改一个版

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

@@ -1,63 +1,72 @@
 package com.example.studentfaceregistry
 
 import android.graphics.Bitmap
-import android.graphics.BitmapFactory
-import android.graphics.ImageFormat
 import android.graphics.Matrix
-import android.graphics.Rect
-import android.graphics.YuvImage
 import android.media.Image
+import androidx.camera.core.ExperimentalGetImage
 import androidx.camera.core.ImageProxy
-import java.io.ByteArrayOutputStream
-import kotlin.math.min
 
+@ExperimentalGetImage
 object ImageUtils {
     fun imageProxyToBitmap(imageProxy: ImageProxy): Bitmap {
         val image = imageProxy.image ?: error("ImageProxy does not contain an image.")
-        val nv21 = yuv420ToNv21(image)
-        val yuvImage = YuvImage(nv21, ImageFormat.NV21, image.width, image.height, null)
-        val output = ByteArrayOutputStream()
-        yuvImage.compressToJpeg(Rect(0, 0, image.width, image.height), 90, output)
-        val bitmap = BitmapFactory.decodeByteArray(output.toByteArray(), 0, output.size())
+        val bitmap = yuv420ToBitmap(image)
+        if (imageProxy.imageInfo.rotationDegrees == 0) {
+            return bitmap
+        }
         val matrix = Matrix().apply {
             postRotate(imageProxy.imageInfo.rotationDegrees.toFloat())
         }
         return Bitmap.createBitmap(bitmap, 0, 0, bitmap.width, bitmap.height, matrix, true)
     }
 
-    private fun yuv420ToNv21(image: Image): ByteArray {
+    private fun yuv420ToBitmap(image: Image): Bitmap {
         val width = image.width
         val height = image.height
-        val ySize = width * height
-        val chromaSize = width * height / 4
-        val nv21 = ByteArray(ySize + chromaSize * 2)
-        copyPlane(image.planes[0], width, height, nv21, 0, 1)
-        copyPlane(image.planes[2], width / 2, height / 2, nv21, ySize, 2)
-        copyPlane(image.planes[1], width / 2, height / 2, nv21, ySize + 1, 2)
-        return nv21
+        val yPlane = image.planes[0]
+        val uPlane = image.planes[1]
+        val vPlane = image.planes[2]
+
+        val yBuffer = yPlane.buffer
+        val uBuffer = uPlane.buffer
+        val vBuffer = vPlane.buffer
+        val pixels = IntArray(width * height)
+
+        val yRowStride = yPlane.rowStride
+        val yPixelStride = yPlane.pixelStride
+        val uRowStride = uPlane.rowStride
+        val uPixelStride = uPlane.pixelStride
+        val vRowStride = vPlane.rowStride
+        val vPixelStride = vPlane.pixelStride
+
+        var pixelIndex = 0
+        for (row in 0 until height) {
+            val yRow = row * yRowStride
+            val uvRow = (row shr 1) * uRowStride
+            val vvRow = (row shr 1) * vRowStride
+            for (col in 0 until width) {
+                val y = (yBuffer.get(yRow + col * yPixelStride).toInt() and 0xFF) - 16
+                val u = (uBuffer.get(uvRow + (col shr 1) * uPixelStride).toInt() and 0xFF) - 128
+                val v = (vBuffer.get(vvRow + (col shr 1) * vPixelStride).toInt() and 0xFF) - 128
+
+                val y1192 = 1192 * y.coerceAtLeast(0)
+                var r = y1192 + 1634 * v
+                var g = y1192 - 833 * v - 400 * u
+                var b = y1192 + 2066 * u
+
+                r = r.coerceIn(0, 262143)
+                g = g.coerceIn(0, 262143)
+                b = b.coerceIn(0, 262143)
+
+                pixels[pixelIndex++] = -0x1000000 or
+                    ((r shl 6) and 0x00FF0000) or
+                    ((g shr 2) and 0x0000FF00) or
+                    ((b shr 10) and 0x000000FF)
+            }
         }
 
-    private fun copyPlane(
-        plane: Image.Plane,
-        width: Int,
-        height: Int,
-        output: ByteArray,
-        offset: Int,
-        outputPixelStride: Int
-    ) {
-        val buffer = plane.buffer
-        val row = ByteArray(plane.rowStride)
-        var outputIndex = offset
-        for (rowIndex in 0 until height) {
-            val bytesToRead = min(plane.rowStride, buffer.remaining())
-            buffer.get(row, 0, bytesToRead)
-            for (colIndex in 0 until width) {
-                val inputIndex = colIndex * plane.pixelStride
-                if (inputIndex < bytesToRead && outputIndex < output.size) {
-                    output[outputIndex] = row[inputIndex]
-                }
-                outputIndex += outputPixelStride
-            }
+        return Bitmap.createBitmap(width, height, Bitmap.Config.ARGB_8888).apply {
+            setPixels(pixels, 0, width, 0, 0, width, height)
         }
     }
 }

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

@@ -3,6 +3,7 @@ package com.example.studentfaceregistry
 import android.Manifest
 import android.content.SharedPreferences
 import android.content.pm.PackageManager
+import android.graphics.Bitmap
 import android.graphics.Rect
 import android.graphics.RectF
 import android.graphics.Typeface
@@ -22,6 +23,7 @@ import android.widget.Toast
 import androidx.activity.result.contract.ActivityResultContracts
 import androidx.appcompat.app.AppCompatActivity
 import androidx.camera.core.CameraSelector
+import androidx.camera.core.ExperimentalGetImage
 import androidx.camera.core.ImageAnalysis
 import androidx.camera.core.ImageProxy
 import androidx.camera.core.Preview
@@ -46,6 +48,7 @@ import kotlinx.coroutines.launch
 import kotlinx.coroutines.withContext
 import java.util.concurrent.Executors
 
+@ExperimentalGetImage
 class MainActivity : AppCompatActivity() {
     private lateinit var countText: TextView
     private lateinit var previewView: PreviewView
@@ -81,6 +84,8 @@ class MainActivity : AppCompatActivity() {
     @Volatile private var matcher = FaceMatcher(DEFAULT_EUCLIDEAN_THRESHOLD, DEFAULT_COSINE_THRESHOLD)
     private val recognitionHistory = mutableMapOf<Int, ArrayDeque<RecognitionResult>>()
     private val recognitionCache = mutableMapOf<Int, CachedRecognition>()
+    private var lastVisibleDetections: List<DetectionUi> = emptyList()
+    private var lastVisibleDetectionsAt = 0L
 
     private val permissionLauncher = registerForActivityResult(
         ActivityResultContracts.RequestMultiplePermissions()
@@ -412,6 +417,8 @@ class MainActivity : AppCompatActivity() {
         cosineThresholdInput.setText(formatThreshold(cosineThreshold))
         recognitionHistory.clear()
         recognitionCache.clear()
+        lastVisibleDetections = emptyList()
+        lastVisibleDetectionsAt = 0L
         overlayView.update(emptyList())
 
         if (hasCameraPermission()) {
@@ -472,6 +479,8 @@ class MainActivity : AppCompatActivity() {
         cameraProvider = null
         recognitionHistory.clear()
         recognitionCache.clear()
+        lastVisibleDetections = emptyList()
+        lastVisibleDetectionsAt = 0L
         overlayView.update(emptyList())
         processor?.close()
         processor = null
@@ -511,30 +520,33 @@ class MainActivity : AppCompatActivity() {
                 if (faces.isEmpty()) {
                     withContext(Dispatchers.Main) {
                         if (currentMode == AppMode.RECOGNITION) {
+                            if (now - lastVisibleDetectionsAt <= DETECTION_GRACE_MS && lastVisibleDetections.isNotEmpty()) {
+                                overlayView.update(lastVisibleDetections)
+                                updateRecognitionStatus(lastVisibleDetections)
+                            } else {
                                 recognitionHistory.clear()
                                 recognitionCache.clear()
+                                lastVisibleDetections = emptyList()
+                                lastVisibleDetectionsAt = 0L
                                 overlayView.update(emptyList())
                                 updateRecognitionStatus(emptyList())
                             }
                         }
+                    }
                     return@launch
                 }
 
                 val activeTrackIds = faces.mapNotNull { it.trackingId }.toSet()
                 val facesNeedingRefresh = faces.mapIndexedNotNull { index, face ->
-                    if (shouldRefreshRecognition(face.trackingId, now)) index to face else null
+                    if (shouldRefreshRecognition(face, now)) index to face else null
                 }
                 val refreshedResults = mutableMapOf<Int, RecognitionResult>()
 
                 if (facesNeedingRefresh.isNotEmpty()) {
                     val bitmap = ImageUtils.imageProxyToBitmap(image)
                     facesNeedingRefresh.forEach { (index, face) ->
-                        val embedding = faceProcessor.embedFace(bitmap, face)
-                        val refreshed = matcher.findNearest(embedding, students)
+                        val refreshed = refreshRecognition(faceProcessor, bitmap, face)
                         refreshedResults[index] = refreshed
-                        face.trackingId?.let { trackId ->
-                            recognitionCache[trackId] = CachedRecognition(refreshed, now)
-                        }
                     }
                 }
 
@@ -543,17 +555,18 @@ class MainActivity : AppCompatActivity() {
                     RawDetection(
                         bounds = face.boundingBox,
                         result = result,
-                        confidence = result.confidence
+                        confidence = if (result.student != null) result.confidence else null
                     )
                 }
-                recognitionHistory.keys.retainAll(activeTrackIds)
-                recognitionCache.keys.retainAll(activeTrackIds)
+                pruneTrackCache(now, activeTrackIds)
 
                 withContext(Dispatchers.Main) {
                     if (currentMode == AppMode.RECOGNITION) {
                         val detections = mapDetectionsToPreview(image, rawDetections)
-                        overlayView.update(detections.filter { it.result.isMatched })
+                        overlayView.update(detections)
                         updateRecognitionStatus(detections)
+                        lastVisibleDetections = detections
+                        lastVisibleDetectionsAt = now
                     }
                 }
             } catch (e: Exception) {
@@ -578,6 +591,8 @@ class MainActivity : AppCompatActivity() {
 
         val matched = detections.filter { it.result.isMatched }
         val unsure = detections.filter { it.result.matchType == MatchType.UNSURE }
+        val tracking = detections.count { it.result.matchType == MatchType.NO_DATA }
+        val noMatch = detections.count { it.result.matchType == MatchType.NO_MATCH }
         val total = detections.size
 
         if (matched.isNotEmpty()) {
@@ -591,10 +606,21 @@ class MainActivity : AppCompatActivity() {
                     val confidence = (it.confidence!! * 100).toInt()
                     "${student.name}(${confidence}%)"
                 }
-                recognitionStatusText.text = "识别到 ${matched.size}/$total 人：$names"
+                val remaining = total - matched.size
+                recognitionStatusText.text = if (remaining > 0) {
+                    "识别到 ${matched.size}/$total 人：$names，另有 $remaining 人跟踪中"
+                } else {
+                    "识别到 ${matched.size}/$total 人：$names"
+                }
             }
         } else if (unsure.isNotEmpty()) {
             recognitionStatusText.text = "疑似检测到 ${unsure.size} 人（不确定）"
+        } else if (tracking > 0) {
+            recognitionStatusText.text = if (noMatch > 0) {
+                "跟踪到 $total 人，识别中 $tracking 人，$noMatch 人未匹配"
+            } else {
+                "跟踪到 $total 人，识别中 $tracking 人"
+            }
         } else {
             recognitionStatusText.text = "检测到 $total 张未入库人脸"
         }
@@ -633,7 +659,9 @@ class MainActivity : AppCompatActivity() {
         val current = processor
         if (current != null) return current
 
-        val created = runCatching { FaceProcessor(this) }
+        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"
@@ -770,14 +798,86 @@ class MainActivity : AppCompatActivity() {
         val matched = history.filter { it.isMatched }
         if (matched.isEmpty()) return result
 
-        val best = matched.groupBy { it.student?.id }.maxByOrNull { it.value.size } ?: return result
-        return if (best.value.size >= 2) best.value.last() else result
+        val grouped = matched.groupBy { it.student?.id }
+        val best = grouped.maxByOrNull { entry ->
+            entry.value.sumOf { it.confidence.toDouble() }
+        } ?: return result
+        val bestResults = best.value
+        val bestScore = bestResults.sumOf { it.confidence.toDouble() }.toFloat()
+        val bestLatest = bestResults.maxByOrNull { it.confidence } ?: return result
+        val secondScore = grouped
+            .filterKeys { it != best.key }
+            .maxOfOrNull { entry -> entry.value.sumOf { it.confidence.toDouble() }.toFloat() }
+            ?: 0f
+
+        return when {
+            bestResults.size >= 2 -> bestLatest
+            bestLatest.confidence >= 0.92f && secondScore == 0f -> bestLatest
+            bestScore >= 1.25f && (bestScore - secondScore) >= 0.25f -> bestLatest
+            else -> result
+        }
     }
 
-    private fun shouldRefreshRecognition(trackId: Int?, now: Long): Boolean {
-        if (trackId == null) return true
+    private fun shouldRefreshRecognition(face: Face, now: Long): Boolean {
+        val trackId = face.trackingId ?: return true
         val cached = recognitionCache[trackId] ?: return true
-        return now - cached.updatedAtMs >= RECOGNITION_REFRESH_MS
+        if (now - cached.lastSeenAtMs > TRACK_STALE_MS) return true
+        if (hasTrackMovedSignificantly(cached.lastBounds, face.boundingBox)) return true
+
+        val refreshInterval = when {
+            cached.result.isMatched && cached.result.confidence >= MATCH_STABLE_CONFIDENCE -> MATCH_REFRESH_MS
+            cached.result.matchType == MatchType.UNSURE -> UNSURE_REFRESH_MS
+            cached.result.matchType == MatchType.NO_MATCH -> NO_MATCH_REFRESH_MS
+            else -> MATCH_REFRESH_MS
+        }
+        return now - cached.updatedAtMs >= refreshInterval
+    }
+
+    private fun hasTrackMovedSignificantly(previousBounds: RectF, currentBounds: Rect): Boolean {
+        val current = RectF(currentBounds)
+        val widthBase = maxOf(previousBounds.width(), current.width(), 1f)
+        val heightBase = maxOf(previousBounds.height(), current.height(), 1f)
+        val centerDx = kotlin.math.abs(previousBounds.centerX() - current.centerX()) / widthBase
+        val centerDy = kotlin.math.abs(previousBounds.centerY() - current.centerY()) / heightBase
+        val sizeDiff = kotlin.math.abs(previousBounds.width() - current.width()) / widthBase
+        val heightDiff = kotlin.math.abs(previousBounds.height() - current.height()) / heightBase
+        return centerDx > TRACK_MOVE_THRESHOLD || centerDy > TRACK_MOVE_THRESHOLD || sizeDiff > TRACK_SIZE_THRESHOLD || heightDiff > TRACK_SIZE_THRESHOLD
+    }
+
+    private fun pruneTrackCache(now: Long, activeTrackIds: Set<Int>) {
+        recognitionCache.entries.removeAll { (trackId, state) ->
+            trackId !in activeTrackIds && now - state.lastSeenAtMs > TRACK_STALE_MS
+        }
+        recognitionHistory.entries.removeAll { (trackId, _) ->
+            trackId !in recognitionCache
+        }
+    }
+
+    private fun refreshRecognition(
+        faceProcessor: FaceProcessor,
+        bitmap: Bitmap,
+        face: Face
+    ): RecognitionResult {
+        val primary = matcher.findNearest(faceProcessor.embedFace(bitmap, face), students)
+        val cached = face.trackingId?.let { recognitionCache[it] }
+        val needsEscalation = cached == null ||
+            !cached.result.isMatched ||
+            primary.matchType != MatchType.MATCH ||
+            primary.confidence < cached.result.confidence - 0.05f ||
+            hasTrackMovedSignificantly(cached.lastBounds, face.boundingBox)
+
+        if (!needsEscalation) {
+            return primary
+        }
+
+        var best = primary
+        for (embedding in faceProcessor.embedFaceCandidates(bitmap, face)) {
+            val result = matcher.findNearest(embedding, students)
+            if (recognitionScore(result) > recognitionScore(best)) {
+                best = result
+            }
+        }
+        return best
     }
 
     private fun resolveRecognitionResult(
@@ -786,16 +886,55 @@ class MainActivity : AppCompatActivity() {
         now: Long
     ): RecognitionResult {
         val trackId = face.trackingId
+        val previousState = trackId?.let { recognitionCache[it] }
+        val previousResult = previousState?.result
         val directResult = when {
             refreshedResult != null -> refreshedResult
-            trackId != null && recognitionCache.containsKey(trackId) -> recognitionCache.getValue(trackId).result
+            previousResult != null -> previousResult
             else -> RecognitionResult(null, Float.MAX_VALUE, 0f, MatchType.NO_DATA)
         }
         val stabilized = stabilizeRecognition(face, directResult)
-        if (trackId != null) {
-            recognitionCache[trackId] = CachedRecognition(stabilized, now)
+        val finalResult = when {
+            previousResult?.isMatched == true && !stabilized.isMatched -> previousResult
+            previousResult?.isMatched == true &&
+                stabilized.isMatched &&
+                previousResult.student?.id != stabilized.student?.id &&
+                stabilized.confidence < previousResult.confidence + 0.12f -> previousResult
+            previousResult?.isMatched == true &&
+                stabilized.isMatched &&
+                previousResult.student?.id == stabilized.student?.id &&
+                stabilized.confidence + 0.08f < previousResult.confidence -> previousResult
+            else -> stabilized
+        }
+        if (trackId != null) {
+            val stableFrames = when {
+                previousState?.result?.student?.id == finalResult.student?.id && finalResult.isMatched ->
+                    (previousState?.stableFrames ?: 0) + 1
+                finalResult.isMatched -> 1
+                else -> 0
+            }
+            recognitionCache[trackId] = CachedRecognition(
+                result = finalResult,
+                updatedAtMs = if (refreshedResult != null) now else previousState?.updatedAtMs ?: now,
+                lastSeenAtMs = now,
+                lastBounds = RectF(face.boundingBox),
+                stableFrames = stableFrames
+            )
+        }
+        return finalResult
+    }
+
+    private fun recognitionScore(result: RecognitionResult): Float {
+        return matchRank(result.matchType) * 10f + result.confidence + result.cosineSimilarity * 0.1f
+    }
+
+    private fun matchRank(matchType: MatchType): Int {
+        return when (matchType) {
+            MatchType.MATCH -> 3
+            MatchType.UNSURE -> 2
+            MatchType.NO_MATCH -> 1
+            MatchType.NO_DATA -> 0
         }
-        return stabilized
     }
 
     companion object {
@@ -804,8 +943,15 @@ class MainActivity : AppCompatActivity() {
         private const val KEY_COSINE_THRESHOLD = "cosine_threshold"
         private const val DEFAULT_EUCLIDEAN_THRESHOLD = 1.0f
         private const val DEFAULT_COSINE_THRESHOLD = 0.6f
-        private const val RECOGNITION_REFRESH_MS = 400L
-        private const val ANALYSIS_INTERVAL_MS = 120L
+        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 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)
     }
 }
@@ -824,5 +970,8 @@ private data class RawDetection(
 
 private data class CachedRecognition(
     val result: RecognitionResult,
-    val updatedAtMs: Long
+    val updatedAtMs: Long,
+    val lastSeenAtMs: Long,
+    val lastBounds: RectF,
+    val stableFrames: Int
 )

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

@@ -28,6 +28,8 @@ class FaceEmbedder(context: Context, private val modelType: ModelType = ModelTyp
     private val inputSize: Int
     private val outputSize: Int
     private val accelerator: Accelerator
+    private val inputPixels: IntArray
+    private val inputValues: FloatArray
 
     // ArcFace 标准化参数 (RGB)
     private val mean = floatArrayOf(127.5f, 127.5f, 127.5f)
@@ -42,6 +44,8 @@ class FaceEmbedder(context: Context, private val modelType: ModelType = ModelTyp
         val inputElements = inputBuffer.readFloat().size
         inputSize = inferInputSize(inputElements)
         outputSize = outputBuffer.readFloat().size
+        inputPixels = IntArray(inputSize * inputSize)
+        inputValues = FloatArray(inputSize * inputSize * 3)
         Log.i(TAG, "FaceEmbedder initialized with accelerator=$accelerator")
     }
 
@@ -55,41 +59,45 @@ class FaceEmbedder(context: Context, private val modelType: ModelType = ModelTyp
         val resized = Bitmap.createScaledBitmap(face, inputSize, inputSize, true)
 
         // 准备输入数据
-        val input = FloatArray(inputSize * inputSize * 3)
-        val pixels = IntArray(inputSize * inputSize)
-        resized.getPixels(pixels, 0, inputSize, 0, 0, inputSize, inputSize)
+        resized.getPixels(inputPixels, 0, inputSize, 0, 0, inputSize, inputSize)
 
         var index = 0
-        for (color in pixels) {
+        when (modelType) {
+            ModelType.FACENET -> {
+                for (color in inputPixels) {
                     val r = ((color shr 16) and 0xFF).toFloat()
                     val g = ((color shr 8) and 0xFF).toFloat()
                     val b = (color and 0xFF).toFloat()
-
-            when (modelType) {
-                ModelType.FACENET -> {
-                    // FaceNet 归一化：(pixel - 127.5) / 127.5
-                    input[index++] = (r - 127.5f) / 127.5f
-                    input[index++] = (g - 127.5f) / 127.5f
-                    input[index++] = (b - 127.5f) / 127.5f
+                    inputValues[index++] = (r - 127.5f) / 127.5f
+                    inputValues[index++] = (g - 127.5f) / 127.5f
+                    inputValues[index++] = (b - 127.5f) / 127.5f
+                }
             }
             ModelType.ARCFACE -> {
-                    // ArcFace 归一化：(pixel - mean) / std
-                    // 等价于 (pixel - 127.5) / 127.5，但语义更清晰
-                    input[index++] = (r - mean[0]) / std[0]
-                    input[index++] = (g - mean[1]) / std[1]
-                    input[index++] = (b - mean[2]) / std[2]
+                for (color in inputPixels) {
+                    val r = ((color shr 16) and 0xFF).toFloat()
+                    val g = ((color shr 8) and 0xFF).toFloat()
+                    val b = (color and 0xFF).toFloat()
+                    inputValues[index++] = (r - mean[0]) / std[0]
+                    inputValues[index++] = (g - mean[1]) / std[1]
+                    inputValues[index++] = (b - mean[2]) / std[2]
                 }
             }
         }
 
         // 执行推理
-        inputBuffer.writeFloat(input)
+        inputBuffer.writeFloat(inputValues)
         model.run(listOf(inputBuffer), listOf(outputBuffer))
 
         // L2 归一化输出特征
         return l2Normalize(outputBuffer.readFloat())
     }
 
+    fun warmUp() {
+        val bitmap = Bitmap.createBitmap(inputSize, inputSize, Bitmap.Config.ARGB_8888)
+        embed(bitmap)
+    }
+
     override fun close() {
         model.close()
     }

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

@@ -44,7 +44,7 @@ class FaceMatcher(
         // 判断匹配类型
         val matchType = determineMatchType(nearestEuclidean, nearestCosine, secondNearestEuclidean)
 
-        return if (nearest != null && matchType == MatchType.MATCH) {
+        return if (nearest != null && (matchType == MatchType.MATCH || matchType == MatchType.UNSURE)) {
             RecognitionResult(nearest, nearestEuclidean, nearestCosine, matchType)
         } else {
             RecognitionResult(null, nearestEuclidean, nearestCosine, matchType)
@@ -113,21 +113,19 @@ class FaceMatcher(
         cosineSim: Float,
         secondNearestDist: Float
     ): MatchType {
-        // 同时满足欧氏距离和余弦相似度阈值
         val euclideanMatch = euclideanDist <= euclideanThreshold
         val cosineMatch = cosineSim >= cosineThreshold
-
-        // 检查是否是明显最优（与第二名的差距）
         val gap = secondNearestDist - euclideanDist
-        val isClearlyBest = gap > 0.3f
+        val isClearlyBest = gap >= 0.12f
 
         return when {
-            // 同时满足两个阈值，或者满足一个且明显优于其他
-            (euclideanMatch && cosineMatch) || (euclideanMatch && isClearlyBest) || (cosineMatch && isClearlyBest) -> {
+            euclideanMatch && cosineMatch -> {
                 MatchType.MATCH
             }
-            // 只满足一个阈值，但不明显优于其他
-            euclideanMatch || cosineSim > (cosineThreshold * 0.9f) -> {
+            (cosineMatch && isClearlyBest) || (euclideanMatch && isClearlyBest) -> {
+                MatchType.MATCH
+            }
+            euclideanMatch || cosineMatch || cosineSim >= (cosineThreshold - 0.05f) -> {
                 MatchType.UNSURE
             }
             else -> {

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

@@ -23,7 +23,7 @@ class FaceProcessor(context: Context) : AutoCloseable {
     private val realtimeDetector = FaceDetection.getClient(
         FaceDetectorOptions.Builder()
             .setPerformanceMode(FaceDetectorOptions.PERFORMANCE_MODE_FAST)
-            .setMinFaceSize(0.08f)
+            .setMinFaceSize(0.06f)
             .enableTracking()
             .setLandmarkMode(FaceDetectorOptions.LANDMARK_MODE_ALL)
             .build()
@@ -117,9 +117,10 @@ class FaceProcessor(context: Context) : AutoCloseable {
      * 注册用的人脸检测（高精度）
      */
     private suspend fun detectFacesForEnrollment(bitmap: Bitmap): List<Face> {
-        val accurateFaces = enrollmentDetector.process(InputImage.fromBitmap(bitmap, 0)).await()
+        val image = InputImage.fromBitmap(bitmap, 0)
+        val accurateFaces = enrollmentDetector.process(image).await()
         if (accurateFaces.isNotEmpty()) return accurateFaces
-        return realtimeDetector.process(InputImage.fromBitmap(bitmap, 0)).await()
+        return realtimeDetector.process(image).await()
     }
 
     /**
@@ -128,11 +129,27 @@ class FaceProcessor(context: Context) : AutoCloseable {
     fun embedFace(bitmap: Bitmap, face: Face): FloatArray {
         val crop = requireNotNull(safeCropRect(bitmap, face.boundingBox)) { "人脸区域超出图片边界。" }
         val cropped = Bitmap.createBitmap(bitmap, crop.left, crop.top, crop.width(), crop.height())
-        val aligned = alignFaceWithLandmarks(cropped, face, crop.left, crop.top)
+        val aligned = alignFaceForRecognition(cropped, face, crop.left, crop.top)
         val preprocessed = preprocessor.preprocessRealtime(aligned)
         return embedder.embed(preprocessed)
     }
 
+    fun embedFaceCandidates(bitmap: Bitmap, face: Face): List<FloatArray> {
+        val embeddings = mutableListOf<FloatArray>()
+        for (paddingRatio in FACE_PADDING_RATIOS) {
+            val crop = safeCropRect(bitmap, face.boundingBox, paddingRatio) ?: continue
+            val cropped = Bitmap.createBitmap(bitmap, crop.left, crop.top, crop.width(), crop.height())
+            val aligned = alignFaceForRecognition(cropped, face, crop.left, crop.top)
+            val preprocessed = preprocessor.preprocessRealtime(aligned)
+            embeddings.add(embedder.embed(preprocessed))
+        }
+        return embeddings
+    }
+
+    fun warmUp() {
+        embedder.warmUp()
+    }
+
     override fun close() {
         realtimeDetector.close()
         enrollmentDetector.close()
@@ -143,10 +160,14 @@ class FaceProcessor(context: Context) : AutoCloseable {
      * 计算安全的裁剪区域（带 padding）
      */
     private fun safeCropRect(bitmap: Bitmap, box: Rect): Rect? {
+        return safeCropRect(bitmap, box, 0.25f)
+    }
+
+    private fun safeCropRect(bitmap: Bitmap, box: Rect, paddingRatio: Float): Rect? {
         if (box.width() <= 0 || box.height() <= 0) return null
 
         // 增加 padding 比例，保留更多上下文信息
-        val padding = (maxOf(box.width(), box.height()) * 0.25f).toInt()
+        val padding = (maxOf(box.width(), box.height()) * paddingRatio).toInt()
 
         val left = (box.left - padding).coerceIn(0, bitmap.width)
         val top = (box.top - padding).coerceIn(0, bitmap.height)
@@ -160,4 +181,30 @@ class FaceProcessor(context: Context) : AutoCloseable {
 
         return Rect(left, top, right, bottom)
     }
+
+    private fun alignFaceForRecognition(
+        faceBitmap: Bitmap,
+        face: Face,
+        faceLeft: Int,
+        faceTop: Int
+    ): Bitmap {
+        val leftEye = face.getLandmark(FaceLandmark.LEFT_EYE)
+        val rightEye = face.getLandmark(FaceLandmark.RIGHT_EYE)
+        if (leftEye != null && rightEye != null) {
+            return aligner.align(
+                bitmap = faceBitmap,
+                leftEyeX = leftEye.position.x,
+                leftEyeY = leftEye.position.y,
+                rightEyeX = rightEye.position.x,
+                rightEyeY = rightEye.position.y,
+                faceLeft = faceLeft,
+                faceTop = faceTop
+            )
+        }
+        return aligner.alignWithEstimation(faceBitmap, 0, 0, face.headEulerAngleZ)
+    }
+
+    companion object {
+        private val FACE_PADDING_RATIOS = floatArrayOf(0.18f, 0.34f)
+    }
 }

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

@@ -16,22 +16,22 @@ import kotlin.math.min
 class ImagePreprocessor {
 
     /**
-     * 完整的预处理流程：
-     * 1. 亮度/对比度增强
-     * 2. 直方图均衡化（简化版）
-     * 3. 转换为适合模型输入的格式
+     * 统一的人脸预处理流程。
+     *
+     * 识别和注册都走同一套轻量增强，减少两条链路的特征分布差异。
      */
     fun preprocess(bitmap: Bitmap): Bitmap {
-        // 先进行亮度对比度增强
-        val enhanced = enhanceBrightnessContrast(bitmap)
-        // 再进行直方图均衡化
-        return histogramEqualize(enhanced)
+        return normalizeForEmbedding(bitmap)
     }
 
     /**
-     * 实时识别场景使用轻量预处理，减少每帧 CPU 开销。
+     * 实时识别场景使用同一套轻量预处理，避免额外的帧间波动。
      */
     fun preprocessRealtime(bitmap: Bitmap): Bitmap {
+        return normalizeForEmbedding(bitmap)
+    }
+
+    private fun normalizeForEmbedding(bitmap: Bitmap): Bitmap {
         return enhanceBrightnessContrast(bitmap, brightness = 0.06f, contrast = 1.08f)
     }
 

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

@@ -79,17 +79,20 @@ class SmartEnrollment(context: Context) : AutoCloseable {
                 faceInfo.faceArea
             }
 
-            // 裁剪人脸
-            val croppedFace = cropFace(bitmap, largestFace.cropRect)
-
-            // 对齐
-            val alignedFace = alignFace(croppedFace, largestFace.face, largestFace.cropRect)
-
-            // 预处理
+            val embeddings = mutableListOf<FloatArray>()
+            for (paddingRatio in IMAGE_PADDING_RATIOS) {
+                val cropRect = safeCropRect(bitmap, largestFace.face.boundingBox, paddingRatio) ?: continue
+                val croppedFace = cropFace(bitmap, cropRect)
+                val alignedFace = alignFace(croppedFace, largestFace.face, cropRect)
                 val preprocessedFace = preprocessor.preprocess(alignedFace)
+                embeddings.add(embedder.embed(preprocessedFace))
+            }
 
-            // 提取特征
-            val embedding = embedder.embed(preprocessedFace)
+            if (embeddings.isEmpty()) {
+                return EnrollmentResult.Failure("未检测到可用人脸，请换更清晰或更正面的照片。")
+            }
+
+            val embedding = fuseEmbeddings(embeddings)
 
             EnrollmentResult.Success(embedding, EnrollmentSourceType.IMAGE)
         } catch (e: Exception) {
@@ -344,9 +347,13 @@ class SmartEnrollment(context: Context) : AutoCloseable {
     }
 
     private fun safeCropRect(bitmap: Bitmap, box: Rect): Rect? {
+        return safeCropRect(bitmap, box, 0.25f)
+    }
+
+    private fun safeCropRect(bitmap: Bitmap, box: Rect, paddingRatio: Float): Rect? {
         if (box.width() <= 0 || box.height() <= 0) return null
 
-        val padding = (maxOf(box.width(), box.height()) * 0.25f).toInt()
+        val padding = (maxOf(box.width(), box.height()) * paddingRatio).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)
@@ -422,6 +429,7 @@ class SmartEnrollment(context: Context) : AutoCloseable {
     companion object {
         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)
     }
 }
 

app/src/main/java/com/example/studentfaceregistry/ui/OverlayView.kt

@@ -48,7 +48,7 @@ class OverlayView @JvmOverloads constructor(
     private val colorMatch = Color.rgb(34, 197, 94)      // 绿色 - 匹配成功
     private val colorUnsure = Color.rgb(234, 179, 8)    // 黄色 - 不确定
     private val colorNoMatch = Color.rgb(239, 68, 68)   // 红色 - 未匹配
-    private val colorUnknown = Color.rgb(59, 130, 246)  // 蓝色 - 未入库
+    private val colorUnknown = Color.rgb(59, 130, 246)  // 蓝色 - 识别中
 
     private var detections: List<DetectionUi> = emptyList()
 
@@ -136,7 +136,7 @@ data class DetectionUi(
                 result.isMatched -> "${result.student?.name} (${result.student?.studentNo})"
                 result.matchType == MatchType.UNSURE -> "疑似：${result.student?.name}"
                 result.student == null && result.distance != Float.MAX_VALUE -> "未匹配"
-                else -> "未入库"
+                else -> "识别中"
             }
         }
 }