高性能百度OCR ONNX Runtime C#实现

RapidOCRSharpOnnx 是一个基于OpenCV与ONNX Runtime 实现的PaddleOCR C#开源推理库，参考了RapidOCR 的实现，重新使用C#实现，并做了大量性能优化与架构重新设计，方便多种ONNX Runtime 执行提供程序部署

特性

• 不依赖于具体的模型，直接导入需要的PP-OCR模型，识别指定的语言，PP-OCRv5模型支持识别106种语言，只需要将PP-OCRv5模型导出为onnx格式，就可以离线运行
• 支持的ONNX Runtime 执行提供程序：Execution Provider CPU, CUDA / TensorRT, OpenVINO, CoreML, DirectML
• 支持批量识别，并提供异步执行性能优化，大幅提高批量执行性能
• 使用OpenCVSharp4 进行图像处理操作
• 绘制识别结果图片使用SkiaSharp
• PP-OCR 版本支持：PP-OCRv5, PP-OCRv4
• 推理引擎：ONNX Runtime 是一个跨平台的机器学习模型加速器

OCR 示例

 

 

 

RapidOCRSharpOnnx使用

1 导出模型为onnx格式

如何转换PP-OCR模型为onnx格式，可以参考PP-OCR官网Obtaining ONNX Models, 或者直接从RapidOCR的魔塔社区下载Model List.

2 安装RapidOCRSharpOnnx组件并加载模型

Install Nuget packages RapidOCRSharpOnnx, OnnxRuntime, OpenCvSharp4.runtime

 

CPU 推理

dotnet add package RapidOCRSharpOnnx
dotnet add package OpenCvSharp4.runtime.win
dotnet add package Microsoft.ML.OnnxRuntime

using RapidOCRSharp ocr = new RapidOCRSharp(new ExecutionProviderCPU(new OcrConfig(detectPath, recPath, LangRec.CH, OCRVersion.PPOCRV5, clsMobilePath)));

CoreML 推理

dotnet add package RapidOCRSharpOnnx
dotnet add package OpenCvSharp4.runtime.osx.10.15-x64
dotnet add package Microsoft.ML.OnnxRuntime

using RapidOCRSharp ocr = new RapidOCRSharp(new ExecutionProviderCoreML(new OcrConfig(detectPath, recPath, LangRec.CH, OCRVersion.PPOCRV5, clsMobilePath)));

CUDA/TensorRT 推理

dotnet add package RapidOCRSharpOnnx
dotnet add package OpenCvSharp4.runtime.win
dotnet add package Microsoft.ML.OnnxRuntime.Gpu.Windows

using RapidOCRSharp ocr = new RapidOCRSharp(new ExecutionProviderCUDA(new OcrConfig(detectPath, recogPath, LangRec.CH, OCRVersion.PPOCRV5, clsPath), deviceId));
using RapidOCRSharp ocr = new RapidOCRSharp(new ExecutionProviderTensorRT(new OcrConfig(detectPath, recogPath, LangRec.CH, OCRVersion.PPOCRV5, clsPath), deviceId));

DirectML 推理

dotnet add package RapidOCRSharpOnnx
dotnet add package OpenCvSharp4.runtime.win
dotnet add package Microsoft.ML.OnnxRuntime.DirectML

using RapidOCRSharp ocr = new RapidOCRSharp(new ExecutionProviderDirectML(new OcrConfig(detectPath, recogPath, LangRec.EN, OCRVersion.PPOCRV5, clsPath), deviceId));

OpenVINO 推理

 

dotnet add package RapidOCRSharpOnnx
dotnet add package OpenCvSharp4.runtime.win
dotnet add package Intel.ML.OnnxRuntime.OpenVino

using RapidOCRSharp ocr = new RapidOCRSharp(new ExecutionProviderOpenVINO(new OcrConfig(detectPath, recogPath, LangRec.EN, OCRVersion.PPOCRV5, clsPath), IntelDeviceType.NPU));

 

基本的API使用

using RapidOCRSharp ocr = new RapidOCRSharp(new ExecutionProviderDirectML(new OcrConfig(detectPath, recogPath, LangRec.EN, OCRVersion.PPOCRV5, clsPath), _deviceId));
string savePath = $"res_{Path.GetFileName(imgPath)}";
var result = ocr.RecognizeText(imgPath, savePath);
Console.WriteLine($"result: {result.ToString()}");

批量识别图片

 using RapidOCRSharp ocr = new RapidOCRSharp(new ExecutionProviderDirectML(new OcrConfig(detectPath, recogPath, LangRec.CH, OCRVersion.PPOCRV5, clsPath), _deviceId));
 var list = Directory.GetFiles(@"C:\code\model\OCRTestImages");
 Stopwatch sw = new Stopwatch();
 sw.Start();
 var resPath = ocr.BatchParallelAsync(list.ToList(), saveDir, receiveAction: ReceiveResult);
 sw.Stop();
 Console.WriteLine($"BatchAsync Time: {sw.ElapsedMilliseconds} ms");

private static void ReceiveResult(OcrBatchResult batchResult)
{
    Console.WriteLine(batchResult.ToString());
    Console.WriteLine("------------------------------------------------------------");
}

批量识别使用Foreach API

        private static async Task TestBatchForeachAsync()
        {

            string detectPath = @"D:\code\RapidOCRSharpOnnx\RapidOCRSharpOnnx.TestCommon\Models\ch_PP-OCRv5_det_mobile.onnx";
            string recogPath = @"D:\code\RapidOCRSharpOnnx\RapidOCRSharpOnnx.TestCommon\Models\ch_PP-OCRv5_rec_mobile.onnx";
            string clsPath = @"D:\code\RapidOCRSharpOnnx\RapidOCRSharpOnnx.TestCommon\Models\ch_PP-LCNet_x0_25_textline_ori_cls_mobile.onnx";

            using RapidOCRSharp ocr = new RapidOCRSharp(new ExecutionProviderDirectML(new OcrConfig(detectPath, recogPath, LangRec.CH, OCRVersion.PPOCRV5, clsPath), _deviceId));
            var list = Directory.GetFiles(@"D:\code\model\OCRTestImages");
            var res = ocr.BatchForeachAsync(list.ToList(), @"D:\code\model\OCRTestImagesResults");

            await foreach (var item in res)
            {
                Console.WriteLine(item.TextBlocks);
            }

        }

 

性能测试

OCR组件库性能对比测试 CPU推理测试

OCR library	Version	language	Inference Engine
PaddleSharp	3.0.1	Paddle Inference C API .NET binding	Sdcb.PaddleInference
PaddleOCR	3.5.0	python	paddlepaddle
RapidOCR	3.8.1	python	openvino
RapidOCRSharpOnnx	1.0.0	C#	Intel.ML.OnnxRuntime.OpenVino

 

测试电脑

Windows 11 Pro OS Version 25H2

CPU： Intel Core Ultra 9 285k 3.7GHz

内存：DDR5 128GB speed 4400MT/s

硬盘：SSD 2TB

 

测试数据

图片: 60 张图片 (图片大小: 1180x92)

PP-OCR 模型: ch_PP-OCRv5_det_mobile, ch_PP-OCRv5_rec_mobile, ch_PP-LCNet_x0_25_textline_ori_cls_mobile

PaddleSharp 测试结果

CPU 推理时间 : 48.1769278s

 

PaddleOCR 测试结果

CPU 推理时间 : 62.6685s

 

RapidOCR 测试结果

CPU 推理时间 : 17.9634s

 

RapidOCRSharpOnnx 测试结果

CPU 推理时间 : 9.2447s

 

性能测试结果

OCR library	Version	language	Inference Engine	Elapsed Time
PaddleSharp	3.0.1	Paddle Inference C API .NET binding	Sdcb.PaddleInference.runtime.win64.mkl version 3.1.0.54 CPU	48.1769s
PaddleOCR	3.5.0	python	paddlepaddle version 3.2.0 CPU	62.6685s
RapidOCR	3.8.1	python	openvino version 2026.1.0 21367 CPU	17.9634s
RapidOCRSharpOnnx	1.0.0	C#	Intel.ML.OnnxRuntime.OpenVino CPU version 1.24.1	9.2447s