bf04400852
OCR 模型按需下载（P8 瘦身）： - 移除 plugins/ppocr/assets/ 内置模型（det 9.8MB + rec 21MB + dict），APK 减包 ~31MB - 新增 modelDownloader.ts：优先从 HuggingFace/CDN 下载，兜底从 APK assets 拷贝 - OcrModule.kt 新增 setModelDir，支持从 filesystem 加载模型，回退 assets 兼容旧用户 - settingsStore 持久化 ocrModelVersion / ocrModelDir - 自动化页集成模型状态检查与一键下载 UI OCR 三层独立控制： - OcrProcessorConfig 从单一 aiVisionEnabled 拆分为 layer1/2/3 三个独立开关 - 设置页可按层启停（L1 正则规则 / L2 本地 OCR / L3 AI Vision） - OCR 处理增加耗时与字符数日志 日志系统升级： - logger.ts 新增 LogFileBackend 抽象，支持磁盘持久化（按日期 app-YYYY-MM-DD.log） - 新增 logBackend.ts（ExpoLogFileBackend）+ 日志中心页 settings/logs.tsx - 日志中心：实时缓冲 + 历史文件、4 级过滤、Tag/关键词搜索、JSON 展开、分享导出、7 天过期清理 - _layout.tsx 启动时初始化文件后端 + 日志脱敏（验证码/卡号） 原生浮层 UI 主题同步（P6）： - 新增 floatingUiConfig.ts：JS 侧从 theme tokens + i18n 构建 FloatingUiConfig 推送原生 - 新增 FloatingUiConfigStore.kt：SharedPreferences 存储，三浮层组件读取 - FloatingBillView 重设计：颜色/文案走配置、新增币种 chip、金额校验改 BigDecimal - FloatingHelper / FloatingTip 同步适配 - _layout.tsx 新增 FloatingUiConfigSyncer，主题/语言切换自动推送 UI 与组件增强： - FormModal 新增 select/dropdown 控件、行内布局（row/flex）、联动回调 onValuesChange - 新增 Touchable 通用触摸组件、AccountCreateModal 快速建账弹窗 - 信用卡页展示账单周期/到期还款日/本期应还/剩余可用额度，关联账户改下拉选择 - AI 设置页重做：OpenAI/Gemini/DeepSeek 预设 + 默认 URL/模型 - 引导页新增 Android 权限检查步骤（无障碍/通知/短信/存储/悬浮窗） 去重优化： - 对手方匹配改为模糊包含（includes），双方均无对手方时判定低置信度重复 - DedupResult 新增 matchedItem 返回匹配对比项 文档重构： - README.md 重写为入口索引（品牌更新 + 模块概览 + 文档导航表） - 新增 AGENTS.md（AI 助手贡献指南）、docs/architecture.md（Mermaid 数据流/分层/OCR 级联图） - 新增 docs/development.md（环境/命令/编码规范/测试/提交规范）、plugins/README.md - UI 重设计文档（design spec + p1-p8）移入 docs/design/ 其他： - i18n 新增权限/信用卡详情/日志中心/AI 设置等翻译键 - ppocr Config Plugin 修复 import 注入去重；size-optimization 增强 - 新增测试：logger.test.ts、floating-ui-config.test.ts
706 lines
29 KiB
Kotlin
706 lines
29 KiB
Kotlin
package com.beancount.mobile.ppocr
|
||
|
||
import android.graphics.Bitmap
|
||
import android.graphics.BitmapFactory
|
||
import android.util.Base64
|
||
import android.util.Log
|
||
import ai.onnxruntime.OnnxTensor
|
||
import ai.onnxruntime.OrtEnvironment
|
||
import ai.onnxruntime.OrtSession
|
||
import com.facebook.react.bridge.Arguments
|
||
import com.facebook.react.bridge.Promise
|
||
import com.facebook.react.bridge.ReactApplicationContext
|
||
import com.facebook.react.bridge.ReactContextBaseJavaModule
|
||
import com.facebook.react.bridge.ReactMethod
|
||
import com.facebook.react.bridge.ReadableArray
|
||
import com.facebook.react.bridge.WritableMap
|
||
import com.facebook.react.bridge.WritableNativeArray
|
||
import com.facebook.react.bridge.WritableNativeMap
|
||
import com.facebook.react.module.annotations.ReactModule
|
||
import kotlinx.coroutines.CoroutineScope
|
||
import kotlinx.coroutines.Dispatchers
|
||
import kotlinx.coroutines.SupervisorJob
|
||
import kotlinx.coroutines.launch
|
||
import java.io.BufferedReader
|
||
import java.io.InputStreamReader
|
||
import java.nio.FloatBuffer
|
||
import java.util.concurrent.locks.ReentrantLock
|
||
import kotlin.math.max
|
||
import kotlin.math.min
|
||
|
||
/**
|
||
* PP-OCRv6 (ONNX Runtime) React Native Bridge（plan.md「3.4 Layer 2」+「决策 4 Config Plugin」）。
|
||
*
|
||
* 引擎：ONNX Runtime（跨平台、微软官方、Windows 友好），替代 NCNN 路线。
|
||
* 模型：ppocrv6_det.onnx + ppocrv6_rec.onnx（PP-OCRv6 small，从 PaddlePaddle 官方 HuggingFace 下载）。
|
||
* 字典：ppocrv6_dict.txt（PP-OCRv6 多语言字典，18708 字符；rec 模型 18710 维输出 = 字典 + blank + 特殊位）。
|
||
*
|
||
* 流水线：
|
||
* 1. det（文本检测）：bitmap → DB 后处理得到文本框
|
||
* 2. rec（文本识别）：每个框 crop → resize 到 48px 高 → CTC 解码
|
||
*
|
||
* 性能优化（参考 AutoAccounting OcrProcessor.kt）：
|
||
* - 短边压缩到 720px（像素量比 1440p 减少约 75%）
|
||
* - CPU 执行（兼容性最稳，部分设备 GPU 会崩溃）
|
||
* - det 最大边限制 960（PaddleOCR 默认 limit_max_side_len）
|
||
*
|
||
* JS 层通过 NativeModules.PpOcr.recognizeText(base64) 调用。
|
||
*/
|
||
const val OCR_MODULE_NAME = "PpOcr"
|
||
|
||
@ReactModule(name = OCR_MODULE_NAME)
|
||
class OcrModule(private val context: ReactApplicationContext) :
|
||
ReactContextBaseJavaModule(context) {
|
||
|
||
private val scope = CoroutineScope(SupervisorJob() + Dispatchers.IO)
|
||
private val lock = ReentrantLock()
|
||
|
||
private var ortEnv: OrtEnvironment? = null
|
||
private var detSession: OrtSession? = null
|
||
private var recSession: OrtSession? = null
|
||
private var dictionary: List<String> = emptyList()
|
||
@Volatile private var initialized = false
|
||
@Volatile private var initFailed = false
|
||
|
||
/** 模型文件目录（filesystem 绝对路径）。非空时从该目录加载模型，否则回退到 assets。 */
|
||
@Volatile private var modelDir: String? = null
|
||
|
||
override fun getName(): String = OCR_MODULE_NAME
|
||
|
||
override fun initialize() {
|
||
super.initialize()
|
||
// 异步加载模型，避免阻塞 RN 桥初始化
|
||
scope.launch { initEngine() }
|
||
}
|
||
|
||
/** 从 assets 或 filesystem 加载 det/rec ONNX 模型与字典。 */
|
||
private fun initEngine() {
|
||
lock.lock()
|
||
try {
|
||
if (initialized || initFailed) return
|
||
val env = OrtEnvironment.getEnvironment()
|
||
val opts = OrtSession.SessionOptions().apply {
|
||
// CPU 线程数：2 是兼容性/性能的稳妥折中（高端机可调高）
|
||
setInterOpNumThreads(2)
|
||
setIntraOpNumThreads(2)
|
||
// 移动端关闭内存优化里的图优化级别过高（部分模型会崩）
|
||
setOptimizationLevel(OrtSession.SessionOptions.OptLevel.BASIC_OPT)
|
||
}
|
||
val dir = modelDir
|
||
val (det, rec, dict) = if (dir != null) {
|
||
// 从 filesystem 加载（P8：模型按需下载到本地目录）
|
||
val detPath = dir + java.io.File.separator + ASSET_DET_MODEL
|
||
val recPath = dir + java.io.File.separator + ASSET_REC_MODEL
|
||
val dictPath = dir + java.io.File.separator + ASSET_DICT
|
||
Log.i(OCR_MODULE_NAME, "从 filesystem 加载模型: det=$detPath, rec=$recPath")
|
||
Triple(
|
||
env.createSession(detPath, opts),
|
||
env.createSession(recPath, opts),
|
||
loadDictionaryFromFile(dictPath)
|
||
)
|
||
} else {
|
||
// 回退：从 assets 加载（兼容未迁移的老用户）
|
||
val detBytes = context.assets.open(ASSET_DET_MODEL).use { it.readBytes() }
|
||
val recBytes = context.assets.open(ASSET_REC_MODEL).use { it.readBytes() }
|
||
Log.i(OCR_MODULE_NAME, "从 assets 加载模型（兼容模式）")
|
||
Triple(
|
||
env.createSession(detBytes, opts),
|
||
env.createSession(recBytes, opts),
|
||
loadDictionaryFromAssets()
|
||
)
|
||
}
|
||
|
||
detSession = det
|
||
recSession = rec
|
||
ortEnv = env
|
||
dictionary = dict
|
||
initialized = true
|
||
Log.i(OCR_MODULE_NAME, "PP-OCRv6 ONNX 模型加载成功（det+rec, dict=${dict.size}）")
|
||
} catch (e: Exception) {
|
||
initFailed = true
|
||
Log.e(OCR_MODULE_NAME, "OCR 初始化失败: ${e.message}", e)
|
||
val dir = modelDir
|
||
if (dir != null) {
|
||
Log.e(OCR_MODULE_NAME, "请确认 $dir 下存在 $ASSET_DET_MODEL / $ASSET_REC_MODEL / $ASSET_DICT")
|
||
} else {
|
||
Log.e(OCR_MODULE_NAME, "请确认 assets 下存在 $ASSET_DET_MODEL / $ASSET_REC_MODEL / $ASSET_DICT")
|
||
}
|
||
} finally {
|
||
lock.unlock()
|
||
}
|
||
}
|
||
|
||
/** rec 推理用的 OrtEnvironment（复用 ortEnv 单例）。 */
|
||
private val recEnv: OrtEnvironment? get() = ortEnv
|
||
|
||
/**
|
||
* 从 assets 加载 ppocrv6_dict.txt 字典。
|
||
*
|
||
* PaddleOCR CTC 约定：模型输出 logits 的 index 0 是 blank，字符从 index 1 开始；
|
||
* 字典条目 dictionary[i] 对应模型输出 index i+1。解码时 dictIdx = argmaxIdx - 1。
|
||
* 字典本身不含 blank，运行时固定用 index 0 作 blank（见 ctcGreedyDecode）。
|
||
*/
|
||
private fun loadDictionaryFromAssets(): List<String> {
|
||
val words = mutableListOf<String>()
|
||
context.assets.open(ASSET_DICT).use { stream ->
|
||
BufferedReader(InputStreamReader(stream, Charsets.UTF_8)).useLines { lines ->
|
||
lines.forEach { line ->
|
||
// PaddleOCR 字典每行一个字符（去掉行尾换行）
|
||
words.add(line.trimEnd('\r', '\n'))
|
||
}
|
||
}
|
||
}
|
||
return words
|
||
}
|
||
|
||
/** 从 filesystem 路径加载字典文件。 */
|
||
private fun loadDictionaryFromFile(path: String): List<String> {
|
||
val words = mutableListOf<String>()
|
||
java.io.File(path).bufferedReader(Charsets.UTF_8).useLines { lines ->
|
||
lines.forEach { line ->
|
||
words.add(line.trimEnd('\r', '\n'))
|
||
}
|
||
}
|
||
return words
|
||
}
|
||
|
||
/**
|
||
* 识别图片文本。
|
||
* @param imageBase64 base64 编码的图片（JPEG/PNG）
|
||
* @return 识别出的纯文本（所有行用 \n 连接）
|
||
*/
|
||
@ReactMethod
|
||
fun recognizeText(imageBase64: String, promise: Promise) {
|
||
scope.launch {
|
||
var bitmap: Bitmap? = null
|
||
var scaled: Bitmap? = null
|
||
try {
|
||
ensureReady()
|
||
bitmap = decodeBase64(imageBase64)
|
||
if (bitmap == null) {
|
||
promise.reject("DECODE_FAILED", "base64 解码失败")
|
||
return@launch
|
||
}
|
||
scaled = scaleDownForOcr(bitmap, OCR_MAX_SHORT_EDGE)
|
||
val blocks = runInference(scaled)
|
||
val text = blocks.joinToString("\n") { it.text }
|
||
promise.resolve(text)
|
||
} catch (e: Exception) {
|
||
Log.e(OCR_MODULE_NAME, "recognizeText 异常: ${e.message}", e)
|
||
promise.reject("OCR_ERROR", e.message)
|
||
} finally {
|
||
bitmap?.recycle()
|
||
if (scaled !== bitmap) {
|
||
scaled?.recycle()
|
||
}
|
||
}
|
||
}
|
||
}
|
||
|
||
/**
|
||
* 识别并返回带坐标的文本块（用于复杂版面）。
|
||
* @return JSON 数组字符串：[{text, x, y, width, height, confidence}]
|
||
*/
|
||
@ReactMethod
|
||
fun recognizeTextBlocks(imageBase64: String, promise: Promise) {
|
||
scope.launch {
|
||
var bitmap: Bitmap? = null
|
||
var scaled: Bitmap? = null
|
||
try {
|
||
ensureReady()
|
||
bitmap = decodeBase64(imageBase64)
|
||
if (bitmap == null) {
|
||
promise.reject("DECODE_FAILED", "base64 解码失败")
|
||
return@launch
|
||
}
|
||
scaled = scaleDownForOcr(bitmap, OCR_MAX_SHORT_EDGE)
|
||
val blocks = runInference(scaled)
|
||
// 序列化为 RN WritableArray
|
||
val arr = WritableNativeArray()
|
||
for (b in blocks) {
|
||
val map: WritableMap = WritableNativeMap()
|
||
map.putString("text", b.text)
|
||
map.putDouble("x", b.x.toDouble())
|
||
map.putDouble("y", b.y.toDouble())
|
||
map.putDouble("width", b.width.toDouble())
|
||
map.putDouble("height", b.height.toDouble())
|
||
map.putDouble("confidence", b.confidence.toDouble())
|
||
arr.pushMap(map)
|
||
}
|
||
promise.resolve(arr)
|
||
} catch (e: Exception) {
|
||
Log.e(OCR_MODULE_NAME, "recognizeTextBlocks 异常: ${e.message}", e)
|
||
promise.reject("OCR_ERROR", e.message)
|
||
} finally {
|
||
bitmap?.recycle()
|
||
if (scaled !== bitmap) {
|
||
scaled?.recycle()
|
||
}
|
||
}
|
||
}
|
||
}
|
||
|
||
/** 引擎是否已就绪（模型加载完成）。 */
|
||
@ReactMethod
|
||
fun isReady(promise: Promise) {
|
||
promise.resolve(initialized)
|
||
}
|
||
|
||
/** 设置模型文件目录（绝对路径）。若引擎已初始化则释放并重新加载。 */
|
||
@ReactMethod
|
||
fun setModelDir(dir: String, promise: Promise) {
|
||
modelDir = dir
|
||
if (initialized) {
|
||
release()
|
||
initFailed = false
|
||
scope.launch { initEngine() }
|
||
}
|
||
promise.resolve(true)
|
||
}
|
||
|
||
// ============== 推理流水线 ==============
|
||
|
||
private fun ensureReady() {
|
||
if (!initialized && !initFailed) initEngine()
|
||
if (!initialized) throw IllegalStateException("OCR 引擎未就绪（模型未加载，${if (initFailed) "初始化失败" else "加载中"}）")
|
||
}
|
||
|
||
/** 完整推理：det 检测文本框 → 对每个框 rec 识别 → 返回带坐标的文本块。 */
|
||
private fun runInference(bitmap: Bitmap): List<OcrBlock> {
|
||
Log.i(OCR_MODULE_NAME, "runInference 开始: bitmap 尺寸 = ${bitmap.width}x${bitmap.height}")
|
||
val det = detSession ?: run {
|
||
Log.e(OCR_MODULE_NAME, "detSession 为空，放弃推理")
|
||
return emptyList()
|
||
}
|
||
val rec = recSession ?: run {
|
||
Log.e(OCR_MODULE_NAME, "recSession 为空，放弃推理")
|
||
return emptyList()
|
||
}
|
||
|
||
var resized: Bitmap? = null
|
||
var detInputTensor: OnnxTensor? = null
|
||
var detOutputs: OrtSession.Result? = null
|
||
val results = mutableListOf<OcrBlock>()
|
||
|
||
try {
|
||
// ---- 1. 文本检测（DB）----
|
||
resized = resizeForDet(bitmap, DET_LIMIT_MAX_SIDE)
|
||
Log.i(OCR_MODULE_NAME, "det 图像缩放后尺寸 = ${resized.width}x${resized.height}")
|
||
val ratioX = bitmap.width.toFloat() / resized.width
|
||
val ratioY = bitmap.height.toFloat() / resized.height
|
||
|
||
val detInput = preprocessDet(resized)
|
||
detInputTensor = OnnxTensor.createTensor(recEnv, FloatBuffer.wrap(detInput.data), longArrayOf(1L, 3L, detInput.h.toLong(), detInput.w.toLong()))
|
||
val detInputs = mapOf("x" to detInputTensor)
|
||
detOutputs = det.run(detInputs)
|
||
@Suppress("UNCHECKED_CAST")
|
||
val detProb = (detOutputs[0].value as Array<Array<Array<FloatArray>>>)[0][0] // [H,W]
|
||
Log.i(OCR_MODULE_NAME, "det 推理完成，概率图尺寸 = ${detProb.size}x${detProb[0].size}")
|
||
|
||
// DB 后处理：threshold → 轮廓 → 最小外接矩形
|
||
val boxes = dbPostprocess(detProb, detInput.h, detInput.w, ratioX, ratioY)
|
||
Log.i(OCR_MODULE_NAME, "dbPostprocess 后处理完成，检测到文本框数量 = ${boxes.size}")
|
||
|
||
if (boxes.isEmpty()) return emptyList()
|
||
|
||
// ---- 2. 文本识别（CRNN+CTC）----
|
||
for ((idx, box) in boxes.withIndex()) {
|
||
var crop: Bitmap? = null
|
||
var recInputTensor: OnnxTensor? = null
|
||
var recOutputs: OrtSession.Result? = null
|
||
try {
|
||
crop = cropBox(bitmap, box)
|
||
if (crop == null) {
|
||
Log.w(OCR_MODULE_NAME, "裁剪文本框失败 (index = $idx)")
|
||
continue
|
||
}
|
||
val recInput = preprocessRec(crop)
|
||
recInputTensor = OnnxTensor.createTensor(recEnv, FloatBuffer.wrap(recInput.data), longArrayOf(1L, 3L, REC_IMAGE_HEIGHT.toLong(), recInput.w.toLong()))
|
||
val recInputs = mapOf("x" to recInputTensor)
|
||
recOutputs = rec.run(recInputs)
|
||
// 输出 shape: [1, T, numClasses]
|
||
@Suppress("UNCHECKED_CAST")
|
||
val logits = (recOutputs[0].value as Array<Array<FloatArray>>)[0]
|
||
|
||
val (text, conf) = ctcGreedyDecode(logits)
|
||
val xs = box.map { it[0] }
|
||
val ys = box.map { it[1] }
|
||
val minX = (xs.minOrNull() ?: 0f).toInt()
|
||
val minY = (ys.minOrNull() ?: 0f).toInt()
|
||
val maxX = (xs.maxOrNull() ?: 0f).toInt()
|
||
val maxY = (ys.maxOrNull() ?: 0f).toInt()
|
||
val w = maxX - minX
|
||
val h = maxY - minY
|
||
Log.i(OCR_MODULE_NAME, "文本框 $idx 识别结果 = '$text', 坐标 = ($minX, $minY, $w, $h), 置信度 = $conf")
|
||
if (text.isNotEmpty()) {
|
||
results.add(OcrBlock(text, minX.toFloat(), minY.toFloat(), w.toFloat(), h.toFloat(), conf))
|
||
}
|
||
} finally {
|
||
crop?.recycle()
|
||
recInputTensor?.close()
|
||
recOutputs?.close()
|
||
}
|
||
}
|
||
} finally {
|
||
if (resized !== bitmap) {
|
||
resized?.recycle()
|
||
}
|
||
detInputTensor?.close()
|
||
detOutputs?.close()
|
||
}
|
||
|
||
return results
|
||
}
|
||
|
||
// ============== 前处理 ==============
|
||
|
||
/** det 前处理：resize → BCHW → normalize(mean=[0.485,0.456,0.406], std=[0.229,0.224,0.225])。 */
|
||
private fun preprocessDet(bmp: Bitmap): TensorData {
|
||
val w = bmp.width
|
||
val h = bmp.height
|
||
val pixels = IntArray(w * h)
|
||
bmp.getPixels(pixels, 0, w, 0, 0, w, h)
|
||
val data = FloatArray(3 * w * h)
|
||
// BCHW 顺序
|
||
val means = floatArrayOf(0.485f, 0.456f, 0.406f)
|
||
val stds = floatArrayOf(0.229f, 0.224f, 0.225f)
|
||
for (c in 0..2) {
|
||
val mean = means[c]
|
||
val std = stds[c]
|
||
for (y in 0 until h) {
|
||
for (x in 0 until w) {
|
||
val px = pixels[y * w + x]
|
||
// 提取 R/G/B（c=0→R, 1→G, 2→B）
|
||
val channelVal = (px shr (16 - 8 * c)) and 0xFF
|
||
data[c * w * h + y * w + x] = (channelVal / 255.0f - mean) / std
|
||
}
|
||
}
|
||
}
|
||
return TensorData(data, w, h)
|
||
}
|
||
|
||
/** rec 前处理：crop → resize 到 48 高（保持宽高比）→ pad 到整除 4 → normalize。 */
|
||
private fun preprocessRec(bmp: Bitmap): TensorData {
|
||
var w = bmp.width
|
||
val h = bmp.height
|
||
// resize 到高度 48，宽度等比缩放
|
||
var resizedW = (w.toFloat() / h * REC_IMAGE_HEIGHT).toInt()
|
||
// 宽度上限，避免单行过长爆显存
|
||
resizedW = min(resizedW, REC_MAX_WIDTH)
|
||
resizedW = max(resizedW, 1)
|
||
val resized = if (resizedW == w && h == REC_IMAGE_HEIGHT) bmp
|
||
else Bitmap.createScaledBitmap(bmp, resizedW, REC_IMAGE_HEIGHT, true)
|
||
w = resized.width
|
||
val pixels = IntArray(w * REC_IMAGE_HEIGHT)
|
||
resized.getPixels(pixels, 0, w, 0, 0, w, REC_IMAGE_HEIGHT)
|
||
val data = FloatArray(3 * w * REC_IMAGE_HEIGHT)
|
||
val means = floatArrayOf(0.5f, 0.5f, 0.5f)
|
||
val stds = floatArrayOf(0.5f, 0.5f, 0.5f)
|
||
for (c in 0..2) {
|
||
val mean = means[c]
|
||
val std = stds[c]
|
||
for (y in 0 until REC_IMAGE_HEIGHT) {
|
||
for (x in 0 until w) {
|
||
val px = pixels[y * w + x]
|
||
val channelVal = (px shr (16 - 8 * c)) and 0xFF
|
||
data[c * w * REC_IMAGE_HEIGHT + y * w + x] = (channelVal / 255.0f - mean) / std
|
||
}
|
||
}
|
||
}
|
||
if (resized !== bmp) resized.recycle()
|
||
return TensorData(data, w, REC_IMAGE_HEIGHT)
|
||
}
|
||
|
||
// ============== DB 后处理（简化版） ==============
|
||
// 参考 PaddleOCR db_postprocess：sigmoid → threshold → 连通域 → 最小外接矩形
|
||
// 这里用轻量实现：逐像素阈值化后用投影法估框，对常见单/多行账单足够。
|
||
|
||
/**
|
||
* DB 后处理：sigmoid + 阈值 0.3 → 二值图 → 连通域外接矩形。
|
||
* 简化版：用水平投影切行 + 垂直投影切列，得到矩形框（对账单类版面够用）。
|
||
*/
|
||
private fun dbPostprocess(prob: Array<FloatArray>, h: Int, w: Int, ratioX: Float, ratioY: Float): List<List<FloatArray>> {
|
||
// 自动检测是否需要 Sigmoid
|
||
var minVal = Float.MAX_VALUE
|
||
var maxVal = Float.MIN_VALUE
|
||
for (y in 0 until h) {
|
||
for (x in 0 until w) {
|
||
val v = prob[y][x]
|
||
if (v < minVal) minVal = v
|
||
if (v > maxVal) maxVal = v
|
||
}
|
||
}
|
||
val needSigmoid = minVal < -0.05f || maxVal > 1.05f
|
||
Log.i(OCR_MODULE_NAME, "DB prob min=$minVal, max=$maxVal, needSigmoid=$needSigmoid")
|
||
|
||
// 过滤状态栏（前8%）和导航栏（后8%），避免其上的干扰字符（如电池、时间、返回键）影响识别或导致行粘连
|
||
val startY = (h * 0.08).toInt()
|
||
val endY = (h * 0.92).toInt()
|
||
|
||
val binMask = Array(h) { IntArray(w) }
|
||
var activeCountTotal = 0
|
||
for (y in 0 until h) {
|
||
for (x in 0 until w) {
|
||
if (y < startY || y > endY) {
|
||
binMask[y][x] = 0
|
||
continue
|
||
}
|
||
val raw = prob[y][x]
|
||
val sig = if (needSigmoid) {
|
||
1.0f / (1.0f + Math.exp(-raw.toDouble()).toFloat())
|
||
} else {
|
||
raw
|
||
}
|
||
val isActive = if (sig > DET_THRESH) 1 else 0
|
||
binMask[y][x] = isActive
|
||
if (isActive == 1) activeCountTotal++
|
||
}
|
||
}
|
||
Log.i(OCR_MODULE_NAME, "二值化完成: 活跃像素 = $activeCountTotal / ${w * h}")
|
||
|
||
// 清理垂直干扰线（如滚动条、背景边框线）：如果某列在文本有效区域内的活跃像素超过该区域高度的 30%，视为干扰列，整列清零
|
||
val maxColActive = ((endY - startY) * 0.3).toInt()
|
||
var clearedColsCount = 0
|
||
for (x in 0 until w) {
|
||
var colActive = 0
|
||
for (y in startY..endY) {
|
||
if (binMask[y][x] == 1) colActive++
|
||
}
|
||
if (colActive > maxColActive) {
|
||
clearedColsCount++
|
||
for (y in 0 until h) {
|
||
binMask[y][x] = 0
|
||
}
|
||
}
|
||
}
|
||
Log.i(OCR_MODULE_NAME, "垂直线噪清理完成: 清理了 $clearedColsCount / $w 列")
|
||
|
||
// 水平投影：按行找文本行
|
||
val rowHits = IntArray(h)
|
||
for (y in 0 until h) {
|
||
var sum = 0
|
||
for (x in 0 until w) sum += binMask[y][x]
|
||
rowHits[y] = sum
|
||
}
|
||
val minRowWidth = max(1, w / 20) // 一行至少要有这么多像素才算文本
|
||
val rowRanges = mutableListOf<IntArray>()
|
||
var inLine = false
|
||
var lineStart = 0
|
||
for (y in 0 until h) {
|
||
val isText = rowHits[y] >= minRowWidth
|
||
if (isText && !inLine) { inLine = true; lineStart = y }
|
||
else if (!isText && inLine) {
|
||
rowRanges.add(intArrayOf(lineStart, y - 1))
|
||
inLine = false
|
||
}
|
||
}
|
||
if (inLine) rowRanges.add(intArrayOf(lineStart, h - 1))
|
||
Log.i(OCR_MODULE_NAME, "dbPostprocess 水平分割完成，找到行Ranges数 = ${rowRanges.size}")
|
||
|
||
val boxes = mutableListOf<List<FloatArray>>()
|
||
// 对每行做垂直投影切列（账单每行通常是连续一段或多段）
|
||
for ((y0, y1) in rowRanges.map { it[0] to it[1] }) {
|
||
val colHits = IntArray(w)
|
||
for (x in 0 until w) {
|
||
var sum = 0
|
||
for (y in y0..y1) sum += binMask[y][x]
|
||
colHits[x] = sum
|
||
}
|
||
val minColHeight = max(1, (y1 - y0 + 1) / 12)
|
||
var inSeg = false
|
||
var segStart = 0
|
||
var segs = mutableListOf<IntArray>()
|
||
for (x in 0 until w) {
|
||
val isText = colHits[x] >= minColHeight
|
||
if (isText && !inSeg) { inSeg = true; segStart = x }
|
||
else if (!isText && inSeg) {
|
||
// 合并间隔很近的段
|
||
if (segs.isNotEmpty() && segStart - segs.last()[1] < DET_MERGE_GAP) {
|
||
segs.last()[1] = x - 1
|
||
} else {
|
||
segs.add(intArrayOf(segStart, x - 1))
|
||
}
|
||
inSeg = false
|
||
}
|
||
}
|
||
if (inSeg) {
|
||
if (segs.isNotEmpty() && (w - 1) - segs.last()[1] < DET_MERGE_GAP) {
|
||
segs.last()[1] = w - 1
|
||
} else {
|
||
segs.add(intArrayOf(segStart, w - 1))
|
||
}
|
||
}
|
||
for ((x0, x1) in segs.map { it[0] to it[1] }) {
|
||
// 过滤过小的框
|
||
val boxW = x1 - x0 + 1
|
||
val boxH = y1 - y0 + 1
|
||
if (boxW < 4 || boxH < 2) continue
|
||
// 映射回原图坐标（4 个角点）
|
||
val fx0 = x0 * ratioX
|
||
val fx1 = x1 * ratioX
|
||
val fy0 = y0 * ratioY
|
||
val fy1 = y1 * ratioY
|
||
boxes.add(listOf(
|
||
floatArrayOf(fx0, fy0),
|
||
floatArrayOf(fx1, fy0),
|
||
floatArrayOf(fx1, fy1),
|
||
floatArrayOf(fx0, fy1),
|
||
))
|
||
}
|
||
}
|
||
return boxes
|
||
}
|
||
|
||
// ============== CTC 解码 ==============
|
||
|
||
/**
|
||
* CTC greedy decode：每个时间步取 argmax，去 blank 去重复。返回 (text, avgConfidence)。
|
||
*
|
||
* PaddleOCR 约定：logits 的 index 0 固定是 blank，字符从 index 1 起，
|
||
* dictionary[i] 对应模型输出 index i+1。因此 dictIdx = argmaxIdx - 1。
|
||
*
|
||
* PP-OCRv6 模型输出已经是概率分布（值域 [0,1]），无需额外 softmax。
|
||
* 直接取 argmax 对应的值作为该时间步的置信度。
|
||
*/
|
||
private fun ctcGreedyDecode(logits: Array<FloatArray>): Pair<String, Float> {
|
||
if (logits.isEmpty()) return "" to 0f
|
||
val numClasses = logits[0].size
|
||
val blankIdx = 0 // PaddleOCR CTC：blank 固定在 index 0
|
||
|
||
val sb = StringBuilder()
|
||
var lastIdx = -1
|
||
var confSum = 0.0f
|
||
var confCount = 0
|
||
for (t in logits.indices) {
|
||
var maxIdx = 0
|
||
var maxVal = logits[t][0]
|
||
for (i in 1 until numClasses) {
|
||
if (logits[t][i] > maxVal) { maxVal = logits[t][i]; maxIdx = i }
|
||
}
|
||
|
||
if (maxIdx != blankIdx && maxIdx != lastIdx) {
|
||
val dictIdx = maxIdx - 1 // index 1..N → dictionary[0..N-1]
|
||
if (dictIdx in 0 until dictionary.size) {
|
||
sb.append(dictionary[dictIdx])
|
||
confSum += maxVal // v6 输出已是概率，直接用
|
||
confCount++
|
||
}
|
||
}
|
||
lastIdx = maxIdx
|
||
}
|
||
val avgConf = if (confCount > 0) confSum / confCount else 0f
|
||
return sb.toString() to avgConf
|
||
}
|
||
|
||
// ============== Bitmap 工具 ==============
|
||
|
||
private fun cropBox(bmp: Bitmap, box: List<FloatArray>): Bitmap? {
|
||
val xs = box.map { it[0] }
|
||
val ys = box.map { it[1] }
|
||
val paddingX = 4
|
||
val paddingY = 2
|
||
val minX = max(0, (xs.minOrNull() ?: 0f).toInt() - paddingX)
|
||
val minY = max(0, (ys.minOrNull() ?: 0f).toInt() - paddingY)
|
||
val maxX = min(bmp.width, ((xs.maxOrNull() ?: 0f) + 1).toInt() + paddingX)
|
||
val maxY = min(bmp.height, ((ys.maxOrNull() ?: 0f) + 1).toInt() + paddingY)
|
||
val w = maxX - minX
|
||
val h = maxY - minY
|
||
if (w < 2 || h < 2) return null
|
||
return Bitmap.createBitmap(bmp, minX, minY, w, h)
|
||
}
|
||
|
||
private fun resizeForDet(bmp: Bitmap, maxSide: Int): Bitmap {
|
||
val ratio = maxSide.toFloat() / max(bmp.width, bmp.height)
|
||
if (ratio >= 1f) return bmp
|
||
val newW = (bmp.width * ratio).toInt()
|
||
val newH = (bmp.height * ratio).toInt()
|
||
// 确保尺寸是 32 的倍数（det 模型下采样要求）
|
||
val alignedW = (newW / 32) * 32
|
||
val alignedH = (newH / 32) * 32
|
||
if (alignedW < 32 || alignedH < 32) return bmp
|
||
return Bitmap.createScaledBitmap(bmp, alignedW, alignedH, true)
|
||
}
|
||
|
||
// ============== 公共工具（复用） ==============
|
||
|
||
/** 解码 base64 图片为 Bitmap。 */
|
||
private fun decodeBase64(base64: String): Bitmap? {
|
||
return try {
|
||
// 去除 data:image/...;base64, 前缀
|
||
val data = if (base64.contains(",")) base64.substringAfter(",") else base64
|
||
val bytes = Base64.decode(data, Base64.DEFAULT)
|
||
BitmapFactory.decodeByteArray(bytes, 0, bytes.size)
|
||
} catch (e: Exception) {
|
||
Log.e(OCR_MODULE_NAME, "base64 解码失败: ${e.message}")
|
||
null
|
||
}
|
||
}
|
||
|
||
/**
|
||
* 短边压缩到 maxShortEdge（参考 AutoAccounting scaleDownForOcr）。
|
||
* 像素量比 1440p 减少约 75%，识别速度大幅提升。
|
||
*/
|
||
private fun scaleDownForOcr(bitmap: Bitmap, maxShortEdge: Int): Bitmap {
|
||
val width = bitmap.width
|
||
val height = bitmap.height
|
||
val shortEdge = minOf(width, height)
|
||
if (shortEdge <= maxShortEdge) return bitmap
|
||
val scale = maxShortEdge.toFloat() / shortEdge
|
||
val newWidth = (width * scale).toInt()
|
||
val newHeight = (height * scale).toInt()
|
||
return Bitmap.createScaledBitmap(bitmap, newWidth, newHeight, true)
|
||
}
|
||
|
||
override fun onCatalystInstanceDestroy() {
|
||
super.onCatalystInstanceDestroy()
|
||
release()
|
||
}
|
||
|
||
override fun invalidate() {
|
||
super.invalidate()
|
||
release()
|
||
}
|
||
|
||
private fun release() {
|
||
lock.lock()
|
||
try {
|
||
detSession?.close()
|
||
recSession?.close()
|
||
// OrtEnvironment 是单例，不主动 close（进程级）
|
||
detSession = null
|
||
recSession = null
|
||
ortEnv = null
|
||
initialized = false
|
||
} catch (_: Exception) {
|
||
} finally {
|
||
lock.unlock()
|
||
}
|
||
}
|
||
|
||
/** 识别结果块。 */
|
||
private data class OcrBlock(val text: String, val x: Float, val y: Float, val width: Float, val height: Float, val confidence: Float)
|
||
/** 预处理后的张量数据 + 宽高。 */
|
||
private data class TensorData(val data: FloatArray, val w: Int, val h: Int)
|
||
|
||
companion object {
|
||
/** OCR 最大短边（参考 AutoAccounting OCR_MAX_SHORT_EDGE）。 */
|
||
private const val OCR_MAX_SHORT_EDGE = 720
|
||
/** det resize 最大边（PaddleOCR limit_max_side_len 默认值）。 */
|
||
private const val DET_LIMIT_MAX_SIDE = 960
|
||
/** DB 二值化阈值。 */
|
||
private const val DET_THRESH = 0.3f
|
||
/** 投影法合并相邻文本段的间隔（像素）。 */
|
||
private const val DET_MERGE_GAP = 10
|
||
/** rec 固定图像高度。 */
|
||
private const val REC_IMAGE_HEIGHT = 48
|
||
/** rec 单行最大宽度。 */
|
||
private const val REC_MAX_WIDTH = 320
|
||
/** assets 中的模型/字典文件名。 */
|
||
private const val ASSET_DET_MODEL = "ppocrv6_det.onnx"
|
||
private const val ASSET_REC_MODEL = "ppocrv6_rec.onnx"
|
||
// PP-OCRv6 多语言识别模型的配套字典（18708 字符 + 运行时 1 blank = 18710 维输出）。
|
||
// 注意：必须与 rec 模型配套，错用 v5 字典（18383）会导致 CTC 解码丢字符。
|
||
private const val ASSET_DICT = "ppocrv6_dict.txt"
|
||
}
|
||
}
|