change model with a sc model

app/src/main/assets/README.md

@@ -1,4 +1,5 @@
-Place a FaceNet-compatible TensorFlow Lite model here as facenet.tflite.
+Place a face recognition TensorFlow Lite model here as mobilefacenet.tflite
+(MobileFaceNet@WebFace600K, 112x112 input, 512-d output).
 
-Expected input: [1, 160, 160, 3].
-Expected output: a float embedding vector.
+Expected input: [1, 112, 112, 3] (float32, RGB, values normalized to [-1, 1]).
+Expected output: a 512-d L2-normalized float embedding vector.

app/src/main/java/com/example/studentfaceregistry/MainActivity.kt

@@ -86,6 +86,9 @@ class MainActivity : AppCompatActivity() {
     private val recognitionCache = mutableMapOf<Int, CachedRecognition>()
     private var lastVisibleDetections: List<DetectionUi> = emptyList()
     private var lastVisibleDetectionsAt = 0L
+    private var perfFrameCount = 0
+    private var perfDetectMsSum = 0L
+    private var perfRefreshMsSum = 0L
 
     private val permissionLauncher = registerForActivityResult(
         ActivityResultContracts.RequestMultiplePermissions()
@@ -285,6 +288,10 @@ class MainActivity : AppCompatActivity() {
                 text = "保存识别参数"
                 setOnClickListener { saveMatcherConfig() }
             })
+            matcherConfigPanel.addView(Button(this@MainActivity).apply {
+                text = "自动校准阈值"
+                setOnClickListener { applySuggestedThresholds() }
+            })
 
             addView(matcherConfigPanel)
 
@@ -431,7 +438,39 @@ class MainActivity : AppCompatActivity() {
     private fun startRecognitionSession() {
         if (currentMode != AppMode.RECOGNITION) return
 
-        val faceProcessor = ensureProcessor() ?: return
+        recognitionStatusText.text = "正在加载模型..."
+        lifecycleScope.launch(Dispatchers.IO) {
+            val existing = processor
+            if (existing != null) {
+                withContext(Dispatchers.Main) {
+                    if (currentMode == AppMode.RECOGNITION) bindRecognitionCamera(existing)
+                }
+                return@launch
+            }
+
+            val created = runCatching {
+                FaceProcessor(this@MainActivity).also { it.warmUp() }
+            }
+            withContext(Dispatchers.Main) {
+                if (currentMode != AppMode.RECOGNITION) {
+                    created.getOrNull()?.close()
+                    return@withContext
+                }
+                val faceProcessor = created.getOrNull()
+                if (faceProcessor == null) {
+                    val error = created.exceptionOrNull()
+                    val message = error?.message ?: error?.javaClass?.simpleName ?: "unknown error"
+                    recognitionStatusText.text = "模型初始化失败：$message"
+                    toast("模型初始化失败：$message")
+                    return@withContext
+                }
+                processor = faceProcessor
+                bindRecognitionCamera(faceProcessor)
+            }
+        }
+    }
+
+    private fun bindRecognitionCamera(faceProcessor: FaceProcessor) {
         previewView.post {
             val providerFuture = ProcessCameraProvider.getInstance(this)
             providerFuture.addListener({
@@ -515,7 +554,9 @@ class MainActivity : AppCompatActivity() {
             try {
                 val mediaImage = image.image ?: return@launch
                 val inputImage = InputImage.fromMediaImage(mediaImage, image.imageInfo.rotationDegrees)
+                val detectStart = SystemClock.elapsedRealtime()
                 val faces = faceProcessor.detectFaces(inputImage)
+                val detectMs = SystemClock.elapsedRealtime() - detectStart
 
                 if (faces.isEmpty()) {
                     withContext(Dispatchers.Main) {
@@ -543,11 +584,27 @@ class MainActivity : AppCompatActivity() {
                 val refreshedResults = mutableMapOf<Int, RecognitionResult>()
 
                 if (facesNeedingRefresh.isNotEmpty()) {
+                    val refreshStart = SystemClock.elapsedRealtime()
                     val bitmap = ImageUtils.imageProxyToBitmap(image)
                     facesNeedingRefresh.forEach { (index, face) ->
                         val refreshed = refreshRecognition(faceProcessor, bitmap, face)
                         refreshedResults[index] = refreshed
                     }
+                    val refreshMs = SystemClock.elapsedRealtime() - refreshStart
+                    perfFrameCount++
+                    perfDetectMsSum += detectMs
+                    perfRefreshMsSum += refreshMs
+                    if (perfFrameCount >= 60) {
+                        Log.i(
+                            "MainActivity",
+                            "FacePerf 检测平均=${perfDetectMsSum / perfFrameCount}ms " +
+                                "刷新推理平均=${perfRefreshMsSum / perfFrameCount}ms " +
+                                "人脸数=${faces.size}"
+                        )
+                        perfFrameCount = 0
+                        perfDetectMsSum = 0L
+                        perfRefreshMsSum = 0L
+                    }
                 }
 
                 val rawDetections = faces.mapIndexed { index, face ->
@@ -655,23 +712,6 @@ class MainActivity : AppCompatActivity() {
         uploadUrlText.text = "-"
     }
 
-    private fun ensureProcessor(): FaceProcessor? {
-        val current = processor
-        if (current != null) return current
-
-        val created = runCatching {
-            FaceProcessor(this).also { it.warmUp() }
-        }
-            .onFailure {
-                val message = it.message ?: "unknown error"
-                recognitionStatusText.text = "模型初始化失败：$message"
-                toast("模型初始化失败：$message")
-            }
-            .getOrNull()
-        processor = created
-        return created
-    }
-
     private fun hasCameraPermission(): Boolean {
         return ContextCompat.checkSelfPermission(this, Manifest.permission.CAMERA) == PackageManager.PERMISSION_GRANTED
     }
@@ -722,6 +762,20 @@ class MainActivity : AppCompatActivity() {
         return "当前参数：欧氏 <= ${formatThreshold(euclideanThreshold)}，余弦 >= ${formatThreshold(cosineThreshold)}"
     }
 
+    private fun applySuggestedThresholds() {
+        if (students.size < 3) {
+            toast("至少需要 3 名已注册学生才能校准阈值")
+            return
+        }
+        val suggested = matcher.suggestThresholds(students)
+        euclideanThresholdInput.setText(formatThreshold(suggested.euclidean))
+        cosineThresholdInput.setText(formatThreshold(suggested.cosine))
+        matcherConfigSummaryText.text =
+            "建议参数（基于 ${students.size} 名学生）：欧氏 <= ${formatThreshold(suggested.euclidean)}，" +
+                "余弦 >= ${formatThreshold(suggested.cosine)}，点击“保存识别参数”生效"
+        toast("已填入建议阈值，确认后点击保存")
+    }
+
     private fun formatThreshold(value: Float): String {
         return String.format(java.util.Locale.US, "%.2f", value)
     }
@@ -860,11 +914,11 @@ class MainActivity : AppCompatActivity() {
     ): RecognitionResult {
         val primary = matcher.findNearest(faceProcessor.embedFace(bitmap, face), students)
         val cached = face.trackingId?.let { recognitionCache[it] }
+        // 未匹配时只做单次推理；仅在疑似匹配或人脸明显移动时才用多裁剪升级，
+        // 避免对陌生面孔持续做 3 次推理
         val needsEscalation = cached == null ||
-            !cached.result.isMatched ||
-            primary.matchType != MatchType.MATCH ||
-            primary.confidence < cached.result.confidence - 0.05f ||
-            hasTrackMovedSignificantly(cached.lastBounds, face.boundingBox)
+            hasTrackMovedSignificantly(cached.lastBounds, face.boundingBox) ||
+            primary.matchType != MatchType.NO_MATCH
 
         if (!needsEscalation) {
             return primary
@@ -944,15 +998,15 @@ class MainActivity : AppCompatActivity() {
         private const val DEFAULT_EUCLIDEAN_THRESHOLD = 1.0f
         private const val DEFAULT_COSINE_THRESHOLD = 0.6f
         private const val MATCH_REFRESH_MS = 900L
-        private const val UNSURE_REFRESH_MS = 260L
-        private const val NO_MATCH_REFRESH_MS = 160L
+        private const val UNSURE_REFRESH_MS = 500L
+        private const val NO_MATCH_REFRESH_MS = 800L
         private const val TRACK_STALE_MS = 900L
         private const val TRACK_MOVE_THRESHOLD = 0.16f
         private const val TRACK_SIZE_THRESHOLD = 0.20f
         private const val MATCH_STABLE_CONFIDENCE = 0.82f
         private const val ANALYSIS_INTERVAL_MS = 80L
         private const val DETECTION_GRACE_MS = 700L
-        private val ANALYSIS_SIZE = Size(640, 480)
+        private val ANALYSIS_SIZE = Size(1280, 720)
     }
 }
 

app/src/main/java/com/example/studentfaceregistry/data/Student.kt

@@ -6,5 +6,6 @@ data class Student(
     val name: String,
     val photoUri: String,
     val embedding: FloatArray,
+    val extraEmbeddings: List<FloatArray> = emptyList(),
     val createdAt: Long = System.currentTimeMillis()
 )

app/src/main/java/com/example/studentfaceregistry/data/StudentRepository.kt

@@ -39,6 +39,7 @@ class StudentRepository(context: Context) {
                         name = obj.getString("name"),
                         photoUri = obj.getString("photoUri"),
                         embedding = decodeEmbedding(obj.getJSONArray("embedding")),
+                        extraEmbeddings = decodeEmbeddings(obj.optJSONArray("extraEmbeddings")),
                         createdAt = obj.optLong("createdAt", System.currentTimeMillis())
                     )
                 )
@@ -57,6 +58,10 @@ class StudentRepository(context: Context) {
                     put("photoUri", student.photoUri)
                     put("createdAt", student.createdAt)
                     put("embedding", JSONArray(student.embedding.toList()))
+                    put(
+                        "extraEmbeddings",
+                        JSONArray(student.extraEmbeddings.map { JSONArray(it.toList()) })
+                    )
                 }
             )
         }
@@ -66,4 +71,13 @@ class StudentRepository(context: Context) {
     private fun decodeEmbedding(array: JSONArray): FloatArray {
         return FloatArray(array.length()) { index -> array.getDouble(index).toFloat() }
     }
+
+    private fun decodeEmbeddings(array: JSONArray?): List<FloatArray> {
+        if (array == null) return emptyList()
+        return buildList {
+            for (index in 0 until array.length()) {
+                add(decodeEmbedding(array.getJSONArray(index)))
+            }
+        }
+    }
 }

app/src/main/java/com/example/studentfaceregistry/face/FaceEmbedder.kt

@@ -125,7 +125,7 @@ class FaceEmbedder(context: Context, private val modelType: ModelType = ModelTyp
 
     companion object {
         private const val TAG = "FaceEmbedder"
-        const val MODEL_ASSET_NAME = "facenet.tflite"
+        const val MODEL_ASSET_NAME = "mobilefacenet.tflite"
     }
 
     private fun inferInputSize(floatCount: Int): Int {

app/src/main/java/com/example/studentfaceregistry/face/FaceMatcher.kt

@@ -28,7 +28,13 @@ class FaceMatcher(
         var secondNearestEuclidean = Float.MAX_VALUE
 
         for (student in students) {
-            val cosineSim = cosineSimilarity(embedding, student.embedding)
+            // 对该学生存储的所有特征（主特征 + 多角度补充特征）取最高相似度
+            var bestCosine = -Float.MAX_VALUE
+            for (candidate in student.allEmbeddings()) {
+                val cosineSim = cosineSimilarity(embedding, candidate)
+                if (cosineSim > bestCosine) bestCosine = cosineSim
+            }
+            val cosineSim = bestCosine
             val euclideanDist = euclideanDistanceFromCosine(cosineSim)
 
             if (euclideanDist < nearestEuclidean) {
@@ -51,26 +57,6 @@ class FaceMatcher(
         }
     }
 
-    /**
-     * 查找所有可能的匹配（返回多个候选）
-     */
-    fun findAllMatches(
-        embedding: FloatArray,
-        students: List<Student>,
-        topK: Int = 3
-    ): List<MatchCandidate> {
-        val candidates = students.map { student ->
-            val cosineSim = cosineSimilarity(embedding, student.embedding)
-            MatchCandidate(
-                student = student,
-                euclideanDistance = euclideanDistanceFromCosine(cosineSim),
-                cosineSimilarity = cosineSim
-            )
-        }
-
-        return candidates.sortedBy { it.euclideanDistance }.take(topK)
-    }
-
     /**
      * 计算欧氏距离
      */
@@ -84,6 +70,8 @@ class FaceMatcher(
         return sqrt(sum)
     }
 
+    private fun Student.allEmbeddings(): List<FloatArray> = listOf(embedding) + extraEmbeddings
+
     /**
      * 对已 L2 归一化向量，通过余弦相似度推导欧氏距离，避免重复遍历
      */
@@ -136,14 +124,17 @@ class FaceMatcher(
 
     /**
      * 根据一组距离自适应调整阈值
-     * 适用于首次导入大量数据后优化识别效果
+     * 基于已注册学生两两之间的（异类）距离分布给出保守建议：
+     * 阈值取异类距离分布的低分位，保证绝大多数异类对被正确拒绝。
+     * 注意：本方法没有同人距离数据，只能给出"上界"参考，
+     * 现场采集同人数据后可进一步放宽。
      */
     fun suggestThresholds(students: List<Student>): SuggestedThresholds {
         if (students.size < 3) {
             return SuggestedThresholds(euclideanThreshold, cosineThreshold)
         }
 
-        // 计算所有学生两两之间的距离（采样）
+        // 计算所有学生两两之间的距离（采样，上限 200 对）
         val distances = mutableListOf<Float>()
         val sampleSize = minOf(students.size, 50)
         val step = maxOf(1, students.size / sampleSize)
@@ -161,13 +152,15 @@ class FaceMatcher(
 
         distances.sort()
 
-        // 建议使用较小百分位作为阈值（假设同一个人多次录入的距离较小）
+        // 使用低分位异类距离作为欧氏阈值上界：仅约 10% 的异类对会低于该距离
         val p10Index = (distances.size * 0.1).toInt().coerceIn(0, distances.size - 1)
-        val p20Index = (distances.size * 0.2).toInt().coerceIn(0, distances.size - 1)
+        val p05Index = (distances.size * 0.05).toInt().coerceIn(0, distances.size - 1)
+        val suggestedEuclidean = distances[p10Index]
 
         return SuggestedThresholds(
-            euclidean = distances[p10Index],
-            cosine = 1f - distances[p20Index] * 0.5f
+            euclidean = suggestedEuclidean,
+            // 对 L2 归一化向量：cos = 1 - d^2 / 2，用更保守的 p05 距离推导余弦阈值
+            cosine = 1f - (distances[p05Index] * distances[p05Index]) / 2f
         )
     }
 }
@@ -182,15 +175,6 @@ enum class MatchType {
     NO_DATA     // 没有数据
 }
 
-/**
- * 匹配候选
- */
-data class MatchCandidate(
-    val student: Student,
-    val euclideanDistance: Float,
-    val cosineSimilarity: Float
-)
-
 /**
  * 建议的阈值
  */

app/src/main/java/com/example/studentfaceregistry/face/SmartEnrollment.kt

@@ -12,8 +12,10 @@ import com.google.mlkit.vision.face.FaceDetection
 import com.google.mlkit.vision.face.FaceDetectorOptions
 import com.google.mlkit.vision.face.FaceLandmark
 import kotlinx.coroutines.tasks.await
+import kotlin.math.abs
 import kotlin.math.PI
 import kotlin.math.atan2
+import kotlin.math.roundToInt
 import kotlin.math.sqrt
 
 /**
@@ -79,6 +81,28 @@ class SmartEnrollment(context: Context) : AutoCloseable {
                 faceInfo.faceArea
             }
 
+            // 质量检查：人脸大小
+            val faceBox = largestFace.face.boundingBox
+            if (faceBox.width() < MIN_ENROLL_FACE_PX || faceBox.height() < MIN_ENROLL_FACE_PX) {
+                return EnrollmentResult.Failure(
+                    "照片中的人脸太小（${faceBox.width()}px），请换更近或更高清的照片。"
+                )
+            }
+
+            // 质量检查：角度
+            val yaw = yawAngle(largestFace.face)
+            val pitch = pitchAngle(largestFace.face)
+            if (abs(yaw) > MAX_ENROLL_YAW || abs(pitch) > MAX_ENROLL_PITCH) {
+                return EnrollmentResult.Failure(
+                    "请使用正面人脸照片（当前左右偏转 ${yaw.roundToInt()}°、上下 ${pitch.roundToInt()}°）。"
+                )
+            }
+
+            // 质量检查：清晰度
+            if (!isSharpEnough(cropFace(bitmap, largestFace.cropRect))) {
+                return EnrollmentResult.Failure("照片模糊，请换更清晰的照片。")
+            }
+
             val embeddings = mutableListOf<FloatArray>()
             for (paddingRatio in IMAGE_PADDING_RATIOS) {
                 val cropRect = safeCropRect(bitmap, largestFace.face.boundingBox, paddingRatio) ?: continue
@@ -131,6 +155,9 @@ class SmartEnrollment(context: Context) : AutoCloseable {
                     faceInfo.faceArea
                 }
 
+                if (largestFace.face.boundingBox.width() < MIN_ENROLL_FACE_PX) return@mapNotNull null
+                if (!isSharpEnough(cropFace(bitmap, largestFace.cropRect))) return@mapNotNull null
+
                 FaceFrameInfo(bitmap, largestFace.face, largestFace.cropRect, largestFace.faceArea)
             }
 
@@ -147,15 +174,15 @@ class SmartEnrollment(context: Context) : AutoCloseable {
             val selectedFrames = selectBestFramesFromGroups(groupedFrames, maxFrames)
 
             // 提取每帧的特征
-            val embeddings = selectedFrames.map { frameInfo ->
-                val cropped = cropFace(frameInfo.bitmap, frameInfo.cropRect)
-                val aligned = alignFace(cropped, frameInfo.face, frameInfo.cropRect)
-                val preprocessed = preprocessor.preprocess(aligned)
-                embedder.embed(preprocessed)
+            val frameEmbeddings = selectedFrames.map { frameInfo ->
+                frameInfo to embedFrame(frameInfo)
             }
 
-            // 融合特征
-            val fusedEmbedding = fuseEmbeddings(embeddings)
+            // 按偏航角分组：正脸融合为主特征，左右侧脸融合为补充特征
+            val (primaryEmbeddings, extraEmbeddings) = splitByYaw(frameEmbeddings)
+            val fusedEmbedding = fuseEmbeddings(primaryEmbeddings)
+            // splitByYaw 返回的补充特征已经是每组融合后的向量
+            val fusedExtras = extraEmbeddings
 
             // 计算角度覆盖
             val yawRange = calculateAngleRange(selectedFrames, { f -> yawAngle(f.face) })
@@ -166,7 +193,8 @@ class SmartEnrollment(context: Context) : AutoCloseable {
                 sourceType = EnrollmentSourceType.VIDEO,
                 frameCount = selectedFrames.size,
                 yawRange = yawRange,
-                pitchRange = pitchRange
+                pitchRange = pitchRange,
+                extraEmbeddings = fusedExtras
             )
         } catch (e: Exception) {
             EnrollmentResult.Failure("视频处理失败：${e.message}")
@@ -337,6 +365,82 @@ class SmartEnrollment(context: Context) : AutoCloseable {
         return Bitmap.createBitmap(bitmap, box.left, box.top, box.width(), box.height())
     }
 
+    /**
+     * 对单帧提取特征（裁剪 + 对齐 + 预处理 + 推理）
+     */
+    private fun embedFrame(frameInfo: FaceFrameInfo): FloatArray {
+        val cropped = cropFace(frameInfo.bitmap, frameInfo.cropRect)
+        val aligned = alignFace(cropped, frameInfo.face, frameInfo.cropRect)
+        val preprocessed = preprocessor.preprocess(aligned)
+        return embedder.embed(preprocessed)
+    }
+
+    /**
+     * 按偏航角将帧特征分组：正脸（主特征）、左右侧脸（补充特征）
+     */
+    private fun splitByYaw(
+        frames: List<Pair<FaceFrameInfo, FloatArray>>
+    ): Pair<List<FloatArray>, List<FloatArray>> {
+        val front = mutableListOf<FloatArray>()
+        val left = mutableListOf<FloatArray>()
+        val right = mutableListOf<FloatArray>()
+
+        for ((frameInfo, embedding) in frames) {
+            val yaw = yawAngle(frameInfo.face)
+            when {
+                yaw < -10f -> left.add(embedding)
+                yaw > 10f -> right.add(embedding)
+                else -> front.add(embedding)
+            }
+        }
+
+        val extras = mutableListOf<FloatArray>()
+        if (left.size >= 2) extras.add(fuseEmbeddings(left))
+        if (right.size >= 2) extras.add(fuseEmbeddings(right))
+        val primary = if (front.isNotEmpty()) front else frames.map { it.second }
+        return primary to extras
+    }
+
+    /**
+     * 基于 Laplacian 方差判断人脸区域是否清晰
+     */
+    private fun isSharpEnough(bitmap: Bitmap, threshold: Double = BLUR_VARIANCE_THRESHOLD): Boolean {
+        val width = bitmap.width
+        val height = bitmap.height
+        if (width < 3 || height < 3) return false
+
+        val pixels = IntArray(width * height)
+        bitmap.getPixels(pixels, 0, width, 0, 0, width, height)
+
+        var sum = 0.0
+        var sumSquares = 0.0
+        var count = 0L
+        for (y in 1 until height - 1) {
+            var index = y * width + 1
+            for (x in 1 until width - 1) {
+                val lap = 4 * luma(pixels[index]) -
+                    luma(pixels[index - 1]) - luma(pixels[index + 1]) -
+                    luma(pixels[index - width]) - luma(pixels[index + width])
+                sum += lap
+                sumSquares += lap * lap
+                count++
+                index++
+            }
+        }
+        if (count == 0L) return false
+
+        val mean = sum / count
+        val variance = sumSquares / count - mean * mean
+        return variance >= threshold
+    }
+
+    private fun luma(pixel: Int): Int {
+        val r = (pixel shr 16) and 0xFF
+        val g = (pixel shr 8) and 0xFF
+        val b = pixel and 0xFF
+        return (r * 299 + g * 587 + b * 114) / 1000
+    }
+
     private suspend fun detectUsableFaces(bitmap: Bitmap): List<DetectedFace> {
         return detector.process(InputImage.fromBitmap(bitmap, 0)).await()
             .mapNotNull { face ->
@@ -430,6 +534,10 @@ class SmartEnrollment(context: Context) : AutoCloseable {
         private val VIDEO_EXTENSIONS = setOf(".mp4", ".avi", ".mov", ".wmv", ".flv", ".mkv", ".webm", ".3gp")
         private val VIDEO_MIME_TYPES = setOf("application/mp4")
         private val IMAGE_PADDING_RATIOS = floatArrayOf(0.18f, 0.25f, 0.34f)
+        private const val MIN_ENROLL_FACE_PX = 96
+        private const val MAX_ENROLL_YAW = 30f
+        private const val MAX_ENROLL_PITCH = 15f
+        private const val BLUR_VARIANCE_THRESHOLD = 40.0
     }
 }
 
@@ -476,7 +584,8 @@ sealed class EnrollmentResult {
         val sourceType: EnrollmentSourceType = EnrollmentSourceType.IMAGE,
         val frameCount: Int = 1,
         val yawRange: Float = 0f,
-        val pitchRange: Float = 0f
+        val pitchRange: Float = 0f,
+        val extraEmbeddings: List<FloatArray> = emptyList()
     ) : EnrollmentResult() {
         override fun equals(other: Any?): Boolean {
             if (this === other) return true
@@ -489,6 +598,10 @@ sealed class EnrollmentResult {
             if (pitchRange != other.pitchRange) return false
             if (!embedding.contentEquals(other.embedding)) return false
             if (sourceType != other.sourceType) return false
+            if (extraEmbeddings.size != other.extraEmbeddings.size) return false
+            if (extraEmbeddings.indices.any {
+                    !extraEmbeddings[it].contentEquals(other.extraEmbeddings[it])
+                }) return false
 
             return true
         }
@@ -499,6 +612,7 @@ sealed class EnrollmentResult {
             result = 31 * result + pitchRange.hashCode()
             result = 31 * result + embedding.contentHashCode()
             result = 31 * result + sourceType.hashCode()
+            result = 31 * result + extraEmbeddings.sumOf { it.contentHashCode() }
             return result
         }
     }

app/src/main/java/com/example/studentfaceregistry/face/VideoEnrollment.kt

@@ -1,388 +0,0 @@
-package com.example.studentfaceregistry.face
-
-import android.graphics.Bitmap
-import android.graphics.Rect
-import com.google.mlkit.vision.common.InputImage
-import com.google.mlkit.vision.face.Face
-import com.google.mlkit.vision.face.FaceDetection
-import com.google.mlkit.vision.face.FaceDetectorOptions
-import com.google.mlkit.vision.face.FaceLandmark
-import kotlinx.coroutines.tasks.await
-import kotlin.math.PI
-import kotlin.math.abs
-import kotlin.math.atan2
-import kotlin.math.cos
-import kotlin.math.sin
-import kotlin.math.sqrt
-
-/**
- * 视频人脸注册器
- * 从视频中提取多角度人脸帧，融合生成更鲁棒的特征向量
- *
- * 使用场景：
- * - 学生录制视频时转头、点头
- * - 自动选择质量好的帧
- * - 融合多帧特征，提升识别准确率
- */
-class VideoEnrollment(private val context: android.content.Context) : AutoCloseable {
-
-    private val detector = FaceDetection.getClient(
-        FaceDetectorOptions.Builder()
-            .setPerformanceMode(FaceDetectorOptions.PERFORMANCE_MODE_ACCURATE)
-            .setLandmarkMode(FaceDetectorOptions.LANDMARK_MODE_ALL)
-            .setClassificationMode(FaceDetectorOptions.CLASSIFICATION_MODE_ALL)
-            .build()
-    )
-
-    private val embedder = FaceEmbedder(context, FaceEmbedder.ModelType.ARCFACE)
-
-    /**
-     * 从视频帧序列中提取人脸特征
-     * @param frames 视频帧列表（Bitmap）
-     * @param minFrames 最少需要的人脸帧数
-     * @param maxFrames 最多使用的人脸帧数（用于特征融合）
-     * @return 融合后的特征向量
-     */
-    suspend fun enrollFromFrames(
-        frames: List<Bitmap>,
-        minFrames: Int = 5,
-        maxFrames: Int = 20
-    ): VideoEnrollmentResult {
-        val faceFrames = mutableListOf<FaceFrame>()
-
-        // 检测所有帧中的人脸
-        for ((index, frame) in frames.withIndex()) {
-            val faces = detector.process(InputImage.fromBitmap(frame, 0)).await()
-            for (face in faces) {
-                val quality = calculateFaceQuality(face)
-                if (quality >= MIN_QUALITY_THRESHOLD) {
-                    val yaw = calculateYawAngle(face)
-                    val pitch = calculatePitchAngle(face)
-                    faceFrames.add(
-                        FaceFrame(
-                            bitmap = frame,
-                            face = face,
-                            quality = quality,
-                            yaw = yaw,
-                            pitch = pitch,
-                            frameIndex = index
-                        )
-                    )
-                }
-            }
-        }
-
-        if (faceFrames.size < minFrames) {
-            return VideoEnrollmentResult.Failure(
-                reason = "检测到 ${faceFrames.size} 帧有效人脸，需要至少 $minFrames 帧。请确保视频中人脸清晰且有多角度展示。"
-            )
-        }
-
-        // 按角度分组，确保多角度覆盖
-        val groupedFrames = groupFramesByAngle(faceFrames)
-
-        // 从每组选择质量最好的帧
-        val selectedFrames = selectBestFramesFromGroups(groupedFrames, maxFrames)
-
-        // 提取每帧的特征
-        val embeddings = selectedFrames.map { frame ->
-            extractSingleEmbedding(frame)
-        }
-
-        // 融合特征（平均 + L2 归一化）
-        val fusedEmbedding = fuseEmbeddings(embeddings)
-
-        return VideoEnrollmentResult.Success(
-            embedding = fusedEmbedding,
-            frameCount = selectedFrames.size,
-            totalDetected = faceFrames.size,
-            angleCoverage = calculateAngleCoverage(selectedFrames)
-        )
-    }
-
-    /**
-     * 计算人脸质量分数（0-1）
-     */
-    private fun calculateFaceQuality(face: Face): Float {
-        var score = 1.0f
-
-        // 可见度分数
-        val leftEyeVisible = face.getLandmark(FaceLandmark.LEFT_EYE) != null
-        val rightEyeVisible = face.getLandmark(FaceLandmark.RIGHT_EYE) != null
-        if (!leftEyeVisible || !rightEyeVisible) score *= 0.7f
-
-        // 置信度分数
-        val confidence = face.trackingId?.let { 1.0f } ?: 0.9f
-        score *= confidence
-
-        // 人脸大小分数（越大越好）
-        val faceArea = face.boundingBox.width() * face.boundingBox.height()
-        val sizeScore = (faceArea / 100000f).coerceIn(0.5f, 1.0f)
-        score *= sizeScore
-
-        return score
-    }
-
-    /**
-     * 计算人脸偏航角（左右转头，-90 到 90 度）
-     */
-    private fun calculateYawAngle(face: Face): Float {
-        val leftCheek = face.getLandmark(FaceLandmark.LEFT_CHEEK) ?: return 0f
-        val rightCheek = face.getLandmark(FaceLandmark.RIGHT_CHEEK) ?: return 0f
-
-        val dx = rightCheek.position.x - leftCheek.position.x
-        val dy = rightCheek.position.y - leftCheek.position.y
-
-        // 计算角度
-        val angle = atan2(dy.toDouble(), dx.toDouble()).toFloat() * 180f / PI.toFloat()
-
-        return angle.coerceIn(-90f, 90f)
-    }
-
-    /**
-     * 计算人脸俯仰角（上下点头，-90 到 90 度）
-     */
-    private fun calculatePitchAngle(face: Face): Float {
-        val nose = face.getLandmark(FaceLandmark.NOSE_BASE) ?: return 0f
-        val leftEye = face.getLandmark(FaceLandmark.LEFT_EYE) ?: return 0f
-        val rightEye = face.getLandmark(FaceLandmark.RIGHT_EYE) ?: return 0f
-
-        // 眼睛中心点
-        val eyeCenterY = (leftEye.position.y + rightEye.position.y) / 2f
-
-        val dx = nose.position.x - (leftEye.position.x + rightEye.position.x) / 2f
-        val dy = nose.position.y - eyeCenterY
-
-        val angle = atan2(dy.toDouble(), dx.toDouble()).toFloat() * 180f / PI.toFloat()
-
-        return angle.coerceIn(-90f, 90f)
-    }
-
-    /**
-     * 按角度分组人脸帧
-     */
-    private fun groupFramesByAngle(frames: List<FaceFrame>): Map<VideoAngleBucket, List<FaceFrame>> {
-        val buckets = mutableMapOf<VideoAngleBucket, MutableList<FaceFrame>>()
-
-        for (frame in frames) {
-            val bucket = getAngleBucket(frame.yaw, frame.pitch)
-            if (!buckets.containsKey(bucket)) {
-                buckets[bucket] = mutableListOf()
-            }
-            buckets[bucket]!!.add(frame)
-        }
-
-        return buckets
-    }
-
-    /**
-     * 获取角度分组
-     */
-    private fun getAngleBucket(yaw: Float, pitch: Float): VideoAngleBucket {
-        val yawBucket = when {
-            yaw < -30 -> -2  // 左
-            yaw < -10 -> -1  // 左前
-            yaw < 10 -> 0    // 正
-            yaw < 30 -> 1    // 右前
-            else -> 2        // 右
-        }
-
-        val pitchBucket = when {
-            pitch < -15 -> -1  // 上
-            pitch < 15 -> 0    // 中
-            else -> 1          // 下
-        }
-
-        return VideoAngleBucket(yawBucket, pitchBucket)
-    }
-
-    /**
-     * 从每组中选择质量最好的帧
-     */
-    private fun selectBestFramesFromGroups(
-        groups: Map<VideoAngleBucket, List<FaceFrame>>,
-        maxFrames: Int
-    ): List<FaceFrame> {
-        val selected = mutableListOf<FaceFrame>()
-
-        // 先按组排序（组内按质量降序）
-        val sortedGroups = groups.entries.sortedByDescending { entry ->
-            entry.value.maxOfOrNull { it.quality } ?: 0f
-        }
-
-        for ((_, frames) in sortedGroups) {
-            val sortedFrames = frames.sortedByDescending { it.quality }
-            for (frame in sortedFrames) {
-                if (selected.size >= maxFrames) break
-                // 避免选择同一帧多次
-                if (!selected.any { it.frameIndex == frame.frameIndex }) {
-                    selected.add(frame)
-                }
-            }
-        }
-
-        return selected
-    }
-
-    /**
-     * 提取单帧人脸特征
-     */
-    private fun extractSingleEmbedding(frame: FaceFrame): FloatArray {
-        val bitmap = frame.bitmap
-        val face = frame.face
-
-        // 裁剪人脸
-        val cropped = cropFace(bitmap, face.boundingBox)
-
-        // 对齐
-        val aligned = alignFace(cropped, face)
-
-        // 预处理
-        val preprocessor = ImagePreprocessor()
-        val preprocessed = preprocessor.preprocess(aligned)
-
-        // 提取特征
-        return embedder.embed(preprocessed)
-    }
-
-    /**
-     * 融合多个特征向量
-     */
-    private fun fuseEmbeddings(embeddings: List<FloatArray>): FloatArray {
-        if (embeddings.isEmpty()) throw IllegalArgumentException("embeddings is empty")
-        if (embeddings.size == 1) return embeddings[0]
-
-        val size = embeddings[0].size
-        val sum = FloatArray(size)
-
-        for (embedding in embeddings) {
-            for (i in 0 until size) {
-                sum[i] += embedding[i]
-            }
-        }
-
-        // 平均
-        for (i in 0 until size) {
-            sum[i] /= embeddings.size
-        }
-
-        // L2 归一化
-        return l2Normalize(sum)
-    }
-
-    /**
-     * 计算角度覆盖范围
-     */
-    private fun calculateAngleCoverage(frames: List<FaceFrame>): AngleCoverage {
-        var minYaw = Float.MAX_VALUE
-        var maxYaw = -Float.MAX_VALUE
-        var minPitch = Float.MAX_VALUE
-        var maxPitch = -Float.MAX_VALUE
-
-        for (frame in frames) {
-            minYaw = minOf(minYaw, frame.yaw)
-            maxYaw = maxOf(maxYaw, frame.yaw)
-            minPitch = minOf(minPitch, frame.pitch)
-            maxPitch = maxOf(maxPitch, frame.pitch)
-        }
-
-        return AngleCoverage(
-            yawRange = maxYaw - minYaw,
-            pitchRange = maxPitch - minPitch,
-            minYaw = minYaw,
-            maxYaw = maxYaw,
-            minPitch = minPitch,
-            maxPitch = maxPitch
-        )
-    }
-
-    /**
-     * 裁剪人脸
-     */
-    private fun cropFace(bitmap: Bitmap, box: Rect): Bitmap {
-        val padding = (maxOf(box.width(), box.height()) * 0.25f).toInt()
-        val left = (box.left - padding).coerceIn(0, bitmap.width)
-        val top = (box.top - padding).coerceIn(0, bitmap.height)
-        val right = (box.right + padding).coerceIn(0, bitmap.width)
-        val bottom = (box.bottom + padding).coerceIn(0, bitmap.height)
-        return Bitmap.createBitmap(bitmap, left, top, right - left, bottom - top)
-    }
-
-    /**
-     * 对齐人脸
-     */
-    private fun alignFace(bitmap: Bitmap, face: Face): Bitmap {
-        val leftEye = face.getLandmark(FaceLandmark.LEFT_EYE) ?: return bitmap
-        val rightEye = face.getLandmark(FaceLandmark.RIGHT_EYE) ?: return bitmap
-
-        val aligner = FaceAligner()
-        return aligner.align(
-            bitmap = bitmap,
-            leftEyeX = leftEye.position.x,
-            leftEyeY = leftEye.position.y,
-            rightEyeX = rightEye.position.x,
-            rightEyeY = rightEye.position.y,
-            faceLeft = 0,
-            faceTop = 0
-        )
-    }
-
-    private fun l2Normalize(values: FloatArray): FloatArray {
-        var sum = 0f
-        for (value in values) sum += value * value
-        val norm = sqrt(sum.coerceAtLeast(1e-12f))
-        return FloatArray(values.size) { values[it] / norm }
-    }
-
-    override fun close() {
-        detector.close()
-        embedder.close()
-    }
-
-    companion object {
-        private const val MIN_QUALITY_THRESHOLD = 0.5f
-    }
-}
-
-/**
- * 人脸帧信息
- */
-data class FaceFrame(
-    val bitmap: Bitmap,
-    val face: Face,
-    val quality: Float,
-    val yaw: Float,      // 偏航角（左右转头）
-    val pitch: Float,    // 俯仰角（上下点头）
-    val frameIndex: Int
-)
-
-/**
- * 角度分组
- */
-data class VideoAngleBucket(val yaw: Int, val pitch: Int)
-
-/**
- * 角度覆盖范围
- */
-data class AngleCoverage(
-    val yawRange: Float,
-    val pitchRange: Float,
-    val minYaw: Float,
-    val maxYaw: Float,
-    val minPitch: Float,
-    val maxPitch: Float
-)
-
-/**
- * 注册结果
- */
-sealed class VideoEnrollmentResult {
-    data class Success(
-        val embedding: FloatArray,
-        val frameCount: Int,
-        val totalDetected: Int,
-        val angleCoverage: AngleCoverage
-    ) : VideoEnrollmentResult()
-
-    data class Failure(val reason: String) : VideoEnrollmentResult()
-}

app/src/main/java/com/example/studentfaceregistry/upload/UploadServer.kt

@@ -226,7 +226,8 @@ class UploadServer(
                             studentNo = studentNo,
                             name = name,
                             photoUri = "upload://$fileName",
-                            embedding = result.embedding
+                            embedding = result.embedding,
+                            extraEmbeddings = result.extraEmbeddings
                         )
                     )
                     val sourceInfo = if (result.sourceType == com.example.studentfaceregistry.face.EnrollmentSourceType.VIDEO) {

convert_model.py

@@ -1,134 +1,62 @@
 #!/usr/bin/env python3
 """
-ArcFace ONNX to TFLite 转换脚本
+ONNX 人脸模型转 TFLite 脚本（基于 onnx2tf）
 
-使用方法:
-    1. 先安装依赖：pip install onnx onnxruntime tensorflow
-    2. 运行转换：python convert_model.py
+环境要求（推荐使用已有 onnx2tf 的 conda 环境）：
+    conda activate base   # 已安装 onnx2tf + ai-edge-litert
+    或 pip install onnx2tf ai-edge-litert onnx onnxruntime
+
+用法:
+    python convert_model.py <input.onnx> <output.tflite>
+
+示例:
+    python convert_model.py backup_models/buffalo_sc/w600k_mbf.onnx app/app/src/main/assets/mobilefacenet.tflite
 """
 
 import os
-import numpy as np
+import shutil
 import subprocess
 import sys
+import tempfile
 
-def install_requirements():
-    """安装必要的包"""
-    packages = ['onnx', 'onnxruntime', 'tensorflow', 'tf2onnx', 'onnx-tf']
-    print("检查并安装依赖包...")
-    for pkg in packages:
-        try:
-            __import__(pkg.replace('-', '_'))
-            print(f"  ✓ {pkg} 已安装")
-        except ImportError:
-            print(f"  安装 {pkg}...")
-            subprocess.check_call([sys.executable, '-m', 'pip', 'install', pkg, '-q'])
-    print("依赖包安装完成！\n")
 
-def download_model():
-    """下载 ArcFace 模型"""
-    model_url = "https://drive.google.com/uc?export=download&id=1gnt6P3jaiwfevV4hreWHPu0Mive5VRyP"
-    output_path = "/Users/liushuming/projects/app/app/src/main/assets/arcface_ir50_glint360k.onnx"
+def main() -> int:
+    if len(sys.argv) != 3:
+        print(__doc__)
+        return 1
 
-    print("下载 ArcFace 模型 (IResNet-50, Glint360K 数据集)...")
-    print(f"目标路径：{output_path}")
+    onnx_path = sys.argv[1]
+    tflite_path = sys.argv[2]
 
-    import urllib.request
-    urllib.request.urlretrieve(model_url, output_path)
-
-    size_mb = os.path.getsize(output_path) / 1024 / 1024
-    print(f"下载完成！模型大小：{size_mb:.2f} MB\n")
-
-    return output_path
-
-def convert_onnx_to_tflite(onnx_path, output_path):
-    """将 ONNX 模型转换为 TFLite"""
-    print(f"转换 ONNX -> TFLite...")
-    print(f"输入：{onnx_path}")
-    print(f"输出：{output_path}")
+    if not os.path.isfile(onnx_path):
+        print(f"错误：找不到输入模型 {onnx_path}")
+        return 1
 
+    work_dir = tempfile.mkdtemp(prefix="onnx2tf_")
     try:
-        import onnx
-        from onnx_tf.backend import prepare
+        print(f"转换 {onnx_path} -> {tflite_path}")
+        subprocess.check_call(
+            [sys.executable, "-m", "onnx2tf", "-i", onnx_path, "-o", work_dir]
+        )
 
-        # 加载 ONNX 模型
-        onnx_model = onnx.load(onnx_path)
-        onnx.checker.check_model(onnx_model)
+        base = os.path.splitext(os.path.basename(onnx_path))[0]
+        converted = os.path.join(work_dir, f"{base}_float32.tflite")
+        if not os.path.isfile(converted):
+            candidates = [
+                f for f in os.listdir(work_dir) if f.endswith("_float32.tflite")
+            ]
+            if not candidates:
+                print("错误：onnx2tf 未生成 float32 tflite，请查看上方日志")
+                return 1
+            converted = os.path.join(work_dir, candidates[0])
 
-        # 打印模型信息
-        inputs = onnx_model.graph.input
-        outputs = onnx_model.graph.output
-        print(f"\n模型输入：{inputs[0].name}, 形状：{[d.dim_value for d in inputs[0].type.tensor_type.shape.dim]}")
-        print(f"模型输出：{outputs[0].name}, 形状：{[d.dim_value for d in outputs[0].type.tensor_type.shape.dim]}")
+        os.makedirs(os.path.dirname(tflite_path) or ".", exist_ok=True)
+        shutil.copyfile(converted, tflite_path)
+        print(f"完成：{tflite_path}（{os.path.getsize(tflite_path) / 1024 / 1024:.2f} MB）")
+        return 0
+    finally:
+        shutil.rmtree(work_dir, ignore_errors=True)
 
-        # 转换为 TensorFlow
-        print("\n转换为 TensorFlow 格式...")
-        tf_backend = prepare(onnx_model)
-        tf_graph = tf_backend.tf_graph
-
-        # 保存为 SavedModel
-        import tensorflow as tf
-        saved_model_dir = output_path.replace('.tflite', '_saved_model')
-
-        # 清理已存在的目录
-        import shutil
-        if os.path.exists(saved_model_dir):
-            shutil.rmtree(saved_model_dir)
-
-        # 保存 graph
-        with tf.Graph().as_default() as graph:
-            tf.import_graph_def(tf_graph, name="")
-            tf.saved_model.save(tf.keras.models.Model(inputs=graph.get_tensor_by_name('input:0'),
-                                                      outputs=graph.get_tensor_by_name('output:0')),
-                               saved_model_dir)
-
-        # 转换为 TFLite
-        print("转换为 TFLite 格式...")
-        converter = tf.lite.TFLiteConverter.from_saved_model(saved_model_dir)
-        converter.optimizations = [tf.lite.Optimize.DEFAULT]
-
-        tflite_model = converter.convert()
-
-        with open(output_path, 'wb') as f:
-            f.write(tflite_model)
-
-        # 清理临时文件
-        shutil.rmtree(saved_model_dir)
-
-        size_mb = os.path.getsize(output_path) / 1024 / 1024
-        print(f"\n✓ 转换完成！")
-        print(f"  TFLite 模型：{output_path}")
-        print(f"  模型大小：{size_mb:.2f} MB")
-
-        return True
-
-    except ImportError as e:
-        print(f"\n错误：缺少依赖包 - {e}")
-        print("请运行：pip install onnx-tf tensorflow")
-        return False
-    except Exception as e:
-        print(f"\n转换错误：{e}")
-        return False
-
-def main():
-    print("=" * 60)
-    print("ArcFace 模型下载和转换工具")
-    print("=" * 60)
-
-    # 安装依赖
-    install_requirements()
-
-    # 下载模型
-    onnx_path = download_model()
-
-    # 转换为 TFLite
-    output_path = "/Users/liushuming/projects/app/app/src/main/assets/facenet.tflite"
-    convert_onnx_to_tflite(onnx_path, output_path)
-
-    print("\n" + "=" * 60)
-    print("完成！")
-    print("模型已保存到：app/src/main/assets/facenet.tflite")
-    print("=" * 60)
 
 if __name__ == "__main__":
-    main()
+    sys.exit(main())