opencv DNN模块之Faster RCNN(Tensflow)检测物体

上一节写的是使用caffe的SSD模型来识别物体，能检测20种，本节使用tensflow的faster rcnn模型基于COCO数据集能检测80个类别，代码很相似

原理

本人不太懂，自行百度

---

每种网络模型对应的输入数据

可以查看每种模型的模型二进制文件名称，网络描述文件名称，中至化处理参数，数据样本尺寸，描述label文件名称，rgb通道顺序以及典型应用场景等信息
链接地址:https://github.com/opencv/opencv/blob/master/samples/dnn/models.yml

################################################################################
# Object detection models.
################################################################################

。。。。

# TensorFlow implementation of Faster-RCNN model from https://github.com/tensorflow/models/tree/master/research/object_detection
faster_rcnn_tf:
  model: "faster_rcnn_inception_v2_coco_2018_01_28.pb"
  config: "faster_rcnn_inception_v2_coco_2018_01_28.pbtxt"
  mean: [0, 0, 0]
  scale: 1.0
  width: 800
  height: 600
  rgb: true
  sample: "object_detection"

。。。。

模型文件下载地址：https://github.com/opencv/opencv/wiki/TensorFlow-Object-Detection-API

---

网络输入输出

• 输入图像 [1x3x300x300]通道顺序：RGB、均值为0、放缩：1.0
• 输出数据 [1x1xNx7] 输出1通道 N行 7列数据，N代表N种检测出的可能性，表述如下：
  [image_id,label,conf,x_min,y_min,x_max,ymax]
  [当前图片索引，识别对应标签，置信度，对应矩形框比例位置]

---

代码

#include <opencv2/opencv.hpp>
#include <opencv2/dnn.hpp>
#include <iostream>

using namespace std;
using namespace cv;
using namespace cv::dnn;

#define PIC_PATH "/work/opencv_pic/"
#define PIC_NAME "horse.jpg"

const size_t width = 300;
const size_t height = 300;
string bin_model = "/work/opencv_dnn/faster_rcnn/frozen_inference_graph.pb";
string protxt = "/work/opencv_dnn/faster_rcnn/faster_rcnn_resnet50_coco_2018_01_28.pbtxt";
string label_map = "/work/opencv_dnn/faster_rcnn/mscoco_label_map.pbtxt";
std::map<int, string> readLabelMaps(void);
int dnn_fast_rcnn(void)
{
    string pic = string(PIC_PATH)+string(PIC_NAME);
    Mat src;
    src = imread(pic);
    if(src.empty())
    {
        printf("pic read err\n");
        return -1;
    }

    namedWindow("input image",WINDOW_AUTOSIZE);
    //创建并载入神经网络
    //Net net = readNetFromCaffe(protxt,bin_model);
    Net net = readNetFromTensorflow(bin_model,protxt);
    if(net.empty())
    {
        printf("read caffe model data err\n");
        return -1;
    }


    //设置计算后台
    net.setPreferableBackend(DNN_BACKEND_OPENCV);
    net.setPreferableTarget(DNN_TARGET_CPU);

    //获取各层信息
    vector<string> layers_names = net.getLayerNames();

    for(size_t i=0;i<layers_names.size();i++)
    {
        int id = net.getLayerId(layers_names[i]);
        auto layer = net.getLayer(id);
        printf("layer id:%d,type:%s,name:%s\n",id,layer->type.c_str(),layer->name.c_str());
    }

    //图片格式转换
    Mat blobimage = blobFromImage(src,1, Size(width, height), Scalar(127.5, 127.5, 127.5), true, false);

    //网络输入数据
    net.setInput(blobimage);

    //获取识别数据
    Mat detection = net.forward();

    //获取名称索引
    map<int, string> names = readLabelMaps();

    //图像reshap 输出图像1x1xNx7  中的N参数代表多少种可能 7列参数 类型为浮点型
    Mat detectionMat(detection.size[2],detection.size[3],CV_32F,detection.ptr<float>());
    RNG rng(123785);
    float confidence_threshold = 0.5;   //过滤阈值  有N种可能 尽量过滤不需要的
    for(int i=0;i<detectionMat.rows;i++)
    {
        float confidence = detectionMat.at<float>(i,2);   //获取可能性
        if(confidence>confidence_threshold)
        {
            size_t objindex = (size_t)(detectionMat.at<float>(i,1))+1;   //获取label索引
            cout << "objindex:" << objindex <<endl;

            //获取矩形框的坐标  获取到的是相对图片的比例值  这里应该放缩获取真实值
            //左上坐标
            float tl_x = detectionMat.at<float>(i,3)*src.cols;
            float tl_y = detectionMat.at<float>(i,4)*src.rows;

            //右下坐标
            float br_x = detectionMat.at<float>(i,5)*src.cols;
            float br_y = detectionMat.at<float>(i,6)*src.rows;
            //绘制矩形框
            Rect object_box((int)(tl_x),(int)(tl_y),(int)(br_x-tl_x),(int)(br_y-tl_y));
            rectangle(src,object_box, Scalar(rng.uniform(0,255),rng.uniform(0,255), rng.uniform(0,255)),2,8,0);

            map<int, string>::iterator it = names.find(objindex);
            //显示识别对象名称
            putText(src,(it->second).c_str(),Point(tl_x, tl_y), FONT_HERSHEY_SIMPLEX, 1.0, Scalar(rng.uniform(0,255),rng.uniform(0,255), rng.uniform(0,255)), 2);
        }
    }
    imshow("input image",src);

    waitKey(0);
    destroyAllWindows();
    return 0;
}

std::map<int, string> readLabelMaps()
{
        std::map<int, string> labelNames;
        std::ifstream fp(label_map);
        if (!fp.is_open())
        {
                printf("could not open file...\n");
                exit(-1);
        }
        string one_line;
        string display_name;
        while (!fp.eof())
        {
                std::getline(fp, one_line);
                std::size_t found = one_line.find("id:");
                if (found != std::string::npos) {
                        int index = found;
                        string id = one_line.substr(index + 4, one_line.length() - index);

                        std::getline(fp, display_name);
                        std::size_t  found = display_name.find("display_name:");

                        index = found + 15;
                        string name = display_name.substr(index, display_name.length() - index);
                        name = name.replace(name.length() - 1, name.length(), "");
                        // printf("id : %d, name: %s \n", stoi(id.c_str()), name.c_str());
                        labelNames[stoi(id)] = name;
                }
        }
        fp.close();
        return labelNames;
}

---

效果

---