前边使用python api进行模型推理,这里使用opencv c++ api,模型选择还是前边提到的unet。
opencv 无法训练模型,但它支持载入其他深度学习框架训练好的模型,并使用该模型进行预测 inference;
OpenCV: Deep Neural Network module
1、图像预处理
将原始图像转换为可以直接输入网络的格式,在进行深度学习时,blobFromImage主要是用来对图片进行预处理。
dnn::blobFromImage(InputArray image,
double scalefactor=1.0,
const Size& size = Size(),
const Scalar& mean = Scalar(),
bool swapRB = false,
bool crop = false,
int ddepth = CV_32F)
- image:输入图像(1、3或者4通道)
- scalefactor:图像各通道数值的缩放比例
- size:输出图像的空间尺寸,如size=(200,300)表示高h=300,宽w=200
- mean:用于各通道减去的值,以降低光照的影响(e.g. image为bgr3通道的图像,mean=[104.0, 177.0, 123.0],表示b通道的值-104,g-177,r-123)
- swapRB:交换RB通道,默认为False.(cv2.imread读取的是彩图是bgr通道)
- crop:图像裁剪,默认为False.当值为True时,先按比例缩放,然后从中心裁剪成size尺寸
- ddepth:输出的图像深度,可选CV_32F 或者 CV_8U
2、模型加载
解析onnx文件
dnn::Net net = cv::dnn::readNetFromONNX(net_onnx);
3、模型推理
使用cuda
net.setPreferableBackend(cv::dnn::DNN_BACKEND_CUDA);
net.setPreferableTarget(cv::dnn::DNN_TARGET_CUDA);
单输入单输出
net.setInput(blob); //设置模型的输入
output = net.forward(); //前向传播
代码实现:
#include <iostream>
#include <vector>
#include<opencv2/opencv.hpp>
#include <opencv2/dnn.hpp>
#include <opencv2/dnn/all_layers.hpp>
using namespace cv;
using namespace std;
class unet
{
public:
unet(string modelpath);
void detect(Mat& frame);
private:
dnn::Net net;
const int inpWidth = 256;
const int inpHeight = 256;
void post_processing(Mat& out);
};
unet::unet(string modelpath)
{
this->net = dnn::readNetFromONNX(modelpath);
this->net.setPreferableBackend(cv::dnn::DNN_BACKEND_CUDA);
this->net.setPreferableTarget(cv::dnn::DNN_TARGET_CUDA);
}
void unet::detect(Mat& frame)
{
Mat blob;
dnn::blobFromImage(frame, blob, 1 / 255.0, Size(this->inpWidth, this->inpHeight),
Scalar(0, 0, 0), true, false);
this->net.setInput(blob);
Mat outs;
this->net.forward(outs);
post_processing(outs);
}
void unet::post_processing(Mat& outs)
{
std::cout << "Dimension information of outs:" << std::endl;
cout << outs.size[0] << "x" << outs.size[1] << "x" << outs.size[2] << endl;
outs = outs.reshape(1, { 256 , 256 });
cout << outs.size[0] << "x" << outs.size[1] << endl;
std::vector<std::string> CLASSES = { "ignore", "crack", "spall", "rebar" };
std::vector<cv::Vec3b> PALETTE = { cv::Vec3b(0, 0, 0), cv::Vec3b(255, 0, 0), cv::Vec3b(0, 255, 0), cv::Vec3b(0, 0, 255) }; // bgr
cv::Mat color_seg(outs.rows, outs.cols, CV_8UC3, cv::Scalar(0, 0, 0));
for (int row = 0; row < outs.rows; row++)
{
for (int col = 0; col < outs.cols; col++)
{
float label = outs.at<float>(row, col);
assert(label >= 0 && label < PALETTE.size());
cv::Vec3b color = PALETTE[label];
color_seg.at<cv::Vec3b>(row, col) = color;
}
}
imwrite("C:\\Users\\LYL\\Desktop\\a.jpg", color_seg);
}
int main()
{
unet unet_infer("D:/learning/project/inference_demo/cv_infer/model_softmax_argmax2.onnx");
string imgpath = "D:/learning/project/inference_demo/cv_infer/crack000680.jpg";
Mat srcimg = imread(imgpath);
unet_infer.detect(srcimg);
}