How to do super resolution image with onnx and opencv using c++ ?, its working fine using python

Question

I have animesr.onnx for my super resolution ios/macos app so I need to convert my python codes to c++ codes. These is my python codes:

session = onnxruntime.InferenceSession('animesr.onnx')

img = cv2.imread('imgs/naruto.jpg')
ori_h, ori_w, _ = img.shape
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
img = cv2.resize(img, (512,512))
img = (np.array(img) / 255.0).astype(np.float32)
img = np.transpose(img, (2, 0, 1))
img = np.expand_dims(img, 0)

input_name = session.get_inputs()[0].name
output_name = session.get_outputs()[0].name

input_feed = {input_name: img}
output = session.run([output_name], input_feed)

output = output[0].clip(0, 1) * 255
output = output.astype(np.uint8)
output = np.squeeze(output)
output = np.transpose(output, (1, 2, 0))
output = cv2.cvtColor(output, cv2.COLOR_RGB2BGR)
output = cv2.resize(output, (ori_w*4, ori_h*4))
cv2.imwrite('naruto_animesr.jpg', output)

Its working fine and the result is like this (left is before, right is after)

And this is my c++ codes:

Ort::Session session(env, ORT_TSTR(modelPath), sessionOptions);

cv::Mat inputImage = cv::imread(imagePath, cv::IMREAD_COLOR);
// cv::Mat blob = cv::dnn::blobFromImage(inputImage, 1.0/255, cv::Size(512,512), cv::Scalar(), true);
cv::Mat resizedImage;
cv::resize(inputImage, resizedImage, cv::Size(512,512));
cv::Mat floatImage;
resizedImage.convertTo(floatImage, CV_32FC3, 1.0/255.0);

Ort::MemoryInfo memoryInfo = Ort::MemoryInfo::CreateCpu(OrtArenaAllocator, OrtMemTypeDefault);
std::vector<int64_t> inputShape = {1, 3, 512, 512};
  
Ort::Value inputTensor = Ort::Value::CreateTensor<float>(memoryInfo, (float*) floatImage.data, 3*512*512, inputShape.data(), inputShape.size());

std::vector<const char*> inputNames = {"input"};
std::vector<const char*> outputNames = {"output"};

std::vector<Ort::Value> outputTensor = session.Run(Ort::RunOptions{}, inputNames.data(), &inputTensor, 1, outputNames.data(), outputNames.size());

Ort::TensorTypeAndShapeInfo outputInfo = outputTensor[0].GetTensorTypeAndShapeInfo();
int channels = outputInfo.GetShape()[1]; // 3
int height = outputInfo.GetShape()[2]; // 2048
int width = outputInfo.GetShape()[3]; // 2048
  
const float* outputData = outputTensor[0].GetTensorMutableData<float>();
  
cv::Mat outputImage(height, width, CV_32FC(channels), const_cast<float*>(outputData));
cv::Mat uint8Image;
outputImage.convertTo(uint8Image, CV_8UC3, 255);
cv::Mat bgrOutput;
cv::cvtColor(uint8Image, bgrOutput, cv::COLOR_RGB2BGR);
cv::imwrite(outputPath, bgrOutput);

This is the result and this is the result with cv::ddn::blobFromImage()

Both are suck results, different with my python result.

I thing the problem is from pre-processing image as input before run the model and processing the output to become cv image, I am bad at matrix operation like transform, transpose, reshape, etc.

Can you guys help me provide the correct processing image codes with opencv c++

Thank you