Describe the bug
After upgrading OpenCvSharp to latest version, text recognition is poor in windows. Returned postions do not correspond to image, it looks like image is mirrored before OCR.
Steps to reproduce the bug
Program.cs:
builder.Services.AddSingleton(s => {
RecognizationModel model_rec = RecognizationModel.FromDirectoryV5("latin_PP-OCRv5_mobile_rec_infer");
ClassificationModel model_cls = ClassificationModel.FromDirectory("PP-LCNet_x1_0_textline_ori_infer",ModelVersion.V5);
DetectionModel model_det = DetectionModel.FromDirectory("PP-OCRv5_server_det",ModelVersion.V5);
FullOcrModel model = new(model_det,model_cls,model_rec);
return new QueuedPaddleOcrAll(() => new PaddleOcrAll(model,PaddleDevice.Mkldnn()) {
Enable180Classification = true,
AllowRotateDetection = true,
},consumerCount: 1);
}
);
Controller:
public sealed class OcrLines(QueuedPaddleOcrAll ocr) {
async Task RunAsync(byte[] image) {
using Mat imgSrc = Mat.FromImageData(image);
var result = await ocr.Run(imgSrc);
}
}
Use latest OpenCvSharp nuget package version 4.13.0.20260213 in Windowsi.
Attached image contains text
EESTI PAKENDIRINGLUS OÜ
It is recognized as
EEATKEIDIBIICAO AE
It looks like text is tried to recognized from mirrored image.
Expected behavior
Text should recognized as
EESTI PAKENDIRINGLUS OÜ
If previous OpenCvSharp version 4.11.0.20250507 is used, OCR works properly.
Screenshots
No response
Release version
<PackageReference Include="Sdcb.PaddleInference" Version="3.0.1" />
<PackageReference Include="Sdcb.PaddleInference.runtime.linux-x64" Version="3.1.0.54" />
<PackageReference Include="Sdcb.PaddleInference.runtime.win64.mkl" Version="3.1.0.54" />
<PackageReference Include="Sdcb.PaddleOCR" Version="3.0.1" />
<PackageReference Include="Sdcb.PaddleOCR.Models.LocalV5" Version="3.0.0" />
IDE
Microsoft Visual Studio Community
January 2026 Feature Update
Version: 18.2.2
OS version
Windows 11
Additional context
In Debian Linux where OpenCv is complied from source, using latest OpenCvSharp recognizes text properly.
Describe the bug
After upgrading OpenCvSharp to latest version, text recognition is poor in windows. Returned postions do not correspond to image, it looks like image is mirrored before OCR.
Steps to reproduce the bug
Program.cs:
Controller:
Use latest OpenCvSharp nuget package version 4.13.0.20260213 in Windowsi.
Attached image contains text
It is recognized as
It looks like text is tried to recognized from mirrored image.
Expected behavior
Text should recognized as
If previous OpenCvSharp version 4.11.0.20250507 is used, OCR works properly.
Screenshots
No response
Release version
IDE
Microsoft Visual Studio Community
January 2026 Feature Update
Version: 18.2.2
OS version
Windows 11
Additional context
In Debian Linux where OpenCv is complied from source, using latest OpenCvSharp recognizes text properly.