Qt篇——简单调用yolov3模型识别常见物品

下面仅为一个简单的举例：

一、首先需要引入opencv库。这个就不细讲了，各位可以自行百度。

#在pro文件中引入opencv库，我是用的opencv454 mingw64位的
opencv {
    win32 {
        INCLUDEPATH += $$PWD/3rdparty/opencv454/include
        INCLUDEPATH += $$PWD/3rdparty/opencv454/include/opencv2
        LIBS += $$PWD/3rdparty/opencv454/libopencv_world454.dll.a
    }
}
#这只是一种引入方式，具体大家可以网上搜索，或者评论区留言我给opencv454库下载链接

二、下载yolov3的模型文件yolov3.weights、yolov3.cfg、coco.names。

        网上搜索可以下载到，或者在评论区留言我给模型下载链接

三、代码调用（yolov3模型文件和识别的图片1.png放在exe同级目录下）

dnntest.h头文件

#ifndef DNNTEST_H
#define DNNTEST_H
#include <QString>
#include <opencv2/dnn.hpp>
#include <fstream>

class DnnTest
{
public:
    DnnTest();

    std::vector<std::string> getClasses(const QString &classesFile) {
        std::vector<std::string> classes;
        std::ifstream ifs(classesFile.toStdString().c_str());
        std::string line;
        while (getline(ifs, line)) {
            classes.push_back(line);
        }
        return classes;
    }

    cv::dnn::Net loadModel(QString modelConfiguration, QString modelWeights) {
        return cv::dnn::readNetFromDarknet(modelConfiguration.toStdString(), modelWeights.toStdString());
    }
    void startTest(const cv::Mat &frame, cv::dnn::Net &net, const std::vector<std::string> &classes);
};

#endif // DNNTEST_H

dnntest.cpp源文件

#include "dnntest.h"
#include <QString>
#include <QDebug>
#include <opencv2/imgcodecs.hpp>
#include <QCoreApplication>
#include <QDir>

DnnTest::DnnTest()
{
    QString exeDir = QCoreApplication::applicationDirPath();
    QString filePathPic = QDir(exeDir).filePath("1.png");
    QString filePathClasses = QDir(exeDir).filePath("coco.names");
    QString filePathModelWeights = QDir(exeDir).filePath("yolov3.weights");
    QString filePathModelConfiguration = QDir(exeDir).filePath("yolov3.cfg");
    cv::Mat image = cv::imread(filePathPic.toStdString());
    auto net = loadModel(filePathModelConfiguration, filePathModelWeights);
    auto classes = getClasses(filePathClasses);
    startTest(image, net, classes);
}

void DnnTest::startTest(const cv::Mat &frame, cv::dnn::Net &net, const std::vector<std::string> &classes) {
    cv::Mat blob;
    cv::dnn::blobFromImage(frame, blob, 1 / 255.0, cv::Size(416, 416), cv::Scalar(0, 0, 0), true, false);
    net.setInput(blob);
    std::vector<cv::Mat> outs;
    net.forward(outs, net.getUnconnectedOutLayersNames());
    // 后处理：解析检测结果
    for (const auto &output : outs) {
        auto *data = reinterpret_cast<float*>(output.data);
        for (int j = 0; j < output.rows; ++j, data += output.cols) {
            float confidence = data[4];
            if (confidence > 0.1) { // 过滤低置信度的物体
                int classId = -1;
                float maxConfidence = -1;
                for (int k = 5; k < output.cols; ++k) {
                    if (data[k] > maxConfidence) {
                        maxConfidence = data[k];
                        classId = k - 5;
                    }
                }
                if (classId >= 0 && maxConfidence > 0.1) {
                    QString label = QString::fromStdString(classes[classId]);
                    qDebug() << "x,y,w,h" << data[0] << data[1] << data[2] << data[3];
                    qDebug() << "Detected object:" << label << "with confidence:" << maxConfidence;
                }
            }
        }
    }
}

日志打印如下：

x,y,w,h 0.200591 0.493517 0.17878 0.380748
Detected object: "apple" with confidence: 0.916478
x,y,w,h 0.19109 0.520369 0.193458 0.385598
Detected object: "apple" with confidence: 0.930474
x,y,w,h 0.199505 0.521505 0.17687 0.35982
Detected object: "apple" with confidence: 0.960644
x,y,w,h 0.190564 0.542987 0.198303 0.37466
Detected object: "apple" with confidence: 0.366576
x,y,w,h 0.198655 0.542039 0.19229 0.374069
Detected object: "apple" with confidence: 0.748606

注：x,y,w,h的数值都是基于图片的比例，demo识别的图片1.png内容是：桌上有一个苹果。