return null;
|
||
|
||
var clusterSimilarities = new List<(FaceCluster cluster, double avgSimilarity, double maxSimilarity, int comparisonCount)>();
|
||
|
||
// 检查现有数据库中的聚类
|
||
var existingClusters = await dbContext.FaceClusters
|
||
.Include(c => c.Faces.Take(MAX_COMPARISON_FACES))
|
||
.ToListAsync();
|
||
|
||
foreach (var cluster in existingClusters.Concat(newClusters))
|
||
{
|
||
if (cluster.Faces?.Any() == true)
|
||
{
|
||
var similarities = new List<double>();
|
||
|
||
foreach (var clusterFace in cluster.Faces.Take(MAX_COMPARISON_FACES))
|
||
{
|
||
if (clusterFace.Embedding != null)
|
||
{
|
||
var similarity = CalculateSimilarity(face.Embedding, clusterFace.Embedding);
|
||
similarities.Add(similarity);
|
||
}
|
||
}
|
||
|
||
if (similarities.Any())
|
||
{
|
||
double avgSimilarity = similarities.Average();
|
||
double maxSimilarity = similarities.Max();
|
||
|
||
clusterSimilarities.Add((cluster, avgSimilarity, maxSimilarity, similarities.Count));
|
||
}
|
||
}
|
||
}
|
||
|
||
// 智能选择最佳聚类
|
||
return SelectBestCluster(clusterSimilarities);
|
||
}
|
||
|
||
private FaceCluster? SelectBestCluster(List<(FaceCluster cluster, double avgSimilarity, double maxSimilarity, int comparisonCount)> clusterSimilarities)
|
||
{
|
||
if (!clusterSimilarities.Any()) return null;
|
||
|
||
// 按照综合评分排序
|
||
var rankedClusters = clusterSimilarities
|
||
.Where(cs => cs.avgSimilarity >= BASE_SIMILARITY_THRESHOLD || cs.maxSimilarity >= HIGH_CONFIDENCE_THRESHOLD)
|
||
.Select(cs => new
|
||
{
|
||
cs.cluster,
|
||
cs.avgSimilarity,
|
||
cs.maxSimilarity,
|
||
cs.comparisonCount,
|
||
// 综合评分：平均相似度权重60%，最高相似度权重30%，样本数量权重10%
|
||
Score = cs.avgSimilarity * 0.6 + cs.maxSimilarity * 0.3 +
|
||
Math.Min(cs.comparisonCount / (double)MAX_COMPARISON_FACES, 1.0) * 0.1
|
||
})
|
||
.OrderByDescending(x => x.Score)
|
||
.ToList();
|
||
|
||
if (!rankedClusters.Any()) return null;
|
||
|
||
var bestMatch = rankedClusters.First();
|
||
|
||
// 额外验证：如果最高相似度很高，直接接受
|
||
if (bestMatch.maxSimilarity >= HIGH_CONFIDENCE_THRESHOLD)
|
||
{
|
||
return bestMatch.cluster;
|
||
}
|
||
|
||
// 如果平均相似度足够高且有足够样本，接受
|
||
if (bestMatch.avgSimilarity >= BASE_SIMILARITY_THRESHOLD && bestMatch.comparisonCount >= 2)
|
||
{
|
||
return bestMatch.cluster;
|
||
}
|
||
|
||
return null;
|
||
}
|
||
|
||
public async Task<List<FaceCluster>> ClusterUserFacesAsync(int userId)
|
||
{
|
||
await using var dbContext = await contextFactory.CreateDbContextAsync();
|
||
|
||
// 获取指定用户所有有嵌入向量但未分类的人脸
|
||
var unclusteredFaces = await dbContext.Faces
|
||
.Where(f => f.Embedding != null && f.ClusterId == null && f.Picture.UserId == userId)
|
||
.Include(f => f.Picture)
|
||
.ToListAsync();
|
||
|
||
var clusters = new List<FaceCluster>();
|
||
|
||
foreach (var face in unclusteredFaces)
|
||
{
|
||
var assignedCluster = await FindBestUserClusterAsync(face, userId, clusters, dbContext);
|
||
|
||
if (assignedCluster != null)
|
||
{
|
||
face.ClusterId = assignedCluster.Id;
|
||
}
|
||
else
|
||
{
|
||
var newCluster = new FaceCluster
|
||
{
|
||
Name = $"未知人物 {DateTime.Now:yyyyMMddHHmmss}",
|
||
CreatedAt = DateTime.UtcNow
|
||
};
|
||
|
||
dbContext.FaceClusters.Add(newCluster);
|
||
await dbContext.SaveChangesAsync();
|
||
|
||
face.ClusterId = newCluster.Id;
|
||
clusters.Add(newCluster);
|
||
}
|
||
}
|
||
|
||
await dbContext.SaveChangesAsync();
|
||
logger.LogInformation("用户 {UserId} 人脸聚类完成，共处理 {FaceCount} 个人脸，生成 {ClusterCount} 个聚类",
|
||
userId, unclusteredFaces.Count, clusters.Count);
|
||
|
||
return clusters;
|
||
}
|
||
|
||
private async Task<FaceCluster?> FindBestUserClusterAsync(Face face, int userId, List<FaceCluster> newClusters, MyDbContext dbContext)
|
||
{
|
||
if (face.Embedding == null) return null;
|
||
|
||
var clusterSimilarities = new List<(FaceCluster cluster, double avgSimilarity, double maxSimilarity, int comparisonCount)>();
|
||
|
||
// 检查该用户现有的聚类
|
||
var existingClusters = await dbContext.FaceClusters
|
||
.Where(c => dbContext.Faces.Any(f => f.ClusterId == c.Id && f.Picture.UserId == userId))
|
||
.Include(c => c.Faces.Where(f => f.Picture.UserId == userId).Take(MAX_COMPARISON_FACES))
|
||
.ToListAsync();
|
||
|
||
foreach (var cluster in existingClusters.Concat(newClusters))
|
||
{
|
||
if (cluster.Faces?.Any() == true)
|
||
{
|
||
var similarities = new List<double>();
|
||
|
||
foreach (var clusterFace in cluster.Faces.Take(MAX_COMPARISON_FACES))
|
||
{
|
||
if (clusterFace.Embedding != null)
|
||
{
|
||
var similarity = CalculateSimilarity(face.Embedding, clusterFace.Embedding);
|
||
similarities.Add(similarity);
|
||
}
|
||
}
|
||
|
||
if (similarities.Any())
|
||
{
|
||
double avgSimilarity = similarities.Average();
|
||
double maxSimilarity = similarities.Max();
|
||
|
||
clusterSimilarities.Add((cluster, avgSimilarity, maxSimilarity, similarities.Count));
|
||
}
|
||
}
|
||
}
|
||
|
||
return SelectBestCluster(clusterSimilarities);
|
||
}
|
||
|
||
// 新增：聚类质量评估方法
|
||
public async Task<ClusterQualityMetrics> EvaluateClusterQualityAsync(int clusterId)
|
||
{
|
||
await using var dbContext = await contextFactory.CreateDbContextAsync();
|
||
|
||
var cluster = await dbContext.FaceClusters
|
||
.Include(c => c.Faces.Where(f => f.Embedding != null))
|
||
.FirstOrDefaultAsync(c => c.Id == clusterId);
|
||
|
||
if (cluster?.Faces == null || !cluster.Faces.Any())
|
||
{
|
||
return new ClusterQualityMetrics { IsValid = false };
|
||
}
|
||
|
||
var embeddings = cluster.Faces.Select(f => f.Embedding).Where(e => e != null).ToArray();
|
||
if (embeddings.Length < 2)
|
||
{
|
||
return new ClusterQualityMetrics { IsValid = true, InternalSimilarity = 1.0, FaceCount = embeddings.Length };
|
||
}
|
||
|
||
// 计算内部相似度
|
||
var similarities = new List<double>();
|
||
for (int i = 0; i < embeddings.Length; i++)
|
||
{
|
||
for (int j = i + 1; j < embeddings.Length; j++)
|
||
{
|
||
similarities.Add(CalculateSimilarity(embeddings[i]!, embeddings[j]!));
|
||
}
|
||
}
|
||
|
||
return new ClusterQualityMetrics
|
||
{
|
||
IsValid = true,
|
||
InternalSimilarity = similarities.Average(),
|
||
MinSimilarity = similarities.Min(),
|
||
MaxSimilarity = similarities.Max(),
|
||
FaceCount = embeddings.Length,
|
||
SimilarityStandardDeviation = CalculateStandardDeviation(similarities)
|
||
};
|
||
}
|
||
|
||
private double CalculateStandardDeviation(List<double> values)
|
||
{
|
||
if (!values.Any()) return 0;
|
||
|
||
double mean = values.Average();
|
||
double sumSquaredDifferences = values.Sum(v => Math.Pow(v - mean, 2));
|
||
return Math.Sqrt(sumSquaredDifferences / values.Count);
|
||
}
|
||
|
||
// 新增：动态阈值调整
|
||
public async Task<double> CalculateOptimalThresholdAsync(int userId)
|
||
{
|
||
await using var dbContext = await contextFactory.CreateDbContextAsync();
|
||
|
||
var userClusters = await dbContext.FaceClusters
|
||
.Where(c => dbContext.Faces.Any(f => f.ClusterId == c.Id && f.Picture.UserId == userId))
|
||
.Include(c => c.Faces.Where(f => f.Picture.UserId == userId && f.Embedding != null))
|
||
.ToListAsync();
|
||
|
||
var intraClusterSimilarities = new List<double>();
|
||
var interClusterSimilarities = new List<double>();
|
||
|
||
// 计算聚类内相似度
|
||
foreach (var cluster in userClusters.Where(c => c.Faces.Count > 1))
|
||
{
|
||
var faces = cluster.Faces.Where(f => f.Embedding != null).ToArray();
|
||
for (int i = 0; i < faces.Length; i++)
|
||
{
|
||
for (int j = i + 1; j < faces.Length; j++)
|
||
{
|
||
intraClusterSimilarities.Add(CalculateSimilarity(faces[i].Embedding!, faces[j].Embedding!));
|
||
}
|
||
}
|
||
}
|
||
|
||
// 计算聚类间相似度
|
||
for (int i = 0; i < userClusters.Count; i++)
|
||
{
|
||
for (int j = i + 1; j < userClusters.Count; j++)
|
||
{
|
||
var cluster1Faces = userClusters[i].Faces.Where(f => f.Embedding != null).Take(5).ToArray();
|
||
var cluster2Faces = userClusters[j].Faces.Where(f => f.Embedding != null).Take(5).ToArray();
|
||
|
||
foreach (var face1 in cluster1Faces)
|
||
{
|
||
foreach (var face2 in cluster2Faces)
|
||
{
|
||
interClusterSimilarities.Add(CalculateSimilarity(face1.Embedding!, face2.Embedding!));
|
||
}
|
||
}
|
||
}
|
||
}
|
||
|
||
if (!intraClusterSimilarities.Any() || !interClusterSimilarities.Any())
|
||
{
|
||
return BASE_SIMILARITY_THRESHOLD;
|
||
}
|
||
|
||
// 找到最优分割点
|
||
double minIntra = intraClusterSimilarities.Min();
|
||
double maxInter = interClusterSimilarities.Max();
|
||
|
||
// 理想阈值应该在聚类间最大相似度和聚类内最小相似度之间
|
||
double optimalThreshold = (minIntra + maxInter) / 2.0;
|
||
|
||
// 确保在合理范围内
|
||
return Math.Max(0.4, Math.Min(0.9, optimalThreshold));
|
||
}
|
||
}
|
||
|
||
// 新增：聚类质量评估结果类
|
||
public class ClusterQualityMetrics
|
||
{
|
||
public bool IsValid { get; set; }
|
||
public double InternalSimilarity { get; set; }
|
||
public double MinSimilarity { get; set; }
|
||
public double MaxSimilarity { get; set; }
|
||
public int FaceCount { get; set; }
|
||
public double SimilarityStandardDeviation { get; set; }
|
||
} |