采用ONNX格式的deeplabv3+网络进行语义分割的C++代码

以下是采用ONNX格式的deeplabv3+网络进行语义分割的C++代码：

#include <onnxruntime_cxx_api.h>
#include <opencv2/opencv.hpp>

int main() {

    // Load the ONNX model
    Ort::SessionOptions session_options;
    Ort::Env env(ORT_LOGGING_LEVEL_WARNING, "test");
    Ort::Session session(env, "deeplabv3plus.onnx", session_options);

    // Get the input and output names and shapes
    Ort::AllocatorWithDefaultOptions allocator;
    std::vector<const char*> input_names = {"image"};
    std::vector<const char*> output_names = {"output"};
    std::vector<int64_t> input_shape = {1, 3, 512, 512};
    std::vector<int64_t> output_shape;
    Ort::TypeInfo output_type_info = session.GetOutputTypeInfo(0);
    auto output_tensor_info = output_type_info.GetTensorTypeAndShapeInfo();
    output_tensor_info.GetShape(output_shape);

    // Create the input and output tensors
    auto input_tensor = Ort::Value::CreateTensor<float>(allocator, input_shape.data(), input_shape.size());
    auto output_tensor = Ort::Value::CreateTensor<float>(allocator, output_shape.data(), output_shape.size());

    // Read the input image
    cv::Mat image = cv::imread("test.png");
    cv::resize(image, image, cv::Size(512, 512));
    image.convertTo(image, CV_32FC3, 1.f / 255.f);
    cv::cvtColor(image, image, cv::COLOR_BGR2RGB);

    // Copy the image data to the input tensor
    float* input_data = input_tensor->GetTensorMutableData<float>();
    memcpy(input_data, image.data, sizeof(float) * input_shape[1] * input_shape[2] * input_shape[3]);

    // Run the inference
    session.Run(Ort::RunOptions{nullptr}, input_names.data(), &input_tensor, 1, output_names.data(), 1, &output_tensor);

    // Copy the output tensor data to an OpenCV Mat
    cv::Mat output_image(output_shape[2], output_shape[3], CV_32FC1, output_tensor->GetTensorMutableData<float>());
    cv::Mat output_image_normalized;
    cv::normalize(output_image, output_image_normalized, 0, 255, cv::NORM_MINMAX, CV_8UC1);

    // Display the output image
    cv::imshow("Output", output_image_normalized);
    cv::waitKey(0);

    return 0;
}

这段代码使用了 ONNX Runtime C++ API 来加载 deeplabv3+ 模型并进行推理。输入图像的尺寸为 512x512，输出为一个概率图，每个像素点的值表示属于每个类别的概率。在这里，我们将输出图像归一化到了 0-255 的范围内，并将其展示出来。