RapidAI · xinshoushangdao · Aug 6, 2026 · Aug 7, 2026
diff --git a/.codegraph/.gitignore b/.codegraph/.gitignore
@@ -0,0 +1,16 @@
+# CodeGraph data files
+# These are local to each machine and should not be committed
+
+# Database
+*.db
+*.db-wal
+*.db-shm
+
+# Cache
+cache/
+
+# Logs
+*.log
+
+# Hook markers
+.dirty
diff --git a/RapidOCRConsole/Assets/1.png b/RapidOCRConsole/Assets/1.png
diff --git a/RapidOCRConsole/Program.cs b/RapidOCRConsole/Program.cs
@@ -1,46 +1,110 @@
 
-using Emgu.CV;
 using RapidOCRLib;
-using System.Diagnostics;
-using System.Drawing.Imaging;
-using System.Threading.Tasks;
 
-var path = Path.Combine(AppContext.BaseDirectory, "models");
-//Initialize model.
-OcrLite ocrEngin = new OcrLite()
+var winformModels = Path.Combine(AppContext.BaseDirectory, "..", "..", "..", "..", "RapidOCRWinform", "models");
+winformModels = Path.GetFullPath(winformModels);
+var detModel = Path.Combine(winformModels, "PP-OCRv6_det_medium.onnx");
+var clsModel = Path.Combine(winformModels, "ch_ppocr_mobile_v2.0_cls_infer.onnx");
+var recModel = Path.Combine(winformModels, "PP-OCRv6_rec_medium.onnx");
+var dictFile = Path.Combine(winformModels, "ppocrv6_dict.txt");
+var imgFile = Path.Combine(AppContext.BaseDirectory, "Assets", "demo.png");
+
+const int padding = 50;
+const float boxScoreThresh = 0.5f;
+const float boxThresh = 0.3f;
+const float unClipRatio = 1.6f;
+const bool mostAngle = false;
+
+TextWriter cout = Console.Out;
+void Log(string msg) { cout.WriteLine(msg); cout.Flush(); }
+
+int optimalThr = (int)(Environment.ProcessorCount * 0.7);
+var configs = new (int ThreadNum, int MaxSideLen, bool DoAngle, bool UseGpu, string Label)[]
 {
-    DetPath = Path.Combine(path, "ch_PP-OCRv5_mobile_det.onnx"),
-    ClsPath = Path.Combine(path, "ch_ppocr_mobile_v2.0_cls_infer.onnx"),
-    RecPath = Path.Combine(path, "ch_PP-OCRv5_rec_mobile_infer.onnx"),
-    KeyDicPath = Path.Combine(path, "ppocrv5_dict.txt"),
+    (ThreadNum: optimalThr, MaxSideLen: 1024, DoAngle: true,  UseGpu: false, Label: $"A[CPU]: thr={optimalThr} | max=1024 | doAngle=T (Baseline)"),
+    (ThreadNum: optimalThr, MaxSideLen: 800,  DoAngle: false, UseGpu: false, Label: $"D[CPU]: thr={optimalThr} | max=800  | doAngle=F (最优CPU配置)"),
+    (ThreadNum: optimalThr, MaxSideLen: 800,  DoAngle: false, UseGpu: true,  Label: $"G[GPU]: max=800 | doAngle=F (与D同参数，GPU加速对比)"),
 };
+const int warmupRuns = 1;
+const int testRuns = 2;
+
+Log($"Process64={Environment.Is64BitProcess}, ProcessorCount={Environment.ProcessorCount}, OS={Environment.OSVersion}");
+Log("Ready.");
 
-await ocrEngin.InitModels();
-//get the image will be ocr.
-var demoWillOCRFile = Path.Combine(AppContext.BaseDirectory, "Assets", "demo.png");
-//ocr detect
-var result = ocrEngin.Detect(demoWillOCRFile, 50);
-if (result != null)
+var results = new List<(string Label, double AvgTotal, double Db, double Ang, double Rec, int Blocks, string Engine)>();
+foreach (var cfg in configs)
 {
-    Console.WriteLine(result.ToString());
-    var img = result.BoxImg.ToBitmap();
-    var desFile = Path.Combine(AppContext.BaseDirectory!, "Assets", Path.GetFileNameWithoutExtension(Path.GetRandomFileName()) + ".png")!;
-    #pragma warning disable CA1416
-    img?.Save(desFile!);
-    #pragma warning restore CA1416
-    if (File.Exists(desFile))
+    Log("");
+    Log("=========================================================================");
+    Log($"CONFIG {cfg.Label}");
+    Log("=========================================================================");
+    var ocr = new OcrLite
     {
-        ProcessStartInfo psi = new ProcessStartInfo
-        {
-            FileName = desFile,
-            UseShellExecute = true
-        };
+        DetPath = detModel,
+        ClsPath = clsModel,
+        RecPath = recModel,
+        KeyDicPath = dictFile,
+        ThreadNum = cfg.ThreadNum,
+    };
+    await ocr.InitModels(ocr.DetPath, ocr.ClsPath, ocr.RecPath, ocr.KeyDicPath, cfg.ThreadNum, cfg.UseGpu, gpuDeviceId: 0);
+    // 初始化后立即取一次引擎状态（Detect 会把 EngineProvider 写入结果）
+    Log($"InitModels done. UseGpu={cfg.UseGpu}, Begin warmup ({warmupRuns}).");
 
-        Process.Start(psi);
+    // 预热
+    for (int w = 0; w < warmupRuns; w++)
+    {
+        var r0 = ocr.Detect(imgFile, padding, cfg.MaxSideLen, boxScoreThresh, boxThresh, unClipRatio, cfg.DoAngle, mostAngle);
+        if (r0 != null) Log($"  Warmup DetectTime={r0.DetectTime:F0}ms (Det={r0.DbNetTime:F0}ms Ang={r0.AngleNetTime:F0}ms Rec={r0.CrnnNetTime:F0}ms)");
     }
-    Console.WriteLine("显示strResult");
-    Console.WriteLine(result.StrRes);
+    Log("Warmup done. Begin tests.");
+
+    List<double> totals = new(), dbs = new(), angs = new(), recs = new();
+    int blocks = 0;
+    string engineLine = "unknown";
+    for (int i = 0; i < testRuns; i++)
+    {
+        var r = ocr.Detect(imgFile, padding, cfg.MaxSideLen, boxScoreThresh, boxThresh, unClipRatio, cfg.DoAngle, mostAngle);
+        if (r == null) continue;
+        engineLine = r.EngineProvider ?? "n/a";
+        Log($"  Run {i + 1}/{testRuns}: total={r.DetectTime:F0}ms  Det={r.DbNetTime:F0}ms  Ang={r.AngleNetTime:F0}ms  Rec={r.CrnnNetTime:F0}ms  Blocks={r.TextBlocks.Count}  [{engineLine}]");
+        totals.Add(r.DetectTime);
+        dbs.Add(r.DbNetTime);
+        angs.Add(r.AngleNetTime);
+        recs.Add(r.CrnnNetTime);
+        blocks = r.TextBlocks.Count;
+    }
+    results.Add((cfg.Label, totals.Average(), dbs.Average(), angs.Average(), recs.Average(), blocks, engineLine));
 }
 
+Log("");
+Log("#########################################################################");
+Log("  SUMMARY REPORT (demo.png, PP-OCRv6_medium, CPU vs GPU)");
+Log("#########################################################################");
+Log(string.Format("{0,-58} | {1,8} | {2,8} | {3,8} | {4,8} | {5,7} | {6,8} | {7}",
+    "Config", "Total(ms)", "Det(ms)", "Ang(ms)", "Rec(ms)", "Blocks", "SpeedUp", "Engine"));
+Log("-------------------------------------------------------------------------");
+double baselineTotal = results[0].AvgTotal;
+foreach (var r in results)
+{
+    double speedup = baselineTotal > 0 ? baselineTotal / r.AvgTotal : double.NaN;
+    Log(string.Format("{0,-58} | {1,8:F0} | {2,8:F0} | {3,8:F0} | {4,8:F0} | {5,7} | x{6:F2} | {7}",
+        r.Label, r.AvgTotal, r.Db, r.Ang, r.Rec, r.Blocks, speedup, r.Engine));
+}
+Log("#########################################################################");
 
-Console.ReadKey();
+// 如果 GPU 方案仍显示为 CPU，给用户排查清单
+var gpuResult = results.FirstOrDefault(x => x.Label.StartsWith("G["));
+if (!string.IsNullOrEmpty(gpuResult.Engine) && !gpuResult.Engine.Contains("DML("))
+{
+    Log("");
+    Log("! GPU 方案没有真正启用 DirectML，原因已在 Engine 列打印。按下面步骤排查：");
+    Log("  1) 确认 Windows 版本 ≥ 10.0.1903（Win+R → winver）");
+    Log("  2) 更新显卡驱动到最新版（NVIDIA/AMD/Intel 官网都有）");
+    Log("  3) 本项目已 PlatformTarget=x64，确认编译输出为 x64 而非 32 位");
+    Log("  4) 确认 RapidOCRLib.csproj 引用了 Microsoft.ML.OnnxRuntime.DirectML 1.24.4");
+}
+else if (!string.IsNullOrEmpty(gpuResult.Engine) && gpuResult.Engine.Contains("DML("))
+{
+    Log("");
+    Log($">> DirectML GPU 启用成功！与最优 CPU 方案 D 对比：{baselineTotal / results[1].AvgTotal:F2}x(CPU内部) → {baselineTotal / gpuResult.AvgTotal:F2}x(含GPU)");
+}
diff --git a/RapidOCRConsole/RapidOCRConsole.csproj b/RapidOCRConsole/RapidOCRConsole.csproj
@@ -1,10 +1,13 @@
-<Project Sdk="Microsoft.NET.Sdk">
+<Project Sdk="Microsoft.NET.Sdk">
 
   <PropertyGroup>
     <OutputType>Exe</OutputType>
     <TargetFramework>net10.0</TargetFramework>
     <ImplicitUsings>enable</ImplicitUsings>
     <Nullable>enable</Nullable>
+    <!-- CUDA / cuDNN 原生库仅提供 x64 版本 -->
+    <PlatformTarget>x64</PlatformTarget>
+    <Prefer32Bit>false</Prefer32Bit>
   </PropertyGroup>
 
   <ItemGroup>

diff --git a/RapidOCRLib/AngleNet.cs b/RapidOCRLib/AngleNet.cs
@@ -1,4 +1,4 @@
-using Emgu.CV;
+using Emgu.CV;
 using Emgu.CV.CvEnum;
 using Emgu.CV.Structure;
 using Microsoft.ML.OnnxRuntime;
@@ -22,28 +22,29 @@ class AngleNet
         private InferenceSession angleNet;
         private List<string> inputNames;
 
+        public string ProviderInfo { get; private set; } = "NotInitialized";
+
         public AngleNet() { }
 
         ~AngleNet()
         {
-            angleNet.Dispose();
+            angleNet?.Dispose();
         }
 
-        public async Task InitModel(string path, int numThread)
+        public async Task InitModel(string path, int numThread, bool useGpu = false, int gpuDeviceId = 0)
         {
             try
             {
-                SessionOptions op = new SessionOptions();
-                op.GraphOptimizationLevel = GraphOptimizationLevel.ORT_ENABLE_EXTENDED;
-                op.InterOpNumThreads = numThread;
-                op.IntraOpNumThreads = numThread;
+                SessionOptions op = OcrUtils.CreateSessionOptions(OcrUtils.NetKind.Classification, numThread, useGpu, gpuDeviceId, out var providerInfo);
+                ProviderInfo = providerInfo;
+                Console.WriteLine($"[AngleNet] Loading model with provider: {providerInfo}");
                 angleNet = new InferenceSession(path, op);
                 inputNames = angleNet.InputMetadata.Keys.ToList();
 
                 // 从 ONNX 模型 metadata 读取输入尺寸，兼容 v2 (192x48) 和 v5 (160x80)
                 var dims = angleNet.InputMetadata.First().Value.Dimensions;
-                _dstHeight = dims[2];
-                _dstWidth = dims[3];
+                _dstHeight = dims.Length > 2 && dims[2] > 0 ? dims[2] : 48;
+                _dstWidth = dims.Length > 3 && dims[3] > 0 ? dims[3] : 192;
 
                 await Task.CompletedTask;
             }

diff --git a/RapidOCRLib/CrnnNet.cs b/RapidOCRLib/CrnnNet.cs
@@ -1,4 +1,4 @@
-using Emgu.CV;
+using Emgu.CV;
 using Microsoft.ML.OnnxRuntime;
 using Microsoft.ML.OnnxRuntime.Tensors;
 using RapidOCRLib.Models;
@@ -23,24 +23,25 @@ class CrnnNet
         private List<string> keys;
         private List<string> inputNames;
 
+        public string ProviderInfo { get; private set; } = "NotInitialized";
+
         public CrnnNet() { }
 
         ~CrnnNet()
         {
             crnnNet?.Dispose();
         }
 
-        public async Task InitModel(string path, string keysPath, int numThread)
+        public async Task InitModel(string path, string keysPath, int numThread, bool useGpu = false, int gpuDeviceId = 0)
         {
             try
             {
-                SessionOptions op = new SessionOptions();
-                op.GraphOptimizationLevel = GraphOptimizationLevel.ORT_ENABLE_EXTENDED;
-                op.InterOpNumThreads = numThread;
-                op.IntraOpNumThreads = numThread;
+                SessionOptions op = OcrUtils.CreateSessionOptions(OcrUtils.NetKind.Recognition, numThread, useGpu, gpuDeviceId, out var providerInfo);
+                ProviderInfo = providerInfo;
+                Console.WriteLine($"[CrnnNet] Loading model with provider: {providerInfo}");
                 crnnNet = new InferenceSession(path, op);
                 inputNames = crnnNet.InputMetadata.Keys.ToList();
-                keys = InitKeys(keysPath);
+                keys = InitKeys(crnnNet, keysPath);
                 await Task.CompletedTask;
             }
             catch (Exception ex)
@@ -49,7 +50,36 @@ public async Task InitModel(string path, string keysPath, int numThread)
                 throw;
             }
         }
-        private List<string> InitKeys(string path)
+
+        private List<string> InitKeys(InferenceSession session, string keysPath)
+        {
+            // PP-OCRv5/v6 rec models embed the character dictionary in ONNX metadata.
+            if (session.ModelMetadata.CustomMetadataMap.TryGetValue("character", out string characters))
+            {
+                List<string> keys = new List<string>();
+                keys.Add("#");
+                string[] lines = characters.Split('\n');
+                int count = lines.Length;
+                if (count > 0 && lines[count - 1].Length == 0)
+                {
+                    count--;
+                }
+                for (int i = 0; i < count; i++)
+                {
+                    keys.Add(lines[i]);
+                }
+                keys.Add(" ");
+                Console.WriteLine($"keys Size = {keys.Count}");
+                return keys;
+            }
+            if (string.IsNullOrWhiteSpace(keysPath))
+            {
+                throw new Exception("The rec model does not embed a character dictionary and no key dictory file was provided.");
+            }
+            return InitKeysFromFile(keysPath);
+        }
+
+        private List<string> InitKeysFromFile(string path)
         {
             StreamReader sr = new StreamReader(path, Encoding.UTF8);
             List<string> keys = new List<string>();
@@ -149,4 +179,4 @@ private TextLine ScoreToTextLine(float[] srcData, int h, int w)
         }
 
     }
-}
+}
diff --git a/RapidOCRLib/DbNet.cs b/RapidOCRLib/DbNet.cs
@@ -1,4 +1,4 @@
-using Clipper2Lib;
+using Clipper2Lib;
 using Emgu.CV;
 using Emgu.CV.CvEnum;
 using Emgu.CV.Structure;
@@ -23,21 +23,22 @@ class DbNet
 
         private List<string> inputNames;
 
+        public string ProviderInfo { get; private set; } = "NotInitialized";
+
         public DbNet() { }
 
         ~DbNet()
         {
-            dbNet.Dispose();
+            dbNet?.Dispose();
         }
 
-        public async Task InitModel(string path, int numThread)
+        public async Task InitModel(string path, int numThread, bool useGpu = false, int gpuDeviceId = 0)
         {
             try
             {
-                SessionOptions op = new SessionOptions();
-                op.GraphOptimizationLevel = GraphOptimizationLevel.ORT_ENABLE_EXTENDED;
-                op.InterOpNumThreads = numThread;
-                op.IntraOpNumThreads = numThread;
+                SessionOptions op = OcrUtils.CreateSessionOptions(OcrUtils.NetKind.Detection, numThread, useGpu, gpuDeviceId, out var providerInfo);
+                ProviderInfo = providerInfo;
+                Console.WriteLine($"[DbNet] Loading model with provider: {providerInfo}");
                 dbNet = new InferenceSession(path, op);
                 inputNames = dbNet.InputMetadata.Keys.ToList();
                 await Task.CompletedTask;

diff --git a/RapidOCRLib/Models/ModeOptions.cs b/RapidOCRLib/Models/ModeOptions.cs
@@ -1,6 +1,4 @@
-using System;
-using System.Collections.Generic;
-using System.Text;
+using System;
 
 namespace RapidOCRLib.Models
 {
@@ -26,5 +24,13 @@ public class ModeOptions
         /// cpu thread number,default is 70% of toal logic cpu core numbers.
         /// </summary>
         public int ThreadNum { get; set; } = (int)(Environment.ProcessorCount * 0.7);
+        /// <summary>
+        /// Whether to use GPU acceleration via DirectML (Windows built-in DX12 GPU acceleration).
+        /// </summary>
+        public bool UseGpu { get; set; } = false;
+        /// <summary>
+        /// GPU device ID to use when UseGpu is enabled. Default is 0.
+        /// </summary>
+        public int GpuDeviceId { get; set; } = 0;
     }
 }
diff --git a/RapidOCRLib/Models/OcrResult.cs b/RapidOCRLib/Models/OcrResult.cs
@@ -1,4 +1,4 @@
-using Emgu.CV;
+using Emgu.CV;
 using System.Collections.Generic;
 using System.Drawing;
 using System.Text;
@@ -70,16 +70,26 @@ public sealed class OcrResult
     {
         public List<TextBlock> TextBlocks { get; set; }
         public float DbNetTime { get; set; }
+        public float AngleNetTime { get; set; }
+        public float CrnnNetTime { get; set; }
+        public float PreprocessTime { get; set; }
+        public float PostprocessTime { get; set; }
         public Mat BoxImg { get; set; }
         public float DetectTime { get; set; }
         public string StrRes { get; set; }
+        /// <summary>当前推理使用的后端引擎信息，如 "DML(device 0)" / "CPU"。</summary>
+        public string EngineProvider { get; set; } = "Unknown";
 
         public override string ToString()
         {
             StringBuilder sb = new StringBuilder();
-            sb.AppendLine("OcrResult");
+            sb.AppendLine($"OcrResult [Engine: {EngineProvider}]");
             TextBlocks.ForEach(x => sb.Append(x));
+            sb.AppendLine($"├─PreprocessTime({PreprocessTime}ms)");
             sb.AppendLine($"├─DbNetTime({DbNetTime}ms)");
+            sb.AppendLine($"├─AngleNetTime({AngleNetTime}ms)");
+            sb.AppendLine($"├─CrnnNetTime({CrnnNetTime}ms)");
+            sb.AppendLine($"├─PostprocessTime({PostprocessTime}ms)");
             sb.AppendLine($"├─DetectTime({DetectTime}ms)");
             sb.AppendLine($"└─StrRes({StrRes})");
             return sb.ToString();