cnn.cpp

//
// Created by marking on 2024/12/5.
//

#include "cnn.h"

void Figure::getMaxClass(const Mat& probBlob, int* classId, double* classProb)
{
    Mat probMat = probBlob.reshape(1, 1);
    Point classNumber;
    minMaxLoc(probMat, NULL, classProb, NULL, &classNumber);
    *classId = classNumber.x;
}


void Figure::cnn_load()
{
    //***************************************************加载模型********************************************************
    string model = "/Users/marking/Desktop/data/lenet_iter_10000.caffemodel";
    string config = "/Users/marking/Desktop/data/lenet_train_test.prototxt";

    try {
        net = readNetFromCaffe(config, model);
        cout << "cnn数字识别加载模型成功" << endl;
    }
    catch (cv::Exception& ee) {
        cerr << "错误: " << ee.what() << endl;
        if (net.empty()) {
            cout << "无法加载以下模型：" << endl;
            cout << "modelTxt: " << model << endl;
            cout << "modelBin: " << config << endl;
            exit(-1);
        }
    }
}
bool Figure::validateInput(const Mat& frame) {
    // ---------------------- 预处理增强 ----------------------
    Mat gray, binary;
    cvtColor(frame, gray, COLOR_BGR2GRAY);
    imshow("processed1",frame);
   // threshold(frame,)

    // 1. 自适应二值化（增强鲁棒性）
    adaptiveThreshold(gray, binary, 255, ADAPTIVE_THRESH_GAUSSIAN_C, THRESH_BINARY_INV, 11, 2);

    // 2. 形态学去噪（去除小斑点）
    Mat kernel = getStructuringElement(MORPH_ELLIPSE, Size(3, 3));
    morphologyEx(binary, binary, MORPH_OPEN, kernel);
    imshow("processed",binary);

    // 3. 轮廓分析（确保存在有效字符）
    vector<vector<Point>> contours;
    findContours(binary, contours, RETR_EXTERNAL, CHAIN_APPROX_SIMPLE);

    // 规则1：必须有且仅有一个显著轮廓
    if (contours.size() != 1) return false;

    // 规则2：最大轮廓面积占比超过10%
    double maxArea = contourArea(contours[0]);
    if (maxArea < (frame.rows * frame.cols * 0.1)) return false;

    // 规则3：宽高比在合理范围内（0.3~3.0，避免线状噪声）
    Rect bbox = boundingRect(contours[0]);
    float aspectRatio = (float)bbox.width / bbox.height;
    if (aspectRatio < 0.3 || aspectRatio > 3.0) return false;

    return true;
}

void Figure::cnn_recognition(const Mat& frame) {
    // ---------------------- 输入校验 ----------------------
    if (!validateInput(frame)) {
        cout << "无效输入（非数字图像）" << endl;
        return;
    }

    // ---------------------- 预处理标准化 ----------------------
    Mat processed;
    cvtColor(frame, processed, COLOR_BGR2GRAY);
    adaptiveThreshold(processed, processed, 255, ADAPTIVE_THRESH_GAUSSIAN_C, THRESH_BINARY_INV, 11, 2);
    resize(processed, processed, Size(28, 28));

    // ---------------------- 模型推理 ----------------------
    Mat inputBlob = blobFromImage(processed, 1.0, Size(28, 28), false);
    net.setInput(inputBlob, "data");
    Mat pred = net.forward("prob");

    // ---------------------- 后处理校验 ----------------------
    int classId;
    double classProb;
    getMaxClass(pred, &classId, &classProb);

    // 规则1：置信度阈值（可调整）
    const double CONF_THRESHOLD = 0.8;
    if (classProb < CONF_THRESHOLD) {
        cout << "低置信度（" << classProb * 100 << "%），拒绝识别" << endl;
        return;
    }

    // 规则2：概率分布熵检测（避免多个高概率类别）
    Mat probMat = pred.reshape(1, 1);
    Scalar mean, stddev;
    meanStdDev(probMat, mean, stddev);
    if (stddev[0] < 0.2) { // 标准差过小说明概率分布平坦
        cout << "不确定结果，拒绝识别" << endl;
        return;
    }

    // ---------------------- 最终输出 ----------------------
    cout << "识别结果: " << classId << " （置信度: " << classProb * 100 << "%）" << endl;
}
//********************************************old******************************************************
//void Figure::cnn_recognition(const Mat& frame)
//{
//    //***************************************************识别数字***********************************************************
//    double time0 = static_cast<double>(getTickCount());
//    Mat inputBlob = blobFromImage(frame, 1.0, Size(28, 28), false);
//    net.setInput(inputBlob, "data");
//    Mat pred = net.forward("prob");
//    time0 = ((double)getTickCount() - time0) / getTickFrequency();
//    //cout << "\t此方法运行时间为： " << cvRound(time0 * 1000) << "毫秒" << endl;
//    //cout << pred << endl;
//    int classId; double classProb;
//    Figure::getMaxClass(pred, &classId, &classProb);
//    cout << "Best Class: " << classId << endl;
//    //cout << "Probability: " << classProb * 100 << "%" << endl;
//}
//*******************************************************************************************************
void Figure::cnn_fps(const Mat& img)//在图片的左上角显示帧率
{
    double fps,t2,spendtime;
    char string[20];  // 用于存放帧率的字符串
    // 获取当前 tick count
    t2 = (double)getTickCount();
    // 计算从 t1 到 t2 的时间
    spendtime = (t2 - t1) / getTickFrequency();
    fps = 1.0 / spendtime;
    sprintf(string, "%.2f", fps); // 帧率保留两位小数
    std::string fpsString("FPS: ");
    fpsString += string; // 在 "FPS:" 后加入帧率数值字符串
    // 将帧率信息写在输出帧上
    putText(img, fpsString, Point(5, 20), FONT_HERSHEY_SIMPLEX, 0.5, Scalar(255, 255, 255));
    // 更新 t1 为当前的 tick count
    t1 = t2;
}