samples/dnn/text_detection.cpp

/*
文本检测模型：https://github.com/argman/EAST
下载链接：https://www.dropbox.com/s/r2ingd0l3zt8hxs/frozen_east_text_detection.tar.gz?dl=1
 
文本识别模型可在此处直接下载
下载链接：https://drive.google.com/drive/folders/1cTbQ3nuZG-EKWak6emD_s8_hHXWz7lAr?usp=sharing
以及 doc/tutorials/dnn/dnn_text_spotting/dnn_text_spotting.markdown
 
如何从 pb 转换为 onnx
使用此处的类：https://github.com/meijieru/crnn.pytorch/blob/master/models/crnn.py
import torch
from models.crnn import CRNN
model = CRNN(32, 1, 37, 256)
model.load_state_dict(torch.load('crnn.pth'))
dummy_input = torch.randn(1, 1, 32, 100)
torch.onnx.export(model, dummy_input, "crnn.onnx", verbose=True)
 
更多信息，请参考 doc/tutorials/dnn/dnn_text_spotting/dnn_text_spotting.markdown 和 doc/tutorials/dnn/dnn_OCR/dnn_OCR.markdown
*/
#include <iostream>
#include <fstream>
 
#include <opencv2/imgproc.hpp>
#include <opencv2/highgui.hpp>
#include <opencv2/dnn.hpp>
 
using namespace cv;
using namespace cv::dnn;
 
const char* keys =
    "{ help h | | 打印帮助信息。 }"
    "{ input i | | 输入图像或视频文件的路径。跳过此参数即可从摄像头捕捉帧。 }"
    "{ detModel dmp | | 包含已训练检测器网络的二进制 .pb 文件路径。 }"
    "{ width | 320 | 通过调整大小到特定宽度来预处理输入图像。它应该是 32 的倍数。 }"
    "{ height | 320 | 通过调整大小到特定高度来预处理输入图像。它应该是 32 的倍数。 }"
    "{ thr | 0.5 | 置信度阈值。 }"
    "{ nms | 0.4 | 非最大抑制阈值。 }"
    "{ recModel rmp | | 包含已训练 CRNN 文本识别模型的二进制 .onnx 文件路径。 "
        "下载链接在 doc/tutorials/dnn/dnn_text_spotting/dnn_text_spotting.markdown 中提供 }"
    "{ RGBInput rgb |0| 0：使用 flags=IMREAD_GRAYSCALE 的 imread；1：使用 flags=IMREAD_COLOR 的 imread。 }"
    "{ vocabularyPath vp | alphabet_36.txt | 用于评估的基准路径。 "
        "下载链接在 doc/tutorials/dnn/dnn_text_spotting/dnn_text_spotting.markdown 中提供 }";
 
void fourPointsTransform(const Mat& frame, const Point2f vertices[], Mat& result);
 
int main(int argc, char** argv)
{
    // 解析命令行参数。
    CommandLineParser parser(argc, argv, keys);
parser.about("使用此脚本运行 TensorFlow 实现 (https://github.com/argman/EAST) 的 "
                 "EAST：高效且准确的场景文本检测器 (https://arxiv.org/abs/1704.03155v2)");
    if (argc == 1 || parser.has("help"))
    {
parser.printMessage();
        return 0;
    }
 
    float confThreshold = parser.get<float>("thr");
    float nmsThreshold = parser.get<float>("nms");
    int width = parser.get<int>("width");
    int height = parser.get<int>("height");
    int imreadRGB = parser.get<int>("RGBInput");
    String detModelPath = parser.get<String>("detModel");
    String recModelPath = parser.get<String>("recModel");
    String vocPath = parser.get<String>("vocabularyPath");
 
    if (!parser.check())
    {
parser.printErrors();
        return 1;
    }
 
    // 加载网络。
    CV_Assert(!detModelPath.empty() && !recModelPath.empty());
    TextDetectionModel_EAST detector(detModelPath);
detector.setConfidenceThreshold(confThreshold)
.setNMSThreshold(nmsThreshold);
 
    TextRecognitionModel recognizer(recModelPath);
 
    // 加载词汇表
    CV_Assert(!vocPath.empty());
std::ifstream vocFile;
vocFile.open(samples::findFile(vocPath));
    CV_Assert(vocFile.is_open());
    String vocLine;
std::vector<String> vocabulary;
    while (std::getline(vocFile, vocLine)) {
vocabulary.push_back(vocLine);
    }
recognizer.setVocabulary(vocabulary);
recognizer.setDecodeType("CTC-greedy");
 
    // 识别参数
    double recScale = 1.0 / 127.5;
    Scalar recMean = Scalar(127.5, 127.5, 127.5);
    Size recInputSize = Size(100, 32);
recognizer.setInputParams(recScale, recInputSize, recMean);
 
    // 检测参数
    double detScale = 1.0;
    Size detInputSize = Size(width, height);
    Scalar detMean = Scalar(123.68, 116.78, 103.94);
    bool swapRB = true;
detector.setInputParams(detScale, detInputSize, detMean, swapRB);
 
    // 打开视频文件或图像文件或摄像头流。
    VideoCapture cap;
    bool openSuccess = parser.has("input") ? cap.open(parser.get<String>("input")) : cap.open(0);
    CV_Assert(openSuccess);
 
    static const std::string kWinName = "EAST：一种高效且准确的场景文本检测器";
 
    Mat frame;
    while (waitKey(1) < 0)
    {
cap >> frame;
        if (frame.empty())
        {
            waitKey();
            break;
        }
 
std::cout << frame.size << std::endl;
 
        // 检测
std::vector< std::vector<Point> > detResults;
detector.detect(frame, detResults);
        Mat frame2 = frame.clone();
        if (detResults.size() > 0) {
            // 文本识别
            Mat recInput;
            if (!imreadRGB) {
                cvtColor(frame, recInput, cv::COLOR_BGR2GRAY);
} else {
recInput = frame;
            }
std::vector< std::vector<Point> > contours;
            for (uint i = 0; i < detResults.size(); i++)
            {
                const auto& quadrangle = detResults[i];
                CV_CheckEQ(quadrangle.size(), (size_t)4, "");
 
contours.emplace_back(quadrangle);
 
std::vector<Point2f> quadrangle_2f;
                for (int j = 0; j < 4; j++)
quadrangle_2f.emplace_back(quadrangle[j]);
 
                Mat cropped;
fourPointsTransform(recInput, &quadrangle_2f[0], cropped);
 
std::string recognitionResult = recognizer.recognize(cropped);
std::cout << i << ": '" << recognitionResult << "'" << std::endl;
 
                putText(frame2, recognitionResult, quadrangle[3], FONT_HERSHEY_SIMPLEX, 1.5, Scalar(0, 0, 255), 2);
            }
            polylines(frame2, contours, true, Scalar(0, 255, 0), 2);
        }
        imshow(kWinName, frame2);
    }
    return 0;
}
 
void fourPointsTransform(const Mat& frame, const Point2f vertices[], Mat& result)
{
    const Size outputSize = Size(100, 32);
 
    Point2f targetVertices[4] = {
        Point(0, outputSize.height - 1),
        Point(0, 0), Point(outputSize.width - 1, 0),
        Point(outputSize.width - 1, outputSize.height - 1)
    };
    Mat rotationMatrix = getPerspectiveTransform(vertices, targetVertices);
 
    warpPerspective(frame, result, rotationMatrix, outputSize);
}