目录说明效果项目特点为什么选择 FP16适用场景测试体验部署建议时间不稳定设置C#调用源码下载说明在 OCR 项目落地中很多人都会遇到同一个问题模型效果不错但推理速度、部署便利性、GPU 兼容性、内存稳定性很难同时兼顾。最近我们基于 PP-OCR 系列模型完成了一套新的 TensorRT 推理版本lw.TensorRT.PPOCRSharp它采用 C 实现底层高性能推理 DLLC# 负责界面调用和业务集成适合 Windows 环境下对 OCR 速度有要求的场景。效果NVIDIA GeForce RTX4060 Laptop GPU (8 GB)CUDA 12.4TensorRT-10.16.1.11项目特点TensorRT GPU 推理加速底层使用 NVIDIA TensorRT 直接加载转换后的.engine文件进行推理相比通用推理框架在 NVIDIA GPU 上更容易获得更低延迟。C DLL 封装C# 简单调用核心推理逻辑封装在 C DLL 中C# 端通过 P/Invoke 调用既保证性能也方便 WinForms、WPF、业务系统集成。支持 PP-OCR 多版本模型当前测试程序中已加入多组模型选择PP-OCRv5 mobilePP-OCRv5 serverPP-OCRv6 tinyPP-OCRv6 smallPP-OCRv6 medium可以根据速度、精度、显存占用灵活选择。支持 det / rec / cls 三阶段 OCR完整支持文本检测、文本识别、方向分类模型调用。实际使用时也可以关闭 cls进一步提升速度。自动检测 TensorRT engine 文件程序会直接加载已经转换好的.engine文件。如果缺少 engineC# 程序会自动生成trtexec_convert_commands.txt说明文档提示用户如何使用本机trtexec.exe转换 ONNX 模型。因为不同电脑 TensorRT 安装路径不同文档中使用TRTEXEC占位用户只需要替换成本机路径即可。GPU ID 可配置界面保留use_gpu和gpu_id设置适合多显卡环境。底层 C 已经处理 CUDA device 绑定避免多线程推理时跑错 GPU。多 predictor 并发识别识别模型支持多个 predictor 并行处理适合一张图中存在多行文本的场景。实际部署时可以根据显卡性能调整rec_predictor_num并不是越大越好建议从 1、2、4 逐步测试。为什么选择 FP16TensorRT 支持 FP16 和 INT8但 OCR 模型尤其是 rec 识别模型对量化非常敏感。在实际测试中未经充分校准的 INT8 可能出现识别为空、准确率明显下降等问题。因此当前版本选择 FP16 作为默认和推荐方案速度提升明显精度更稳定转换流程简单更适合通用部署适用场景这套方案适合Windows 桌面 OCR 软件工业视觉字符识别票据、表单、截图 OCR批量图片文字识别C# 项目需要调用高性能 OCR DLLNVIDIA GPU 环境下追求更低延迟的 OCR 推理测试体验在 TensorRT FP16 engine 准备好之后程序启动后即可初始化模型并进行识别。典型流程选择 OCR 模型设置 GPU ID、batch、predictor 数量点击初始化选择图片点击识别程序会输出 det、rec、整体 OCR 耗时方便对不同模型、不同参数进行性能对比。部署建议为了获得更稳定的速度建议使用 NVIDIA 独立显卡设置 NVIDIA 电源管理模式为“最高性能优先”Windows 电源模式设置为高性能初始化后先预热几次再统计速度不建议跨机器复用 TensorRT engine不同显卡、驱动、CUDA、TensorRT 版本建议重新转换 engine时间不稳定设置1、查询GPU最高频率2、锁定GPU频率3、低延时模式设置为超高性4、电源管理模式设置为最高性能优先C#调用源码using Newtonsoft.Json; using OpenCvSharp; using System; using System.Collections.Generic; using System.Diagnostics; using System.Drawing; using System.IO; using System.Runtime.InteropServices; using System.Text; using System.Windows.Forms; namespace OCRTest { public partial class Form1 : Form { public Form1() { InitializeComponent(); } const string DllName lw.TensorRT.PPOCRSharp.dll; //初始化 [DllImport(DllName, EntryPoint init, CallingConvention CallingConvention.Cdecl)] public extern static int init(ref IntPtr engine , [MarshalAs(UnmanagedType.I1)] bool use_gpu , int gpu_id , string det_model_dir , int limit_side_len , double det_db_thresh , double det_db_box_thresh , double det_db_unclip_ratio , [MarshalAs(UnmanagedType.I1)] bool use_dilation , [MarshalAs(UnmanagedType.I1)] bool cls , [MarshalAs(UnmanagedType.I1)] bool use_angle_cls , string cls_model_dir , double cls_thresh , double cls_batch_num , string rec_model_dir , string rec_char_dict_path , int rec_batch_num , int rec_img_h , int rec_img_w , int predictor_num , StringBuilder msg); //识别 [DllImport(DllName, EntryPoint ocr, CallingConvention CallingConvention.Cdecl)] public extern static int ocr(IntPtr engine, IntPtr image, StringBuilder msg, out IntPtr ocr_result, out int ocr_result_len); //识别按图像内存传入避免托管层依赖 C Mat ABI [DllImport(DllName, EntryPoint ocr2, CallingConvention CallingConvention.Cdecl)] public extern static int ocr2(IntPtr engine, int rows, int cols, int channels, IntPtr data, StringBuilder msg, out IntPtr ocr_result, out int ocr_result_len); //释放 [DllImport(DllName, EntryPoint destroy, CallingConvention CallingConvention.Cdecl)] public extern static int destroy(IntPtr engine, StringBuilder msg); static IntPtr OCREngine; private Bitmap bmp; private String imgPath null; private ListOCRResult ltOCRResult; private string fileFilter *.*|*.bmp;*.jpg;*.jpeg;*.tiff;*.tif;*.png; private StringBuilder OCRResultInfo new StringBuilder(); private StringBuilder OCRResultAllInfo new StringBuilder(); Pen pen new Pen(Brushes.Red, 2f); /// summary /// 选择图片 /// /summary /// param namesender/param /// param namee/param private void button1_Click(object sender, EventArgs e) { OpenFileDialog ofd new OpenFileDialog(); ofd.Filter fileFilter; if (ofd.ShowDialog() DialogResult.OK) { imgPath ofd.FileName; bmp?.Dispose(); bmp new Bitmap(imgPath); pictureBox1.Image bmp; richTextBox1.Clear(); button2_Click(null, null); } } /// summary /// 识别 /// /summary /// param namesender/param /// param namee/param private void button2_Click(object sender, EventArgs e) { if (OCREngine IntPtr.Zero) { MessageBox.Show(请先初始化); return; } if (imgPath null) { MessageBox.Show(请先选择图片); return; } button1.Enabled false; button2.Enabled false; richTextBox1.Clear(); OCRResultInfo.Clear(); OCRResultAllInfo.Clear(); StringBuilder msgTemp new StringBuilder(128); Stopwatch stopwatch new Stopwatch(); stopwatch.Start(); IntPtr strPtr IntPtr.Zero; int ocr_result_len 0; string ocr_result string.Empty; int res -1; try { using (Mat img Cv2.ImRead(imgPath, ImreadModes.Color)) { res ocr2(OCREngine, img.Rows, img.Cols, img.Channels(), img.Data, msgTemp, out strPtr, out ocr_result_len); } if (strPtr ! IntPtr.Zero ocr_result_len 0) { byte[] buffer new byte[ocr_result_len]; Marshal.Copy(strPtr, buffer, 0, ocr_result_len); ocr_result Encoding.UTF8.GetString(buffer); } stopwatch.Stop(); double totalTime stopwatch.Elapsed.TotalMilliseconds; OCRResultAllInfo.AppendLine($耗时: {totalTime:F2}ms); OCRResultAllInfo.AppendLine(---------------------------); OCRResultInfo.AppendLine($耗时: {totalTime:F2}ms); OCRResultInfo.AppendLine(---------------------------); if (res 0) { ltOCRResult Newtonsoft.Json.JsonConvert.DeserializeObjectListOCRResult(ocr_result); OCRResultAllInfo.Append(JsonConvert.SerializeObject(ltOCRResult, Newtonsoft.Json.Formatting.Indented)); using (Graphics graphics Graphics.FromImage(bmp)) { foreach (OCRResult item in ltOCRResult) { OCRResultInfo.AppendLine(item.text); System.Drawing.Point[] pt new System.Drawing.Point[] { new System.Drawing.Point(item.x1, item.y1) , new System.Drawing.Point(item.x2, item.y2) , new System.Drawing.Point(item.x3, item.y3) , new System.Drawing.Point(item.x4, item.y4) }; graphics.DrawPolygon(pen, pt); } } if (checkBox1.Checked) { richTextBox1.Text OCRResultAllInfo.ToString(); } else { richTextBox1.Text OCRResultInfo.ToString(); } pictureBox1.Image null; pictureBox1.Image bmp; } else { MessageBox.Show(识别失败 msgTemp.ToString()); } } finally { stopwatch.Stop(); if (strPtr ! IntPtr.Zero) { Marshal.FreeCoTaskMem(strPtr); } button1.Enabled true; button2.Enabled true; } } /// summary /// 初始化 /// /summary /// param namesender/param /// param namee/param private void Form1_Load(object sender, EventArgs e) { rdov6tiny.Checked true; chkcls.Checked false; chkuse_gpu.Checked true; txtgpu_id.Text 0; LoadDefaultImage(); } private void checkBox1_CheckedChanged(object sender, EventArgs e) { richTextBox1.Clear(); if (checkBox1.Checked) { richTextBox1.Text OCRResultAllInfo.ToString(); } else { richTextBox1.Text OCRResultInfo.ToString(); } } private void chkcls_CheckedChanged(object sender, EventArgs e) { if (OCREngine IntPtr.Zero) { return; } richTextBox1.Text ; Console.WriteLine(cls设置已改变正在重新初始化模型...); UnloadModel(); LoadModel(); } private void radioButton1_CheckedChanged(object sender, EventArgs e) { RadioButton rb sender as RadioButton; } private void Form1_FormClosing(object sender, FormClosingEventArgs e) { UnloadModel(); } private void btnDestroy_Click(object sender, EventArgs e) { UnloadModel(); } void UnloadModel() { if (OCREngine ! IntPtr.Zero) { StringBuilder msgTemp new StringBuilder(128); destroy(OCREngine, msgTemp); Console.WriteLine(释放成功: msgTemp.ToString()); OCREngine IntPtr.Zero; } } private void btnInit_Click(object sender, EventArgs e) { if (OCREngine ! IntPtr.Zero) { StringBuilder msgTemp new StringBuilder(128); destroy(OCREngine, msgTemp); Console.WriteLine(释放成功: msgTemp.ToString()); OCREngine IntPtr.Zero; LoadModel(); } else { LoadModel(); } } void LoadModel() { StringBuilder msgTemp new StringBuilder(128); bool use_gpu true; int gpu_id 0; string det_model_dir ; int limit_side_len 960; double det_db_thresh 0.3; double det_db_box_thresh 0.6; double det_db_unclip_ratio 1.2; bool use_dilation false; bool cls true; bool use_angle_cls true; string cls_model_dir inference_trt/PP-OCRv5_mobile_cls_onnx.engine; double cls_thresh 0.9; int cls_batch_num 1; string rec_model_dir ; string rec_char_dict_path inference/ppocrv5_dict.txt; int rec_batch_num 8; int rec_img_h 48; int rec_img_w 320; int predictor_num 4; //赋值 if (chkuse_gpu.Checked true) { use_gpu true; } else { use_gpu false; } gpu_id Convert.ToInt32(txtgpu_id.Text.ToString()); det_db_thresh Convert.ToDouble(txtdet_db_thresh.Text.ToString()); det_db_box_thresh Convert.ToDouble(txtdet_db_box_thresh.Text.ToString()); det_db_unclip_ratio Convert.ToDouble(txtdet_db_unclip_ratio.Text.ToString()); if (chkcls.Checked true) { cls true; } else { cls false; } cls_batch_num Convert.ToInt32(txtcls_batch_num.Text.ToString()); rec_batch_num Convert.ToInt32(txtrec_batch_num.Text.ToString()); predictor_num Convert.ToInt32(txtpredictor_num.Text.ToString()); if (rdomobile.Checked) { det_model_dir BuildEnginePath(PP-OCRv5_mobile_det_onnx); rec_model_dir BuildEnginePath(PP-OCRv5_mobile_rec_onnx); cls_model_dir BuildEnginePath(PP-OCRv5_mobile_cls_onnx); } elseif (rdoserver.Checked) { det_model_dir BuildEnginePath(PP-OCRv5_server_det_infer); rec_model_dir BuildEnginePath(PP-OCRv5_server_rec_infer); cls_model_dir BuildEnginePath(PP-OCRv5_server_cls_infer); } elseif (rdov6small.Checked) { det_model_dir BuildEnginePath(PP-OCRv6_small_det); rec_model_dir BuildEnginePath(PP-OCRv6_small_rec); rec_char_dict_path inference/PP-OCRv6_small_rec_dict.txt; } elseif (rdov6medium.Checked) { det_model_dir BuildEnginePath(PP-OCRv6_medium_det); rec_model_dir BuildEnginePath(PP-OCRv6_medium_rec); rec_char_dict_path inference/PP-OCRv6_medium_rec_dict.txt; } else { det_model_dir BuildEnginePath(PP-OCRv6_tiny_det); rec_model_dir BuildEnginePath(PP-OCRv6_tiny_rec); rec_char_dict_path inference/PP-OCRv6_tiny_rec_dict.txt; } if (!EnsureTensorRtEngines(det_model_dir, rec_model_dir, cls ? cls_model_dir : null, rec_batch_num)) { return; } Console.WriteLine(TensorRT precision: FP16); Console.WriteLine(det model: det_model_dir); if (cls !File.Exists(Path.Combine(Application.StartupPath, cls_model_dir.Replace(/, Path.DirectorySeparatorChar)))) { Console.WriteLine(cls model missing, disabled: cls_model_dir); cls false; use_angle_cls false; } Console.WriteLine(cls model: (cls ? cls_model_dir : disabled)); Console.WriteLine(rec model: rec_model_dir); Console.WriteLine(rec dict : rec_char_dict_path); int res init(ref OCREngine , use_gpu , gpu_id , det_model_dir , limit_side_len , det_db_thresh , det_db_box_thresh , det_db_unclip_ratio , use_dilation , cls , use_angle_cls , cls_model_dir , cls_thresh , cls_batch_num , rec_model_dir , rec_char_dict_path , rec_batch_num , rec_img_h , rec_img_w , predictor_num , msgTemp); if (res 0) { Console.WriteLine(模型加载成功); } else { string msg msgTemp.ToString(); Console.WriteLine(模型加载失败 msg); } } private bool EnsureTensorRtEngines(string detEngine, string recEngine, string clsEngine, int recBatchNum) { Liststring missing new Liststring(); foreach (string engine in new[] { detEngine, recEngine, clsEngine }) { if (string.IsNullOrWhiteSpace(engine)) continue; string fullPath Path.Combine(Application.StartupPath, engine.Replace(/, Path.DirectorySeparatorChar)); if (!File.Exists(fullPath)) missing.Add(engine); } if (missing.Count 0) { returntrue; } StringBuilder sb new StringBuilder(); sb.AppendLine(TensorRT FP16 engine 转换说明); sb.AppendLine(); sb.AppendLine(); sb.AppendLine(当前缺少以下 engine 文件请先使用 trtexec.exe 转换。); sb.AppendLine(不同电脑的 trtexec.exe 路径可能不同请把下面命令中的 TRTEXEC 替换为本机 trtexec.exe 的完整路径或把 TensorRT bin 目录加入 PATH 后替换为 trtexec.exe。); sb.AppendLine(); sb.AppendLine(常见位置示例); sb.AppendLine(C:\TensorRT-xx\bin\trtexec.exe); sb.AppendLine(项目目录\TensorRT-10.16.1.11.Windows.amd64.cuda-12.9\TensorRT-10.16.1.11\bin\trtexec.exe); sb.AppendLine(); sb.AppendLine(建议在 exe 所在目录执行以下命令); sb.AppendLine(Application.StartupPath); sb.AppendLine(); foreach (string engine in missing) { string onnx EngineToOnnxPath(engine); string outDir Path.GetDirectoryName(Path.Combine(Application.StartupPath, engine.Replace(/, Path.DirectorySeparatorChar))); Directory.CreateDirectory(outDir); string shape BuildTrtShapeArgs(engine, recBatchNum); sb.AppendLine(缺少 engine); sb.AppendLine(对应 ONNX onnx); sb.AppendLine(转换命令); sb.AppendLine($\TRTEXEC\ --onnx\{onnx.Replace(/, Path.DirectorySeparatorChar)}\ --saveEngine\{engine.Replace(/, Path.DirectorySeparatorChar)}\ --fp16 {shape}); sb.AppendLine(); } sb.AppendLine(转换完成后重新点击初始化。); sb.AppendLine(注意engine 和显卡型号、驱动、CUDA、TensorRT 版本绑定不建议跨机器复用。); string commandFile Path.Combine(Application.StartupPath, trtexec_convert_commands.txt); try { File.WriteAllText(commandFile, sb.ToString(), Encoding.UTF8); sb.AppendLine(); sb.AppendLine(转换命令已写入); sb.AppendLine(commandFile); } catch (Exception ex) { sb.AppendLine(); sb.AppendLine(转换命令写入 txt 失败); sb.AppendLine(ex.Message); } MessageBox.Show(sb.ToString(), 缺少 TensorRT engine, MessageBoxButtons.OK, MessageBoxIcon.Information); Console.WriteLine(sb.ToString()); returnfalse; } private string BuildEnginePath(string modelName) { returninference_trt/ modelName _fp16.engine; } private string EngineToOnnxPath(string engine) { string file Path.GetFileNameWithoutExtension(engine); if (file.EndsWith(_fp16, StringComparison.OrdinalIgnoreCase)) { file file.Substring(0, file.Length - 5); } returninference/ file .onnx; } private string BuildTrtShapeArgs(string engine, int recBatchNum) { string name Path.GetFileNameWithoutExtension(engine).ToLowerInvariant(); if (name.Contains(rec)) { int optBatch Math.Max(1, recBatchNum); int maxBatch Math.Max(optBatch, 16); return $--minShapesx:1x3x48x32 --optShapesx:{optBatch}x3x48x320 --maxShapesx:{maxBatch}x3x48x1280; } if (name.Contains(cls)) { return--minShapesx:1x3x80x160 --optShapesx:8x3x80x160 --maxShapesx:16x3x80x160; } return--minShapesx:1x3x32x32 --optShapesx:1x3x960x960 --maxShapesx:1x3x960x960; } private void LoadDefaultImage() { string defaultImagePath Path.Combine(Application.StartupPath, 3.jpg); if (!File.Exists(defaultImagePath)) { defaultImagePath Path.Combine(AppDomain.CurrentDomain.BaseDirectory, 3.jpg); } if (!File.Exists(defaultImagePath)) { return; } imgPath defaultImagePath; bmp?.Dispose(); bmp new Bitmap(imgPath); pictureBox1.Image bmp; } } }下载通过网盘分享的文件lw.TensorRT.PPOCRSharp_Test 链接: https://pan.baidu.com/s/1bG5n0ZgJ7mW2fQ4dn9sjaQ 提取码: cmvf引入地址