Cat and dog identification application based on Openvino (c++)

1. Code

Link: https://pan.baidu.com/s/1nvgtKIRVPOEa0pJazW8GHQ extraction code: qhkp

Or: https://github.com/caip1299920300/Openvino_dog_cat

2. Environment configuration

1. Opencv environment configuration

opencv version: opencv340

You can use the following tutorial (you can also refer to: VS+OpenCV+OpenVINO2022 detailed configuration - Zhihu (zhihu.com) )

(1) Right-click on My Computer -> Properties -> Advanced System Settings -> Environment Variables -> System Variables -> Path -> Add D:\OpenCV\build\x64\vc15\bin (note that it is based on your own OpenCV Path, the path given here is my path and cannot be copied and used directly. What is added is the bin folder path in the build of the opencv installation directory)

(2) Copy opencv_world420.dll and opencv_world420d.dll in the path: D:\OpenCV\build\x64\vc15\bin\ to the C:\Windows\SysWOW64 directory

(3) Copy opencv_videoio_ffmpeg420_64.dll in the path: D:\OpenCV\build\bin\ to the C:\Windows\System32 directory

If you can't find Windows, you can click View and check the hidden items.

2. Create a new project in VS and configure the project

配置VC++ 目录的”包含目录“和”库目录“
”包含目录“:
C:\Program Files(x86)\Intel\openvino_2022.1.0.643\runtime\include
C:\Program Files(x86)\Intel\openvino_2022.1.0.643\runtime\include\openvino
C:\Program Files(x86)\Intel\openvino_2022.1.0.643\runtime\include\ngraph
C:\Program Files(x86)\Intel\openvino_2022.1.0.643\runtime\include\ie
E:\opencv\build\include
E:\opencv\build\include\opencv2
”库目录“:
E:\opencv\build\x64\vc15\lib
C:\ProgramFiles (x86)\Intel\openvino_2022.1.0.643\runtime\lib\intel64\Release

配置”链接器“的输入：
openvino.lib
openvino_onnx_frontend.lib
openvino_ir_frontend.lib
opencv_world340.lib

三、运行代码

#include <openvino/openvino.hpp>
#include <iostream>
#include <vector>
#include <opencv2/opencv.hpp>


int main() {
    std::string img_path = "E:/project/VS2017/dog_cat/img/dog.jpg";    // 预测图片
    std::string onnx_path = "E:/project/VS2017/dog_cat/model/Cat_dog.onnx"; // 预测模型
    size_t input_batch_size = 1;    // 输入图片的batch_size
    size_t num_channels = 3;        // 输入通道
    size_t h = 224;                    // 输入图片的高
    size_t w = 224;                    // 输入图片的宽
    clock_t startTime, endTime;        // 推理时间记录变量

    // 0、创建IE插件，查询支持硬件设备
    ov::Core core;
    //获取当前支持的所有的AI硬件推理设备
    std::vector<std::string> devices = core.get_available_devices();
    for (int i = 0; i < devices.size(); i++) {
        std::cout << devices[i] << std::endl;
    }
    // 1、加载检测模型
    // 模型加载并编译
    ov::CompiledModel compiled_model = core.compile_model(onnx_path, "AUTO");
    // 创建用于推断已编译模型的推理请求对象  创建的请求分配了输入和输出张量
    ov::InferRequest infer_request = compiled_model.create_infer_request();    
    
    // 2、请求网络输入
    auto input_tensor = infer_request.get_input_tensor(0);

    // 3、指定shape的大小
    input_tensor.set_shape({ input_batch_size, num_channels, w, h }); 
    // 4、获取输入的地址，并传递给指针input_data_host
    float* input_data_host = input_tensor.data<float>();  

    // 对应于pytorch的代码部分
    // 推理开始时间
    startTime = clock();
    // opencv读取图片
    cv::Mat src = cv::imread(img_path);
    int image_height = src.rows;
    int image_width = src.cols;
    // 修改图片大小
    cv::Mat image;
    cv::resize(src, image, cv::Size(w, h));
    int image_area = image.cols * image.rows;
    unsigned char* pimage = image.data;
    float* phost_b = input_data_host + image_area * 0;   // input_data_host和phost_*进行地址关联
    float* phost_g = input_data_host + image_area * 1;
    float* phost_r = input_data_host + image_area * 2;
    // BGR->RGB
    float mean[] = { 0.406, 0.456, 0.485 };
    float std[] = { 0.225, 0.224, 0.229 };
    for (int i = 0; i < image_area; ++i, pimage += 3) {
        // 注意这里的顺序rgb调换了
        *phost_r++ = pimage[0] / 255.;  // 将图片中的像素点进行减去均值除方差，并赋值给input
        *phost_g++ = pimage[1] / 255.;
        *phost_b++ = pimage[2] / 255.;
    }
    
    
    // 5、执行预测
    infer_request.infer();
    // 6、推理结果
    auto output = infer_request.get_output_tensor(0);
    // 对输出结果处理
    float* prob = output.data<float>();
    const int num_classes = 2; // 种类
    int predict_label = std::max_element(prob, prob + num_classes) - prob;  // 确定预测类别的下标
    std::string label;
    if (predict_label == 0)
        label = "cat";
    else
        label = "dog";
    float confidence = prob[predict_label];    // 获得预测值的置信度
    printf("confidence = %f, label = %s\n", confidence, label);
    endTime = clock();//计时结束
    std::cout << "total推理时间: " << (double)(endTime - startTime) / CLOCKS_PER_SEC << "s" << std::endl;

    return 0;
}