PaddleOCR2.7+Qt5

章节一：Windows 下的 PIP 安装

官网安装教程地址

按照里面的教程去安装

如果使用cuda版本的还要安装tensorrt，不然后面运行demo程序的程序会报如下错。

下载TensorRT 8版本，tensorrt下载地址

章节二：编译源码

进入官网源码地址

下载release2.7

下载三个模型

下载推理预测库

官网下载地址，根据自己的情况下载CPU或对应的自己CUDA版本的GPU

用CMAK打开如下文件

添加路径OpenCV路径和预测库路径。

然后Configers，再Generate，如果有报错不用管他，最后在你的构建目录生成了项目，然后开始编译。

但是编译会报错

在utility.cpp中无法打开包括文件"dirent.h":No such file or directory

<dirent.h>是个unix系统下常见的接口，但windows平台的MSVC编译器并没有提供这个接口，对于跨平台的项目开发就会带来一些麻烦，如果在MSVC下编译时可能因为windows平台缺少这个接口就要为windows平台另外写一些代码。

不过大佬已经做了一个windows版本的<dirent.h>，放在了github上面，链接如下：

https://github.com/tronkko/dirent

下载完后加入这个文件夹：

然后重新编译，但是会报错找不到_stat，做如下修改就可以了。

//修改前struct stat s;_stat(dir_name, &s);
//修改后struct _stat64 s;_stat64(dir_name, &s);

最后就可以生成成功，生成成功要将这几个dll复制到程序目录。

onnxruntime和paddle2onnx在这里

将这个文件复制到程序目录

源码里面这个文件的路径也要改一下

然后输入命令运行程序，下面的路径根据自己情况去修改

./ppocr.exe --det_model_dir=F:/Vision/PaddleOCR/2_7/model/ch_PP-OCRv4_det_infer --cls_model_dir=F:/Vision/PaddleOCR/2_7/model/ch_ppocr_mobile_v2.0_cls_infer --rec_model_dir=F:/Vision/PaddleOCR/2_7/model/ch_PP-OCRv4_rec_infer --image_dir=F:/Vision/PaddleOCR/2_7/PaddleOCR-release-2.7/deploy/cpp_infer/build/Release/aa.png

运行结果如下

章节三：在Qt上编译运行

pro文件构建

QT       += core gui widgetsgreaterThan(QT_MAJOR_VERSION, 4): QT += widgetsCONFIG += c++17DEFINES += QT_DEPRECATED_WARNINGS##################################################################
#指定生成的文件存放位置
##################################################################
MOC_DIR         = $$PWD/temp/moc
RCC_DIR         = $$PWD/temp/rcc
UI_DIR          = $$PWD/temp/ui
OBJECTS_DIR     = $$PWD/temp/obj
DESTDIR         = $$PWD/temp/binCONFIG(debug, debug|release) {QMAKE_CXXFLAGS_DEBUG += /MTd
}CONFIG(release, debug|release) {QMAKE_CXXFLAGS_RELEASE += /MT
}PaddleOCR_ROOT = F:/Vision/PaddleOCR/2_7/OCR_Demo/OCR_Demo/ocr
SOURCES += \
#    $$PaddleOCR_ROOT/src/args.cpp \$$PaddleOCR_ROOT/src/clipper.cpp \$$PaddleOCR_ROOT/src/ocr_cls.cpp \$$PaddleOCR_ROOT/src/ocr_det.cpp \$$PaddleOCR_ROOT/src/ocr_rec.cpp \
#    $$PaddleOCR_ROOT/src/paddleocr.cpp \
#    $$PaddleOCR_ROOT/src/paddlestructure.cpp \$$PaddleOCR_ROOT/src/postprocess_op.cpp \$$PaddleOCR_ROOT/src/preprocess_op.cpp \$$PaddleOCR_ROOT/src/structure_layout.cpp \$$PaddleOCR_ROOT/src/structure_table.cpp \$$PaddleOCR_ROOT/src/utility.cpp \ScreenWidget/screen.cpp \ScreenWidget/screenwidget.cpp \main.cpp \mainwindow.cpp \my_config.cpp \my_paddleocr.cppHEADERS += \$$PaddleOCR_ROOT/include/ocr_cls.h \$$PaddleOCR_ROOT/include/ocr_det.h \$$PaddleOCR_ROOT/include/ocr_rec.h \
#    $$PaddleOCR_ROOT/include/paddleocr.h \
#    $$PaddleOCR_ROOT/include/paddlestructure.h \$$PaddleOCR_ROOT/include/postprocess_op.h \$$PaddleOCR_ROOT/include/preprocess_op.h \$$PaddleOCR_ROOT/include/structure_layout.h \$$PaddleOCR_ROOT/include/structure_table.h \$$PaddleOCR_ROOT/include/utility.h \
#    $$PaddleOCR_ROOT/include/args.h \$$PaddleOCR_ROOT/include/clipper.h \$$PaddleOCR_ROOT/include/dirent.h \ScreenWidget/screen.h \ScreenWidget/screenwidget.h \mainwindow.h \my_config.h \my_paddleocr.hFORMS += \mainwindow.uiINCLUDEPATH += $$PaddleOCR_ROOT
INCLUDEPATH += $$PaddleOCR_ROOT/include
INCLUDEPATH += $$PWD\ScreenWidgetInference_ROOT = F:/Vision/PaddleOCR/2_7/prelib/cuda11_0
INCLUDEPATH += $$Inference_ROOT/paddle/include
INCLUDEPATH += $$Inference_ROOT/third_party/install/protobuf/include
INCLUDEPATH += $$Inference_ROOT/third_party/install/glog/include
#INCLUDEPATH += $$Inference_ROOT/third_party/install/gflags/include
INCLUDEPATH += $$Inference_ROOT/third_party/install/xxhash/include
INCLUDEPATH += $$Inference_ROOT/third_party/install/mklml/include
INCLUDEPATH += $$Inference_ROOT/third_party/install/mkldnn/includeLIBS += -L$$Inference_ROOT/paddle/lib -lpaddle_inference
LIBS += -L$$Inference_ROOT/third_party/install/mklml/lib -lmklml
LIBS += -L$$Inference_ROOT/third_party/install/mklml/lib -llibiomp5md
LIBS += -L$$Inference_ROOT/third_party/install/mkldnn/lib -lmkldnn
LIBS += -L$$Inference_ROOT/third_party/install/glog/lib -lglog
#LIBS += -L$$Inference_ROOT/third_party/install/gflags/lib -lgflags_static
LIBS += -L$$Inference_ROOT/third_party/install/protobuf/lib -llibprotobuf
LIBS += -L$$Inference_ROOT/third_party/install/xxhash/lib -lxxhashOpenCV_ROOT = E:/2021_software/OpenCv/OpenCv3_4_15/install/opencv/build
INCLUDEPATH += $$OpenCV_ROOT/include
INCLUDEPATH += $$OpenCV_ROOT/include/opencv
INCLUDEPATH += $$OpenCV_ROOT/include/opencv2
LIBS += -L$$OpenCV_ROOT/x64/vc15/lib -lopencv_world3415

自定义一个config类

#pragma once#include <iomanip>
#include <iostream>
#include <map>
#include <ostream>
#include <string>
#include <vector>#include "include/utility.h"using namespace PaddleOCR;class MY_OCRConfig {
public:explicit MY_OCRConfig(const std::string &config_file);// common argsbool use_gpu = false;bool use_tensorrt = false;int gpu_id = 0;int gpu_mem = 4000;int cpu_threads = 10;bool enable_mkldnn = false;std::string precision = "fp32";bool benchmark = false;std::string output = "./output/";std::string image_dir = "";std::string type = "ocr";// detection relatedstd::string det_model_dir = "";std::string limit_type = "max";int limit_side_len = 960;double det_db_thresh = 0.3;double det_db_box_thresh = 0.6;double det_db_unclip_ratio = 1.5;bool use_dilation = false;std::string det_db_score_mode = "slow";bool visualize = true;// classification relatedbool use_angle_cls = false;std::string cls_model_dir = "";double cls_thresh = 0.9;int cls_batch_num = 1;// recognition relatedstd::string rec_model_dir = "";int rec_batch_num = 6;std::string rec_char_dict_path = "./ppocr_keys_v1.txt";int rec_img_h = 48;int rec_img_w = 320;// layout model relatedstd::string layout_model_dir = "";std::string layout_dict_path = "../../ppocr/utils/dict/layout_dict/layout_publaynet_dict.txt";double layout_score_threshold = 0.5;double layout_nms_threshold = 0.5;// structure model relatedstd::string table_model_dir = "";int table_max_len = 488;int table_batch_num = 1;bool merge_no_span_structure = true;std::string table_char_dict_path = "../../ppocr/utils/dict/table_structure_dict_ch.txt";// ocr forward relatedbool det = true;bool rec = true;bool cls = false;bool table = false;bool layout = false;private:// Load configurationstd::map<std::string, std::string> LoadConfig(const std::string &config_file);std::vector<std::string> split(const std::string &str,const std::string &delim);std::map<std::string, std::string> config_map_;
};#include "my_config.h"
#include <qdebug.h>std::vector<std::string> MY_OCRConfig::split(const std::string &str,const std::string &delim) {std::vector<std::string> res;if ("" == str)return res;int strlen = str.length() + 1;char *strs = new char[strlen];std::strcpy(strs, str.c_str());int delimlen = delim.length() + 1;char *d = new char[delimlen];std::strcpy(d, delim.c_str());char *p = std::strtok(strs, d);while (p) {std::string s = p;res.push_back(s);p = std::strtok(NULL, d);}delete[] strs;delete[] d;return res;
}std::map<std::string, std::string>
MY_OCRConfig::LoadConfig(const std::string &config_path) {auto config = Utility::ReadDict(config_path);std::map<std::string, std::string> dict;for (int i = 0; i < config.size(); i++) {// pass for empty line or commentif (config[i].size() <= 1 || config[i][0] == '#') {continue;}//std::vector<std::string> res = split(config[i], " ");if (res.size() < 2) {dict[res[0]] = "";}else{dict[res[0]] = res[1];}}return dict;
}MY_OCRConfig::MY_OCRConfig(const std::string &config_file)
{config_map_ = LoadConfig(config_file);// common argsthis->use_gpu        = (config_map_["use_gpu"] == "true");this->use_tensorrt   = (config_map_["use_tensorrt"] == "true");this->gpu_id         = stoi(config_map_["gpu_id"]);this->gpu_mem        = stoi(config_map_["gpu_mem"]);this->cpu_threads    = stoi(config_map_["cpu_threads"]);this->enable_mkldnn  = (config_map_["enable_mkldnn"] == "true");this->precision      = config_map_["precision"];this->benchmark      = (config_map_["benchmark"] == "true");this->output         = config_map_["output"];this->image_dir      = config_map_["image_dir"];this->type           = config_map_["type"];// detection relatedthis->det_model_dir       = config_map_["det_model_dir"];this->limit_type          = config_map_["limit_type"];this->limit_side_len      = stoi(config_map_["limit_side_len"]);this->det_db_thresh       = stod(config_map_["det_db_thresh"]);this->det_db_box_thresh   = stod(config_map_["det_db_box_thresh"]);this->det_db_unclip_ratio = stod(config_map_["det_db_unclip_ratio"]);this->use_dilation        = (config_map_["use_dilation"] == "true");this->det_db_score_mode   = config_map_["det_db_score_mode"];this->visualize           = (config_map_["visualize"] == "true");// classification relatedthis->use_angle_cls   = (config_map_["use_angle_cls"] == "true");this->cls_model_dir   = config_map_["cls_model_dir"];this->cls_thresh      = stod(config_map_["cls_thresh"]);this->cls_batch_num   = stoi(config_map_["cls_batch_num"]);// recognition relatedthis->rec_model_dir        = config_map_["rec_model_dir"];this->rec_batch_num        = stoi(config_map_["rec_batch_num"]);this->rec_char_dict_path   = config_map_["rec_char_dict_path"];this->rec_img_h            = stoi(config_map_["rec_img_h"]);this->rec_img_w            = stoi(config_map_["rec_img_w"]);// layout model relatedthis->layout_model_dir        = config_map_["layout_model_dir"];this->layout_dict_path        = config_map_["layout_dict_path"];this->layout_score_threshold  = stod(config_map_["layout_score_threshold"]);this->layout_nms_threshold    = stod(config_map_["layout_nms_threshold"]);// structure model relatedthis->table_model_dir           = config_map_["table_model_dir"];this->table_max_len             = stoi(config_map_["table_max_len"]);this->table_batch_num           = stoi(config_map_["table_batch_num"]);this->merge_no_span_structure   = (config_map_["merge_no_span_structure"] == "true");this->table_char_dict_path      = config_map_["table_char_dict_path"];// ocr forward relatedthis->det     = (config_map_["det"] == "true");this->rec     = (config_map_["rec"] == "true");this->cls     = (config_map_["cls"] == "true");this->table   = (config_map_["table"] == "true");this->layout  = (config_map_["layout"] == "true");qDebug()<<this->det<<config_map_["det"].c_str()<<QString(config_map_["det"].c_str())<<this->rec<<this->cls<<this->rec_model_dir.c_str();
}

自定义一个ocr类

#pragma once#include <include/ocr_cls.h>
#include <include/ocr_det.h>
#include <include/ocr_rec.h>#include "my_config.h"using namespace PaddleOCR;
class MY_PPOCR {
public:explicit MY_PPOCR();~MY_PPOCR();std::vector<std::vector<OCRPredictResult>> ocr(std::vector<cv::Mat> img_list,bool det = true,bool rec = true,bool cls = true);std::vector<OCRPredictResult> ocr(cv::Mat img, bool det = true,bool rec = true, bool cls = true);void reset_timer();void benchmark_log(int img_num);MY_OCRConfig *p_config;
protected:std::vector<double> time_info_det = {0, 0, 0};std::vector<double> time_info_rec = {0, 0, 0};std::vector<double> time_info_cls = {0, 0, 0};void det(cv::Mat img, std::vector<OCRPredictResult> &ocr_results);void rec(std::vector<cv::Mat> img_list,std::vector<OCRPredictResult> &ocr_results);void cls(std::vector<cv::Mat> img_list,std::vector<OCRPredictResult> &ocr_results);private:std::unique_ptr<DBDetector> detector_;std::unique_ptr<Classifier> classifier_;std::unique_ptr<CRNNRecognizer> recognizer_;
};#include "my_paddleocr.h"
#include <qdebug.h>
//#include "auto_log/autolog.h"MY_PPOCR::MY_PPOCR()
{qDebug()<<"aaa1";p_config = new MY_OCRConfig("./ocrconfig.txt");qDebug()<<"aaa2";//if (p_config->det) {this->detector_.reset(new DBDetector(p_config->det_model_dir, p_config->use_gpu, p_config->gpu_id, p_config->gpu_mem,p_config->cpu_threads, p_config->enable_mkldnn, p_config->limit_type,p_config->limit_side_len, p_config->det_db_thresh, p_config->det_db_box_thresh,p_config->det_db_unclip_ratio, p_config->det_db_score_mode, p_config->use_dilation,p_config->use_tensorrt, p_config->precision));}qDebug()<<"aaa3";//if (p_config->cls && p_config->use_angle_cls) {this->classifier_.reset(new Classifier(p_config->cls_model_dir, p_config->use_gpu, p_config->gpu_id, p_config->gpu_mem,p_config->cpu_threads, p_config->enable_mkldnn, p_config->cls_thresh,p_config->use_tensorrt, p_config->precision, p_config->cls_batch_num));}qDebug()<<"aaa4";//if (p_config->rec) {this->recognizer_.reset(new CRNNRecognizer(p_config->rec_model_dir, p_config->use_gpu, p_config->gpu_id, p_config->gpu_mem,p_config->cpu_threads, p_config->enable_mkldnn, p_config->rec_char_dict_path,p_config->use_tensorrt, p_config->precision, p_config->rec_batch_num,p_config->rec_img_h, p_config->rec_img_w));}qDebug()<<"aaa5";
}MY_PPOCR::~MY_PPOCR()
{delete p_config;
}std::vector<std::vector<OCRPredictResult>>
MY_PPOCR::ocr(std::vector<cv::Mat> img_list, bool det, bool rec, bool cls)
{std::vector<std::vector<OCRPredictResult>> ocr_results;if (!det) {std::vector<OCRPredictResult> ocr_result;ocr_result.resize(img_list.size());if (cls && this->classifier_) {this->cls(img_list, ocr_result);for (int i = 0; i < img_list.size(); i++) {if (ocr_result[i].cls_label % 2 == 1 &&ocr_result[i].cls_score > this->classifier_->cls_thresh) {cv::rotate(img_list[i], img_list[i], 1);}}}if (rec) {this->rec(img_list, ocr_result);}for (int i = 0; i < ocr_result.size(); ++i) {std::vector<OCRPredictResult> ocr_result_tmp;ocr_result_tmp.push_back(ocr_result[i]);ocr_results.push_back(ocr_result_tmp);}} else {for (int i = 0; i < img_list.size(); ++i) {std::vector<OCRPredictResult> ocr_result =this->ocr(img_list[i], true, rec, cls);ocr_results.push_back(ocr_result);}}return ocr_results;
}std::vector<OCRPredictResult> MY_PPOCR::ocr(cv::Mat img, bool det, bool rec, bool cls)
{std::vector<OCRPredictResult> ocr_result;// detthis->det(img, ocr_result);// crop imagestd::vector<cv::Mat> img_list;for (int j = 0; j < ocr_result.size(); j++) {cv::Mat crop_img;crop_img = Utility::GetRotateCropImage(img, ocr_result[j].box);img_list.push_back(crop_img);}// clsif (cls && this->classifier_) {this->cls(img_list, ocr_result);for (int i = 0; i < img_list.size(); i++) {if (ocr_result[i].cls_label % 2 == 1 &&ocr_result[i].cls_score > this->classifier_->cls_thresh) {cv::rotate(img_list[i], img_list[i], 1);}}}// recif (rec) {this->rec(img_list, ocr_result);}return ocr_result;
}void MY_PPOCR::det(cv::Mat img, std::vector<OCRPredictResult> &ocr_results)
{std::vector<std::vector<std::vector<int>>> boxes;std::vector<double> det_times;this->detector_->Run(img, boxes, det_times);for (int i = 0; i < boxes.size(); i++) {OCRPredictResult res;res.box = boxes[i];ocr_results.push_back(res);}// sort boex from top to bottom, from left to rightUtility::sorted_boxes(ocr_results);this->time_info_det[0] += det_times[0];this->time_info_det[1] += det_times[1];this->time_info_det[2] += det_times[2];
}void MY_PPOCR::rec(std::vector<cv::Mat> img_list, std::vector<OCRPredictResult> &ocr_results)
{std::vector<std::string> rec_texts(img_list.size(), "");std::vector<float> rec_text_scores(img_list.size(), 0);std::vector<double> rec_times;this->recognizer_->Run(img_list, rec_texts, rec_text_scores, rec_times);// output rec resultsfor (int i = 0; i < rec_texts.size(); i++) {ocr_results[i].text = rec_texts[i];ocr_results[i].score = rec_text_scores[i];}this->time_info_rec[0] += rec_times[0];this->time_info_rec[1] += rec_times[1];this->time_info_rec[2] += rec_times[2];
}void MY_PPOCR::cls(std::vector<cv::Mat> img_list, std::vector<OCRPredictResult> &ocr_results)
{std::vector<int> cls_labels(img_list.size(), 0);std::vector<float> cls_scores(img_list.size(), 0);std::vector<double> cls_times;this->classifier_->Run(img_list, cls_labels, cls_scores, cls_times);// output cls resultsfor (int i = 0; i < cls_labels.size(); i++) {ocr_results[i].cls_label = cls_labels[i];ocr_results[i].cls_score = cls_scores[i];}this->time_info_cls[0] += cls_times[0];this->time_info_cls[1] += cls_times[1];this->time_info_cls[2] += cls_times[2];
}void MY_PPOCR::reset_timer()
{this->time_info_det = {0, 0, 0};this->time_info_rec = {0, 0, 0};this->time_info_cls = {0, 0, 0};
}

核心调用代码

//std::vector<cv::Mat> img_list;cv::Mat srcimg = cv::imread(qstr2str(fileName).data(), cv::IMREAD_COLOR);img_list.push_back(srcimg);p_ocr->reset_timer();//QElapsedTimer  RunTimer;RunTimer.start();std::vector<std::vector<OCRPredictResult>> ocr_results = p_ocr->ocr(img_list, p_ocr->p_config->det, p_ocr->p_config->rec, p_ocr->p_config->cls);ui->textBrowser->append(QString("检测时间:%1ms ---------------------").arg(RunTimer.elapsed()));//for (int i = 0; i < img_list.size(); ++i) {std::vector<OCRPredictResult> &ocr_result = ocr_results[i];for (int i = 0; i < ocr_result.size(); i++) {QString oustr;oustr += QString::number(i) + " ";// det
//            std::vector<std::vector<int>> boxes = ocr_result[i].box;
//            if (boxes.size() > 0) {
//                oustr += "det boxes: [";
//                for (int n = 0; n < boxes.size(); n++) {
//                    oustr +=  "[" + QString::number(boxes[n][0]) + "," + QString::number(boxes[n][1]) + "]";
//                    if (n != boxes.size() - 1) {
//                        oustr +=  ",";
//                    }
//                }
//                oustr +=  "]";
//            }// recif (ocr_result[i].score != -1.0) {oustr += " score: " + QString::number(ocr_result[i].score,'f',2) + " text: " + QString::fromUtf8(ocr_result[i].text.c_str()) + " ";}// clsif (ocr_result[i].cls_label != -1) {oustr += "cls label: " + QString::number(ocr_result[i].cls_label) + " cls score: " + ocr_result[i].cls_score;}//
//            oustr += "\r\n";ui->textBrowser->append(oustr);}//for (int n = 0; n < ocr_result.size(); n++) {cv::Point rook_points[4];//for (int m = 0; m < ocr_result[n].box.size(); m++) {rook_points[m] = cv::Point(int(ocr_result[n].box[m][0]), int(ocr_result[n].box[m][1]));}//const cv::Point *ppt[1] = {rook_points};int npt[] = {4};cv::polylines(img_list[i], ppt, npt, 1, 1, CV_RGB(255, 0, 0), 2, 8, 0);}//QImage outimage;cvMat2QImage(img_list[i], outimage);ui->label_image->setPixmap(QPixmap::fromImage(outimage.scaled(ui->label_image->width(),ui->label_image->height(),Qt::KeepAspectRatio)));}

程序目录包含这个几个文件，如果你没有，在源码路径和预测库路径去搜索

ocrconfig.txt是自定义的配置

# common args
use_gpu false
use_tensorrt false
gpu_id 0
gpu_mem 4000
cpu_threads 10
enable_mkldnn false
precision fp32
benchmark false
output ./output/
image_dir 
type ocr# detection related
det_model_dir F:\Vision\PaddleOCR\2_7\model\ch_PP-OCRv4_det_infer
limit_type max
limit_side_len 960
det_db_thresh 0.3
det_db_box_thresh 0.6
det_db_unclip_ratio 1.5
use_dilation false
det_db_score_mode slow
visualize true# classification related
use_angle_cls false
cls_model_dir F:\Vision\PaddleOCR\2_7\model\ch_ppocr_mobile_v2.0_cls_infer
cls_thresh 0.9
cls_batch_num 1# recognition related
rec_model_dir F:\Vision\PaddleOCR\2_7\model\ch_PP-OCRv4_rec_infer
rec_batch_num 6
rec_char_dict_path ./ppocr_keys_v1.txt
rec_img_h 48
rec_img_w 320# layout model related
layout_model_dir 
layout_dict_path ./layout_publaynet_dict.txt
layout_score_threshold 0.5
layout_nms_threshold 0.5# structure model related
table_model_dir 
table_max_len 488
table_batch_num 1
merge_no_span_structure true
table_char_dict_path ./table_structure_dict_ch.txt# ocr forward related
det true
rec true
cls false
table false
layout false

最后运行结果