fix cpp infer

deploy/cpp_infer/include/ocr_det.h

@@ -36,16 +36,13 @@ namespace PaddleOCR {
 
 class DBDetector {
 public:
-  explicit DBDetector(const std::string &model_dir, const bool &use_gpu = false,
+  explicit DBDetector(const std::string &model_dir, const bool &use_gpu,
-                      const int &gpu_id = 0, const int &gpu_mem = 4000,
+                      const int &gpu_id, const int &gpu_mem,
-                      const int &cpu_math_library_num_threads = 4,
+                      const int &cpu_math_library_num_threads,
-                      const int &max_side_len = 960,
+                      const int &max_side_len, const double &det_db_thresh,
-                      const double &det_db_thresh = 0.3,
+                      const double &det_db_box_thresh,
-                      const double &det_db_box_thresh = 0.5,
+                      const double &det_db_unclip_ratio,
-                      const double &det_db_unclip_ratio = 2.0,
+                      const bool &visualize) {
-                      const bool &visualize = true) {
-    LoadModel(model_dir);
-
     this->use_gpu_ = use_gpu;
     this->gpu_id_ = gpu_id;
     this->gpu_mem_ = gpu_mem;
@@ -58,6 +55,8 @@ public:
     this->det_db_unclip_ratio_ = det_db_unclip_ratio;
 
     this->visualize_ = visualize;
+
+    LoadModel(model_dir);
   }
 
   // Load Paddle inference model

deploy/cpp_infer/include/ocr_rec.h

@@ -35,19 +35,18 @@ namespace PaddleOCR {
 
 class CRNNRecognizer {
 public:
-  explicit CRNNRecognizer(
+  explicit CRNNRecognizer(const std::string &model_dir, const bool &use_gpu,
-      const std::string &model_dir, const bool &use_gpu = false,
+                          const int &gpu_id, const int &gpu_mem,
-      const int &gpu_id = 0, const int &gpu_mem = 4000,
+                          const int &cpu_math_library_num_threads,
-      const int &cpu_math_library_num_threads = 4,
+                          const string &label_path) {
-      const string &label_path = "./tools/ppocr_keys_v1.txt") {
-    LoadModel(model_dir);
-
     this->use_gpu_ = use_gpu;
     this->gpu_id_ = gpu_id;
     this->gpu_mem_ = gpu_mem;
     this->cpu_math_library_num_threads_ = cpu_math_library_num_threads;
 
     this->label_list_ = Utility::ReadDict(label_path);
+
+    LoadModel(model_dir);
   }
 
   // Load Paddle inference model

deploy/cpp_infer/include/postprocess_op.h

@@ -36,21 +36,21 @@ namespace PaddleOCR {
 
 class PostProcessor {
 public:
-  void GetContourArea(float **box, float unclip_ratio, float &distance);
+  void GetContourArea(const std::vector<std::vector<float>> &box,
+                      float unclip_ratio, float &distance);
 
-  cv::RotatedRect UnClip(float **box, const float &unclip_ratio);
+  cv::RotatedRect UnClip(std::vector<std::vector<float>> box,
+                         const float &unclip_ratio);
 
   float **Mat2Vec(cv::Mat mat);
 
-  void quickSort_vector(std::vector<std::vector<int>> &box, int l, int r,
-                        int axis);
-
   std::vector<std::vector<int>>
-  order_points_clockwise(std::vector<std::vector<int>> pts);
+  OrderPointsClockwise(std::vector<std::vector<int>> pts);
 
-  float **GetMiniBoxes(cv::RotatedRect box, float &ssid);
+  std::vector<std::vector<float>> GetMiniBoxes(cv::RotatedRect box,
+                                               float &ssid);
 
-  float BoxScoreFast(float **box_array, cv::Mat pred);
+  float BoxScoreFast(std::vector<std::vector<float>> box_array, cv::Mat pred);
 
   std::vector<std::vector<std::vector<int>>>
   BoxesFromBitmap(const cv::Mat pred, const cv::Mat bitmap,
@@ -61,7 +61,11 @@ public:
                   float ratio_h, float ratio_w, cv::Mat srcimg);
 
 private:
-  void quickSort(float **s, int l, int r);
+  static bool XsortInt(std::vector<int> a, std::vector<int> b);
+
+  static bool XsortFp32(std::vector<float> a, std::vector<float> b);
+
+  std::vector<std::vector<float>> Mat2Vector(cv::Mat mat);
 
   inline int _max(int a, int b) { return a >= b ? a : b; }
 

deploy/cpp_infer/readme.md

@@ -22,7 +22,7 @@ tar -xf 3.4.7.tar.gz
 * 编译opencv，设置opencv源码路径(`root_path`)以及安装路径(`install_path`)。进入opencv源码路径下，按照下面的方式进行编译。
 
 ```shell
-root_path=/paddle/libs/opencv-3.4.7
+root_path=your_opencv_root_path
 install_path=${root_path}/opencv3
 
 rm -rf build
@@ -51,6 +51,9 @@ make -j
 make install
 ```
 
+
+其中`root_path`为下载的opencv源码路径，`install_path`为opencv的安装路径，`make install`完成之后，会在该文件夹下生成opencv头文件和库文件，用于后面的OCR代码编译。
+
 最终在安装路径下的文件结构如下所示。
 
 ```
@@ -62,8 +65,29 @@ opencv3/
 |-- share
 ```
 
-### 1.2 编译Paddle预测库
+### 1.2 下载或者编译Paddle预测库
+
+* 有2种方式获取Paddle预测库，下面进行详细介绍。
+
+#### 1.2.1 直接下载安装
+
+* [Paddle预测库官网](https://www.paddlepaddle.org.cn/documentation/docs/zh/advanced_guide/inference_deployment/inference/build_and_install_lib_cn.html)上提供了不同cuda版本的Linux预测库，可以直接下载使用。
+  * 如果cuda版本为cuda9.0，1.8.2版本的Paddle预测库可以从这里下载：[下载地址](https://paddle-inference-lib.bj.bcebos.com/1.8.2-gpu-cuda9-cudnn7-avx-mkl/fluid_inference.tgz)。
+  * 如果cuda版本为cuda10.0，1.8.2版本的Paddle预测库可以从这里下载：[下载地址](https://paddle-inference-lib.bj.bcebos.com/1.8.2-gpu-cuda10-cudnn7-avx-mkl/fluid_inference.tgz)。
+  * 更多版本的预测库可以在官网查看下载。
+
+* 下载之后使用下面的方法解压
+
+```
+tar -xf fluid_inference.tgz
+```
+
+最终会在当前的文件夹中生成`fluid_inference/`的子文件夹。
+
+
 
+#### 1.2.2 预测库源码编译
+* 如果希望获取最新预测库特性，可以从Paddle github上克隆最新代码，源码编译预测库。
 * 可以参考[Paddle预测库官网](https://www.paddlepaddle.org.cn/documentation/docs/zh/advanced_guide/inference_deployment/inference/build_and_install_lib_cn.html)的说明，从github上获取Paddle代码，然后进行编译，生成最新的预测库。使用git获取代码方法如下。
 
 ```shell
@@ -80,7 +104,7 @@ cd build
 cmake  .. \
     -DWITH_CONTRIB=OFF \
     -DWITH_MKL=ON \
-    -DWITH_MKLDNN=OFF  \
+    -DWITH_MKLDNN=ON  \
     -DWITH_TESTING=OFF \
     -DCMAKE_BUILD_TYPE=Release \
     -DWITH_INFERENCE_API_TEST=OFF \
@@ -132,6 +156,15 @@ inference/
 sh tools/build.sh
 ```
 
+具体地，`tools/build.sh`中内容如下。
+
+```shell
+c
+```
+
+`OPENCV_DIR`为opencv编译安装的地址；`LIB_DIR`为下载(`fluid_inference`文件夹)或者编译生成的Paddle预测库地址(`build/fluid_inference_install_dir`文件夹)；`CUDA_LIB_DIR`为cuda库文件地址，在docker中；为`/usr/local/cuda/lib64`；`CUDNN_LIB_DIR`为cudnn库文件地址，在docker中为`/usr/lib/x86_64-linux-gnu/`。
+
+
 * 编译完成之后，会在`build`文件夹下生成一个名为`ocr_system`的可执行文件。
 
 

deploy/cpp_infer/src/main.cpp

@@ -46,8 +46,6 @@ int main(int argc, char **argv) {
 
   std::string img_path(argv[2]);
 
-  auto start = std::chrono::system_clock::now();
-
   cv::Mat srcimg = cv::imread(img_path, cv::IMREAD_COLOR);
 
   DBDetector det(config.det_model_dir, config.use_gpu, config.gpu_id,
@@ -59,6 +57,7 @@ int main(int argc, char **argv) {
                      config.gpu_mem, config.cpu_math_library_num_threads,
                      config.char_list_file);
 
+  auto start = std::chrono::system_clock::now();
   std::vector<std::vector<std::vector<int>>> boxes;
   det.Run(srcimg, boxes);
 

deploy/cpp_infer/src/ocr_det.cpp

@@ -12,22 +12,8 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#include "opencv2/core.hpp"
-#include "opencv2/imgcodecs.hpp"
-#include "opencv2/imgproc.hpp"
-#include "paddle_api.h"
-#include "paddle_inference_api.h"
-#include <chrono>
-#include <iomanip>
-#include <iostream>
-#include <ostream>
-#include <vector>
-
-#include <cstring>
-#include <fstream>
-#include <numeric>
-
 #include <include/ocr_det.h>
+#include <stdlib.h>
 
 namespace PaddleOCR {
 

deploy/cpp_infer/src/postprocess_op.cpp

@@ -16,8 +16,8 @@
 
 namespace PaddleOCR {
 
-void PostProcessor::GetContourArea(float **box, float unclip_ratio,
+void PostProcessor::GetContourArea(const std::vector<std::vector<float>> &box,
-                                   float &distance) {
+                                   float unclip_ratio, float &distance) {
   int pts_num = 4;
   float area = 0.0f;
   float dist = 0.0f;
@@ -34,7 +34,8 @@ void PostProcessor::GetContourArea(float **box, float unclip_ratio,
   distance = area * unclip_ratio / dist;
 }
 
-cv::RotatedRect PostProcessor::UnClip(float **box, const float &unclip_ratio) {
+cv::RotatedRect PostProcessor::UnClip(std::vector<std::vector<float>> box,
+                                      const float &unclip_ratio) {
   float distance = 1.0;
 
   GetContourArea(box, unclip_ratio, distance);
@@ -74,53 +75,11 @@ float **PostProcessor::Mat2Vec(cv::Mat mat) {
   return array;
 }
 
-void PostProcessor::quickSort(float **s, int l, int r) {
-  if (l < r) {
-    int i = l, j = r;
-    float x = s[l][0];
-    float *xp = s[l];
-    while (i < j) {
-      while (i < j && s[j][0] >= x)
-        j--;
-      if (i < j)
-        std::swap(s[i++], s[j]);
-      while (i < j && s[i][0] < x)
-        i++;
-      if (i < j)
-        std::swap(s[j--], s[i]);
-    }
-    s[i] = xp;
-    quickSort(s, l, i - 1);
-    quickSort(s, i + 1, r);
-  }
-}
-
-void PostProcessor::quickSort_vector(std::vector<std::vector<int>> &box, int l,
-                                     int r, int axis) {
-  if (l < r) {
-    int i = l, j = r;
-    int x = box[l][axis];
-    std::vector<int> xp(box[l]);
-    while (i < j) {
-      while (i < j && box[j][axis] >= x)
-        j--;
-      if (i < j)
-        std::swap(box[i++], box[j]);
-      while (i < j && box[i][axis] < x)
-        i++;
-      if (i < j)
-        std::swap(box[j--], box[i]);
-    }
-    box[i] = xp;
-    quickSort_vector(box, l, i - 1, axis);
-    quickSort_vector(box, i + 1, r, axis);
-  }
-}
-
 std::vector<std::vector<int>>
-PostProcessor::order_points_clockwise(std::vector<std::vector<int>> pts) {
+PostProcessor::OrderPointsClockwise(std::vector<std::vector<int>> pts) {
   std::vector<std::vector<int>> box = pts;
-  quickSort_vector(box, 0, int(box.size() - 1), 0);
+  std::sort(box.begin(), box.end(), XsortInt);
+
   std::vector<std::vector<int>> leftmost = {box[0], box[1]};
   std::vector<std::vector<int>> rightmost = {box[2], box[3]};
 
@@ -135,16 +94,44 @@ PostProcessor::order_points_clockwise(std::vector<std::vector<int>> pts) {
   return rect;
 }
 
-float **PostProcessor::GetMiniBoxes(cv::RotatedRect box, float &ssid) {
+std::vector<std::vector<float>> PostProcessor::Mat2Vector(cv::Mat mat) {
-  ssid = box.size.width >= box.size.height ? box.size.height : box.size.width;
+  std::vector<std::vector<float>> img_vec;
+  std::vector<float> tmp;
+
+  for (int i = 0; i < mat.rows; ++i) {
+    tmp.clear();
+    for (int j = 0; j < mat.cols; ++j) {
+      tmp.push_back(mat.at<float>(i, j));
+    }
+    img_vec.push_back(tmp);
+  }
+  return img_vec;
+}
+
+bool PostProcessor::XsortFp32(std::vector<float> a, std::vector<float> b) {
+  if (a[0] != b[0])
+    return a[0] < b[0];
+  return false;
+}
+
+bool PostProcessor::XsortInt(std::vector<int> a, std::vector<int> b) {
+  if (a[0] != b[0])
+    return a[0] < b[0];
+  return false;
+}
+
+std::vector<std::vector<float>> PostProcessor::GetMiniBoxes(cv::RotatedRect box,
+                                                            float &ssid) {
+  ssid = std::max(box.size.width, box.size.height);
 
   cv::Mat points;
   cv::boxPoints(box, points);
-  // sorted box points
-  auto array = Mat2Vec(points);
-  quickSort(array, 0, 3);
 
-  float *idx1 = array[0], *idx2 = array[1], *idx3 = array[2], *idx4 = array[3];
+  auto array = Mat2Vector(points);
+  std::sort(array.begin(), array.end(), XsortFp32);
+
+  std::vector<float> idx1 = array[0], idx2 = array[1], idx3 = array[2],
+                     idx4 = array[3];
   if (array[3][1] <= array[2][1]) {
     idx2 = array[3];
     idx3 = array[2];
@@ -168,7 +155,8 @@ float **PostProcessor::GetMiniBoxes(cv::RotatedRect box, float &ssid) {
   return array;
 }
 
-float PostProcessor::BoxScoreFast(float **box_array, cv::Mat pred) {
+float PostProcessor::BoxScoreFast(std::vector<std::vector<float>> box_array,
+                                  cv::Mat pred) {
   auto array = box_array;
   int width = pred.cols;
   int height = pred.rows;
@@ -280,7 +268,7 @@ PostProcessor::FilterTagDetRes(std::vector<std::vector<std::vector<int>>> boxes,
 
   std::vector<std::vector<std::vector<int>>> root_points;
   for (int n = 0; n < boxes.size(); n++) {
-    boxes[n] = order_points_clockwise(boxes[n]);
+    boxes[n] = OrderPointsClockwise(boxes[n]);
     for (int m = 0; m < boxes[0].size(); m++) {
       boxes[n][m][0] /= ratio_w;
       boxes[n][m][1] /= ratio_h;

deploy/cpp_infer/tools/build.sh

@@ -1,9 +1,8 @@
 
-OPENCV_DIR=/paddle/libs/opencv-3.4.7/opencv3
+OPENCV_DIR=your_opencv_dir
-LIB_DIR=/paddle/code/gry/Paddle/build/fluid_inference_install_dir/
+LIB_DIR=your_paddle_inference_dir
-CUDA_LIB_DIR=/usr/local/cuda/lib64
+CUDA_LIB_DIR=your_cuda_lib_dir
-CUDNN_LIB_DIR=/usr/lib/x86_64-linux-gnu/
+CUDNN_LIB_DIR=/your_cudnn_lib_dir
-TENSORRT_ROOT_DIR=YOUR_TENSORRT_ROOT_DIR
 
 BUILD_DIR=build
 rm -rf ${BUILD_DIR}
@@ -19,8 +18,5 @@ cmake .. \
     -DOPENCV_DIR=${OPENCV_DIR} \
     -DCUDNN_LIB=${CUDNN_LIB_DIR} \
     -DCUDA_LIB=${CUDA_LIB_DIR} \
-    -DTENSORRT_ROOT=YOUR_TENSORRT_ROOT_DIR
 
 make -j
-
-

deploy/cpp_infer/tools/config.txt

@@ -13,7 +13,7 @@ det_model_dir  ./inference/det_db
 
 # rec config
 rec_model_dir  ./inference/rec_crnn
-char_list_file ./tools/ppocr_keys_v1.txt
+char_list_file ../../ppocr/utils/ppocr_keys_v1.txt
 img_path  ../../doc/imgs/11.jpg
 
 # show the detection results