Merge pull request #283 from LDOUBLEV/fixocr

add deploy lite demo

.clang_format.hook

+#!/bin/bash
+set -e
+
+readonly VERSION="3.8"
+
+version=$(clang-format -version)
+
+if ! [[ $version == *"$VERSION"* ]]; then
+    echo "clang-format version check failed."
+    echo "a version contains '$VERSION' is needed, but get '$version'"
+    echo "you can install the right version, and make an soft-link to '\$PATH' env"
+    exit -1
+fi
+
+clang-format $@

.gitignore

@@ -18,3 +18,5 @@ output/
 *.idea
 
 *.log
+.clang-format
+.clang_format.hook

README.md

@@ -29,7 +29,7 @@ PaddleOCR旨在打造一套丰富、领先、且实用的OCR工具库，助力
 
 - 超轻量级中文OCR在线体验地址：https://www.paddlepaddle.org.cn/hub/scene/ocr
 
-- [**中文OCR模型快速使用**](./doc/doc_ch/quickstart.md) 
+- [**中文OCR模型快速使用**](./doc/doc_ch/quickstart.md)
 
 ## 中文OCR模型列表
 
@@ -50,7 +50,7 @@ PaddleOCR旨在打造一套丰富、领先、且实用的OCR工具库，助力
     - [基于Python预测引擎推理](./doc/doc_ch/inference.md)
     - 基于C++预测引擎推理(comming soon)
     - [服务部署](./doc/doc_ch/serving.md)
-    - 端侧部署(comming soon)
+    - [端侧部署](./deploy/lite/readme.md)
 - [数据集](./doc/doc_ch/datasets.md)
 - [FAQ](#FAQ)
 - 效果展示

deploy/lite/Makefile

+ARM_ABI = arm8
+export ARM_ABI
+
+include ../Makefile.def
+
+LITE_ROOT=../../../
+
+THIRD_PARTY_DIR=${LITE_ROOT}/third_party
+
+OPENCV_VERSION=opencv4.1.0
+
+OPENCV_LIBS = ../../../third_party/${OPENCV_VERSION}/arm64-v8a/libs/libopencv_imgcodecs.a \
+              ../../../third_party/${OPENCV_VERSION}/arm64-v8a/libs/libopencv_imgproc.a \
+              ../../../third_party/${OPENCV_VERSION}/arm64-v8a/libs/libopencv_core.a \
+              ../../../third_party/${OPENCV_VERSION}/arm64-v8a/3rdparty/libs/libtegra_hal.a \
+              ../../../third_party/${OPENCV_VERSION}/arm64-v8a/3rdparty/libs/liblibjpeg-turbo.a \
+              ../../../third_party/${OPENCV_VERSION}/arm64-v8a/3rdparty/libs/liblibwebp.a \
+              ../../../third_party/${OPENCV_VERSION}/arm64-v8a/3rdparty/libs/liblibpng.a \
+              ../../../third_party/${OPENCV_VERSION}/arm64-v8a/3rdparty/libs/liblibjasper.a \
+              ../../../third_party/${OPENCV_VERSION}/arm64-v8a/3rdparty/libs/liblibtiff.a \
+              ../../../third_party/${OPENCV_VERSION}/arm64-v8a/3rdparty/libs/libIlmImf.a \
+              ../../../third_party/${OPENCV_VERSION}/arm64-v8a/3rdparty/libs/libtbb.a \
+              ../../../third_party/${OPENCV_VERSION}/arm64-v8a/3rdparty/libs/libcpufeatures.a
+
+OPENCV_INCLUDE = -I../../../third_party/${OPENCV_VERSION}/arm64-v8a/include
+
+CXX_INCLUDES = $(INCLUDES) ${OPENCV_INCLUDE} -I$(LITE_ROOT)/cxx/include
+
+CXX_LIBS = ${OPENCV_LIBS} -L$(LITE_ROOT)/cxx/lib/ -lpaddle_light_api_shared $(SYSTEM_LIBS)
+
+###############################################################
+# How to use one of static libaray:                           #
+#  `libpaddle_api_full_bundled.a`                             #
+#  `libpaddle_api_light_bundled.a`                            #
+###############################################################
+# Note: default use lite's shared library.                    #
+###############################################################
+# 1. Comment above line using `libpaddle_light_api_shared.so`
+# 2. Undo comment below line using `libpaddle_api_light_bundled.a`
+
+#CXX_LIBS = $(LITE_ROOT)/cxx/lib/libpaddle_api_light_bundled.a $(SYSTEM_LIBS)
+
+ocr_db_crnn: fetch_opencv ocr_db_crnn.o crnn_process.o db_post_process.o clipper.o
+	$(CC) $(SYSROOT_LINK) $(CXXFLAGS_LINK) ocr_db_crnn.o crnn_process.o db_post_process.o clipper.o  -o ocr_db_crnn  $(CXX_LIBS) $(LDFLAGS)
+
+ocr_db_crnn.o: ocr_db_crnn.cc
+	$(CC) $(SYSROOT_COMPLILE) $(CXX_DEFINES) $(CXX_INCLUDES) $(CXX_FLAGS) -o ocr_db_crnn.o -c ocr_db_crnn.cc
+
+crnn_process.o: fetch_opencv crnn_process.cc
+	$(CC) $(SYSROOT_COMPLILE) $(CXX_DEFINES) $(CXX_INCLUDES) $(CXX_FLAGS) -o crnn_process.o -c crnn_process.cc
+
+db_post_process.o: fetch_clipper fetch_opencv db_post_process.cc
+	$(CC) $(SYSROOT_COMPLILE) $(CXX_DEFINES) $(CXX_INCLUDES) $(CXX_FLAGS) -o db_post_process.o -c db_post_process.cc
+
+clipper.o: fetch_clipper
+	$(CC) $(SYSROOT_COMPLILE) $(CXX_DEFINES) $(CXX_INCLUDES) $(CXX_FLAGS) -o clipper.o -c clipper.cpp
+
+fetch_clipper:
+	@test -e clipper.hpp || \
+        ( echo "Fetch clipper " && \
+        wget -c https://paddle-inference-dist.cdn.bcebos.com/PaddleLite/Clipper/clipper.hpp)
+	@ test -e clipper.cpp || \
+        wget -c https://paddle-inference-dist.cdn.bcebos.com/PaddleLite/Clipper/clipper.cpp
+
+fetch_opencv:
+	@ test -d ${THIRD_PARTY_DIR} ||  mkdir ${THIRD_PARTY_DIR}
+	@ test -e ${THIRD_PARTY_DIR}/${OPENCV_VERSION}.tar.gz || \
+      (echo "fetch opencv libs" && \
+      wget -P ${THIRD_PARTY_DIR} https://paddle-inference-dist.bj.bcebos.com/${OPENCV_VERSION}.tar.gz)
+	@ test -d ${THIRD_PARTY_DIR}/${OPENCV_VERSION} || \
+      tar -zxvf ${THIRD_PARTY_DIR}/${OPENCV_VERSION}.tar.gz -C ${THIRD_PARTY_DIR}
+
+
+.PHONY: clean
+clean:
+	rm -f ocr_db_crnn.o clipper.o db_post_process.o crnn_process.o
+	rm -f ocr_db_crnn

deploy/lite/config.txt

+max_side_len  960
+det_db_thresh  0.3
+det_db_box_thresh  0.5
+det_db_unclip_ratio  2.0
\ No newline at end of file

deploy/lite/crnn_process.cc

+// Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "crnn_process.h" //NOLINT
+#include <algorithm>
+#include <memory>
+#include <string>
+
+const std::vector<int> rec_image_shape{3, 32, 320};
+
+cv::Mat CrnnResizeImg(cv::Mat img, float wh_ratio) {
+  int imgC, imgH, imgW;
+  imgC = rec_image_shape[0];
+  imgW = rec_image_shape[2];
+  imgH = rec_image_shape[1];
+
+  imgW = int(32 * wh_ratio);
+
+  float ratio = float(img.cols) / float(img.rows);
+  int resize_w, resize_h;
+  if (ceilf(imgH * ratio) > imgW)
+    resize_w = imgW;
+  else
+    resize_w = int(ceilf(imgH * ratio));
+  cv::Mat resize_img;
+  cv::resize(img, resize_img, cv::Size(resize_w, imgH), 0.f, 0.f,
+             cv::INTER_LINEAR);
+
+  return resize_img;
+}
+
+std::vector<std::string> ReadDict(std::string path) {
+  std::ifstream in(path);
+  std::string filename;
+  std::string line;
+  std::vector<std::string> m_vec;
+  if (in) {
+    while (getline(in, line)) {
+      m_vec.push_back(line);
+    }
+  } else {
+    std::cout << "no such file" << std::endl;
+  }
+  return m_vec;
+}
+
+cv::Mat GetRotateCropImage(cv::Mat srcimage,
+                           std::vector<std::vector<int>> box) {
+  cv::Mat image;
+  srcimage.copyTo(image);
+  std::vector<std::vector<int>> points = box;
+
+  int x_collect[4] = {box[0][0], box[1][0], box[2][0], box[3][0]};
+  int y_collect[4] = {box[0][1], box[1][1], box[2][1], box[3][1]};
+  int left = int(*std::min_element(x_collect, x_collect + 4));
+  int right = int(*std::max_element(x_collect, x_collect + 4));
+  int top = int(*std::min_element(y_collect, y_collect + 4));
+  int bottom = int(*std::max_element(y_collect, y_collect + 4));
+
+  cv::Mat img_crop;
+  image(cv::Rect(left, top, right - left, bottom - top)).copyTo(img_crop);
+
+  for (int i = 0; i < points.size(); i++) {
+    points[i][0] -= left;
+    points[i][1] -= top;
+  }
+
+  int img_crop_width = int(sqrt(pow(points[0][0] - points[1][0], 2) +
+                                pow(points[0][1] - points[1][1], 2)));
+  int img_crop_height = int(sqrt(pow(points[0][0] - points[3][0], 2) +
+                                 pow(points[0][1] - points[3][1], 2)));
+
+  cv::Point2f pts_std[4];
+  pts_std[0] = cv::Point2f(0., 0.);
+  pts_std[1] = cv::Point2f(img_crop_width, 0.);
+  pts_std[2] = cv::Point2f(img_crop_width, img_crop_height);
+  pts_std[3] = cv::Point2f(0.f, img_crop_height);
+
+  cv::Point2f pointsf[4];
+  pointsf[0] = cv::Point2f(points[0][0], points[0][1]);
+  pointsf[1] = cv::Point2f(points[1][0], points[1][1]);
+  pointsf[2] = cv::Point2f(points[2][0], points[2][1]);
+  pointsf[3] = cv::Point2f(points[3][0], points[3][1]);
+
+  cv::Mat M = cv::getPerspectiveTransform(pointsf, pts_std);
+
+  cv::Mat dst_img;
+  cv::warpPerspective(img_crop, dst_img, M,
+                      cv::Size(img_crop_width, img_crop_height),
+                      cv::BORDER_REPLICATE);
+
+  if (float(dst_img.rows) >= float(dst_img.cols) * 1.5) {
+    cv::Mat srcCopy = cv::Mat(dst_img.rows, dst_img.cols, dst_img.depth());
+    cv::transpose(dst_img, srcCopy);
+    cv::flip(srcCopy, srcCopy, 0);
+    return srcCopy;
+  } else {
+    return dst_img;
+  }
+}

deploy/lite/crnn_process.h

+// Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#pragma once
+
+#include <cstring>
+#include <fstream>
+#include <iostream>
+#include <memory>
+#include <string>
+#include <vector>
+
+#include "math.h" //NOLINT
+#include "opencv2/core.hpp"
+#include "opencv2/imgcodecs.hpp"
+#include "opencv2/imgproc.hpp"
+
+cv::Mat CrnnResizeImg(cv::Mat img, float wh_ratio);
+
+std::vector<std::string> ReadDict(std::string path);
+
+cv::Mat GetRotateCropImage(cv::Mat srcimage, std::vector<std::vector<int>> box);
+
+template <class ForwardIterator>
+inline size_t Argmax(ForwardIterator first, ForwardIterator last) {
+  return std::distance(first, std::max_element(first, last));
+}

deploy/lite/db_post_process.cc

+// Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "db_post_process.h" // NOLINT
+#include <algorithm>
+#include <utility>
+
+void GetContourArea(std::vector<std::vector<float>> box, float unclip_ratio,
+                    float &distance) {
+  int pts_num = 4;
+  float area = 0.0f;
+  float dist = 0.0f;
+  for (int i = 0; i < pts_num; i++) {
+    area += box[i][0] * box[(i + 1) % pts_num][1] -
+            box[i][1] * box[(i + 1) % pts_num][0];
+    dist += sqrtf((box[i][0] - box[(i + 1) % pts_num][0]) *
+                      (box[i][0] - box[(i + 1) % pts_num][0]) +
+                  (box[i][1] - box[(i + 1) % pts_num][1]) *
+                      (box[i][1] - box[(i + 1) % pts_num][1]));
+  }
+  area = fabs(float(area / 2.0));
+
+  distance = area * unclip_ratio / dist;
+}
+
+cv::RotatedRect Unclip(std::vector<std::vector<float>> box,
+                       float unclip_ratio) {
+  float distance = 1.0;
+
+  GetContourArea(box, unclip_ratio, distance);
+
+  ClipperLib::ClipperOffset offset;
+  ClipperLib::Path p;
+  p << ClipperLib::IntPoint(int(box[0][0]), int(box[0][1]))
+    << ClipperLib::IntPoint(int(box[1][0]), int(box[1][1]))
+    << ClipperLib::IntPoint(int(box[2][0]), int(box[2][1]))
+    << ClipperLib::IntPoint(int(box[3][0]), int(box[3][1]));
+  offset.AddPath(p, ClipperLib::jtRound, ClipperLib::etClosedPolygon);
+
+  ClipperLib::Paths soln;
+  offset.Execute(soln, distance);
+  std::vector<cv::Point2f> points;
+
+  for (int j = 0; j < soln.size(); j++) {
+    for (int i = 0; i < soln[soln.size() - 1].size(); i++) {
+      points.emplace_back(soln[j][i].X, soln[j][i].Y);
+    }
+  }
+  cv::RotatedRect res = cv::minAreaRect(points);
+
+  return res;
+}
+
+std::vector<std::vector<float>> Mat2Vector(cv::Mat mat) {
+  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 XsortFp32(std::vector<float> a, std::vector<float> b) {
+  if (a[0] != b[0])
+    return a[0] < b[0];
+  return false;
+}
+
+bool 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<int>>
+OrderPointsClockwise(std::vector<std::vector<int>> pts) {
+  std::vector<std::vector<int>> box = pts;
+  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]};
+
+  if (leftmost[0][1] > leftmost[1][1])
+    std::swap(leftmost[0], leftmost[1]);
+
+  if (rightmost[0][1] > rightmost[1][1])
+    std::swap(rightmost[0], rightmost[1]);
+
+  std::vector<std::vector<int>> rect = {leftmost[0], rightmost[0], rightmost[1],
+                                        leftmost[1]};
+  return rect;
+}
+
+std::vector<std::vector<float>> GetMiniBoxes(cv::RotatedRect box, float &ssid) {
+  ssid = std::max(box.size.width, box.size.height);
+
+  cv::Mat points;
+  cv::boxPoints(box, points);
+
+  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];
+  } else {
+    idx2 = array[2];
+    idx3 = array[3];
+  }
+  if (array[1][1] <= array[0][1]) {
+    idx1 = array[1];
+    idx4 = array[0];
+  } else {
+    idx1 = array[0];
+    idx4 = array[1];
+  }
+
+  array[0] = idx1;
+  array[1] = idx2;
+  array[2] = idx3;
+  array[3] = idx4;
+
+  return array;
+}
+
+float BoxScoreFast(std::vector<std::vector<float>> box_array, cv::Mat pred) {
+  auto array = box_array;
+  int width = pred.cols;
+  int height = pred.rows;
+
+  float box_x[4] = {array[0][0], array[1][0], array[2][0], array[3][0]};
+  float box_y[4] = {array[0][1], array[1][1], array[2][1], array[3][1]};
+
+  int xmin = clamp(int(std::floorf(*(std::min_element(box_x, box_x + 4)))), 0,
+                   width - 1);
+  int xmax = clamp(int(std::ceilf(*(std::max_element(box_x, box_x + 4)))), 0,
+                   width - 1);
+  int ymin = clamp(int(std::floorf(*(std::min_element(box_y, box_y + 4)))), 0,
+                   height - 1);
+  int ymax = clamp(int(std::ceilf(*(std::max_element(box_y, box_y + 4)))), 0,
+                   height - 1);
+
+  cv::Mat mask;
+  mask = cv::Mat::zeros(ymax - ymin + 1, xmax - xmin + 1, CV_8UC1);
+
+  cv::Point root_point[4];
+  root_point[0] = cv::Point(int(array[0][0]) - xmin, int(array[0][1]) - ymin);
+  root_point[1] = cv::Point(int(array[1][0]) - xmin, int(array[1][1]) - ymin);
+  root_point[2] = cv::Point(int(array[2][0]) - xmin, int(array[2][1]) - ymin);
+  root_point[3] = cv::Point(int(array[3][0]) - xmin, int(array[3][1]) - ymin);
+  const cv::Point *ppt[1] = {root_point};
+  int npt[] = {4};
+  cv::fillPoly(mask, ppt, npt, 1, cv::Scalar(1));
+
+  cv::Mat croppedImg;
+  pred(cv::Rect(xmin, ymin, xmax - xmin + 1, ymax - ymin + 1))
+      .copyTo(croppedImg);
+
+  auto score = cv::mean(croppedImg, mask)[0];
+  return score;
+}
+
+std::vector<std::vector<std::vector<int>>>
+BoxesFromBitmap(const cv::Mat pred, const cv::Mat bitmap,
+                std::map<std::string, double> Config) {
+  const int min_size = 3;
+  const int max_candidates = 1000;
+  const float box_thresh = float(Config["det_db_box_thresh"]);
+  const float unclip_ratio = float(Config["det_db_unclip_ratio"]);
+
+  int width = bitmap.cols;
+  int height = bitmap.rows;
+
+  std::vector<std::vector<cv::Point>> contours;
+  std::vector<cv::Vec4i> hierarchy;
+
+  cv::findContours(bitmap, contours, hierarchy, cv::RETR_LIST,
+                   cv::CHAIN_APPROX_SIMPLE);
+
+  int num_contours =
+      contours.size() >= max_candidates ? max_candidates : contours.size();
+
+  std::vector<std::vector<std::vector<int>>> boxes;
+
+  for (int i = 0; i < num_contours; i++) {
+    float ssid;
+    cv::RotatedRect box = cv::minAreaRect(contours[i]);
+    auto array = GetMiniBoxes(box, ssid);
+
+    auto box_for_unclip = array;
+    // end get_mini_box
+
+    if (ssid < min_size) {
+      continue;
+    }
+
+    float score;
+    score = BoxScoreFast(array, pred);
+    // end box_score_fast
+    if (score < box_thresh)
+      continue;
+
+    // start for unclip
+    cv::RotatedRect points = Unclip(box_for_unclip, unclip_ratio);
+    // end for unclip
+
+    cv::RotatedRect clipbox = points;
+    auto cliparray = GetMiniBoxes(clipbox, ssid);
+
+    if (ssid < min_size + 2)
+      continue;
+
+    int dest_width = pred.cols;
+    int dest_height = pred.rows;
+    std::vector<std::vector<int>> intcliparray;
+
+    for (int num_pt = 0; num_pt < 4; num_pt++) {
+      std::vector<int> a{int(clamp(roundf(cliparray[num_pt][0] / float(width) *
+                                          float(dest_width)),
+                                   float(0), float(dest_width))),
+                         int(clamp(roundf(cliparray[num_pt][1] / float(height) *
+                                          float(dest_height)),
+                                   float(0), float(dest_height)))};
+      intcliparray.push_back(a);
+    }
+    boxes.push_back(intcliparray);
+
+  } // end for
+  return boxes;
+}
+
+std::vector<std::vector<std::vector<int>>>
+FilterTagDetRes(std::vector<std::vector<std::vector<int>>> boxes, float ratio_h,
+                float ratio_w, cv::Mat srcimg) {
+  int oriimg_h = srcimg.rows;
+  int oriimg_w = srcimg.cols;
+
+  std::vector<std::vector<std::vector<int>>> root_points;
+  for (int n = 0; n < boxes.size(); 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;
+
+      boxes[n][m][0] = int(std::min(std::max(boxes[n][m][0], 0), oriimg_w - 1));
+      boxes[n][m][1] = int(std::min(std::max(boxes[n][m][1], 0), oriimg_h - 1));
+    }
+  }
+
+  for (int n = 0; n < boxes.size(); n++) {
+    int rect_width, rect_height;
+    rect_width = int(sqrt(pow(boxes[n][0][0] - boxes[n][1][0], 2) +
+                          pow(boxes[n][0][1] - boxes[n][1][1], 2)));
+    rect_height = int(sqrt(pow(boxes[n][0][0] - boxes[n][3][0], 2) +
+                           pow(boxes[n][0][1] - boxes[n][3][1], 2)));
+    if (rect_width <= 10 || rect_height <= 10)
+      continue;
+    root_points.push_back(boxes[n]);
+  }
+  return root_points;
+}

deploy/lite/db_post_process.h

+// Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#pragma once
+
+#include <math.h>
+
+#include <iostream>
+#include <map>
+#include <vector>
+
+#include "clipper.hpp"
+#include "opencv2/core.hpp"
+#include "opencv2/imgcodecs.hpp"
+#include "opencv2/imgproc.hpp"
+
+template <class T> T clamp(T x, T min, T max) {
+  if (x > max)
+    return max;
+  if (x < min)
+    return min;
+  return x;
+}
+
+std::vector<std::vector<float>> Mat2Vector(cv::Mat mat);
+
+void GetContourArea(std::vector<std::vector<float>> box, float unclip_ratio,
+                    float &distance);
+
+cv::RotatedRect Unclip(std::vector<std::vector<float>> box, float unclip_ratio);
+
+std::vector<std::vector<float>> Mat2Vector(cv::Mat mat);
+
+bool XsortFp32(std::vector<float> a, std::vector<float> b);
+
+bool XsortInt(std::vector<int> a, std::vector<int> b);
+
+std::vector<std::vector<int>>
+OrderPointsClockwise(std::vector<std::vector<int>> pts);
+
+std::vector<std::vector<float>> GetMiniBoxes(cv::RotatedRect box, float &ssid);
+
+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,
+                std::map<std::string, double> Config);
+
+std::vector<std::vector<std::vector<int>>>
+FilterTagDetRes(std::vector<std::vector<std::vector<int>>> boxes, float ratio_h,
+                float ratio_w, cv::Mat srcimg);

deploy/lite/ocr_db_crnn.cc

+// Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "paddle_api.h" // NOLINT
+#include <chrono>
+
+#include "crnn_process.h"
+#include "db_post_process.h"
+
+using namespace paddle::lite_api; // NOLINT
+using namespace std;
+
+// fill tensor with mean and scale and trans layout: nhwc -> nchw, neon speed up
+void neon_mean_scale(const float *din, float *dout, int size,
+                     const std::vector<float> mean,
+                     const std::vector<float> scale) {
+  if (mean.size() != 3 || scale.size() != 3) {
+    std::cerr << "[ERROR] mean or scale size must equal to 3\n";
+    exit(1);
+  }
+  float32x4_t vmean0 = vdupq_n_f32(mean[0]);
+  float32x4_t vmean1 = vdupq_n_f32(mean[1]);
+  float32x4_t vmean2 = vdupq_n_f32(mean[2]);
+  float32x4_t vscale0 = vdupq_n_f32(scale[0]);
+  float32x4_t vscale1 = vdupq_n_f32(scale[1]);
+  float32x4_t vscale2 = vdupq_n_f32(scale[2]);
+
+  float *dout_c0 = dout;
+  float *dout_c1 = dout + size;
+  float *dout_c2 = dout + size * 2;
+
+  int i = 0;
+  for (; i < size - 3; i += 4) {
+    float32x4x3_t vin3 = vld3q_f32(din);
+    float32x4_t vsub0 = vsubq_f32(vin3.val[0], vmean0);
+    float32x4_t vsub1 = vsubq_f32(vin3.val[1], vmean1);
+    float32x4_t vsub2 = vsubq_f32(vin3.val[2], vmean2);
+    float32x4_t vs0 = vmulq_f32(vsub0, vscale0);
+    float32x4_t vs1 = vmulq_f32(vsub1, vscale1);
+    float32x4_t vs2 = vmulq_f32(vsub2, vscale2);
+    vst1q_f32(dout_c0, vs0);
+    vst1q_f32(dout_c1, vs1);
+    vst1q_f32(dout_c2, vs2);
+
+    din += 12;
+    dout_c0 += 4;
+    dout_c1 += 4;
+    dout_c2 += 4;
+  }
+  for (; i < size; i++) {
+    *(dout_c0++) = (*(din++) - mean[0]) * scale[0];
+    *(dout_c1++) = (*(din++) - mean[1]) * scale[1];
+    *(dout_c2++) = (*(din++) - mean[2]) * scale[2];
+  }
+}
+
+// resize image to a size multiple of 32 which is required by the network
+cv::Mat DetResizeImg(const cv::Mat img, int max_size_len,
+                     std::vector<float> &ratio_hw) {
+  int w = img.cols;
+  int h = img.rows;
+
+  float ratio = 1.f;
+  int max_wh = w >= h ? w : h;
+  if (max_wh > max_size_len) {
+    if (h > w) {
+      ratio = float(max_size_len) / float(h);
+    } else {
+      ratio = float(max_size_len) / float(w);
+    }
+  }
+
+  int resize_h = int(float(h) * ratio);
+  int resize_w = int(float(w) * ratio);
+  if (resize_h % 32 == 0)
+    resize_h = resize_h;
+  else if (resize_h / 32 < 1 + 1e-5)
+    resize_h = 32;
+  else
+    resize_h = (resize_h / 32 - 1) * 32;
+
+  if (resize_w % 32 == 0)
+    resize_w = resize_w;
+  else if (resize_w / 32 < 1 + 1e-5)
+    resize_w = 32;
+  else
+    resize_w = (resize_w / 32 - 1) * 32;
+
+  cv::Mat resize_img;
+  cv::resize(img, resize_img, cv::Size(resize_w, resize_h));
+
+  ratio_hw.push_back(float(resize_h) / float(h));
+  ratio_hw.push_back(float(resize_w) / float(w));
+  return resize_img;
+}
+
+void RunRecModel(std::vector<std::vector<std::vector<int>>> boxes, cv::Mat img,
+                 std::shared_ptr<PaddlePredictor> predictor_crnn,
+                 std::vector<std::string> &rec_text,
+                 std::vector<float> &rec_text_score,
+                 std::vector<std::string> charactor_dict) {
+  std::vector<float> mean = {0.5f, 0.5f, 0.5f};
+  std::vector<float> scale = {1 / 0.5f, 1 / 0.5f, 1 / 0.5f};
+
+  cv::Mat srcimg;
+  img.copyTo(srcimg);
+  cv::Mat crop_img;
+  cv::Mat resize_img;
+
+  int index = 0;
+  for (int i = boxes.size() - 1; i >= 0; i--) {
+    crop_img = GetRotateCropImage(srcimg, boxes[i]);
+    float wh_ratio = float(crop_img.cols) / float(crop_img.rows);
+
+    resize_img = CrnnResizeImg(crop_img, wh_ratio);
+    resize_img.convertTo(resize_img, CV_32FC3, 1 / 255.f);
+
+    const float *dimg = reinterpret_cast<const float *>(resize_img.data);
+
+    std::unique_ptr<Tensor> input_tensor0(
+        std::move(predictor_crnn->GetInput(0)));
+    input_tensor0->Resize({1, 3, resize_img.rows, resize_img.cols});
+    auto *data0 = input_tensor0->mutable_data<float>();
+
+    neon_mean_scale(dimg, data0, resize_img.rows * resize_img.cols, mean,
+                    scale);
+    //// Run CRNN predictor
+    predictor_crnn->Run();
+
+    // Get output and run postprocess
+    std::unique_ptr<const Tensor> output_tensor0(
+        std::move(predictor_crnn->GetOutput(0)));
+    auto *rec_idx = output_tensor0->data<int64>();
+
+    auto rec_idx_lod = output_tensor0->lod();
+    auto shape_out = output_tensor0->shape();
+
+    std::vector<int> pred_idx;
+    for (int n = int(rec_idx_lod[0][0]); n < int(rec_idx_lod[0][1]); n += 1) {
+      pred_idx.push_back(int(rec_idx[n]));
+    }
+
+    if (pred_idx.size() < 1e-3)
+      continue;
+
+    index += 1;
+    std::string pred_txt = "";
+    for (int n = 0; n < pred_idx.size(); n++) {
+      pred_txt += charactor_dict[pred_idx[n]];
+    }
+    rec_text.push_back(pred_txt);
+
+    ////get score
+    std::unique_ptr<const Tensor> output_tensor1(
+        std::move(predictor_crnn->GetOutput(1)));
+    auto *predict_batch = output_tensor1->data<float>();
+    auto predict_shape = output_tensor1->shape();
+
+    auto predict_lod = output_tensor1->lod();
+
+    int argmax_idx;
+    int blank = predict_shape[1];
+    float score = 0.f;
+    int count = 0;
+    float max_value = 0.0f;
+
+    for (int n = predict_lod[0][0]; n < predict_lod[0][1] - 1; n++) {
+      argmax_idx = int(Argmax(&predict_batch[n * predict_shape[1]],
+                              &predict_batch[(n + 1) * predict_shape[1]]));
+      max_value =
+          float(*std::max_element(&predict_batch[n * predict_shape[1]],
+                                  &predict_batch[(n + 1) * predict_shape[1]]));
+
+      if (blank - 1 - argmax_idx > 1e-5) {
+        score += max_value;
+        count += 1;
+      }
+    }
+    score /= count;
+    rec_text_score.push_back(score);
+  }
+}
+
+std::vector<std::vector<std::vector<int>>>
+RunDetModel(std::shared_ptr<PaddlePredictor> predictor, cv::Mat img,
+            std::map<std::string, double> Config) {
+  // Read img
+  int max_side_len = int(Config["max_side_len"]);
+
+  cv::Mat srcimg;
+  img.copyTo(srcimg);
+
+  std::vector<float> ratio_hw;
+  img = DetResizeImg(img, max_side_len, ratio_hw);
+  cv::Mat img_fp;
+  img.convertTo(img_fp, CV_32FC3, 1.0 / 255.f);
+
+  // Prepare input data from image
+  std::unique_ptr<Tensor> input_tensor0(std::move(predictor->GetInput(0)));
+  input_tensor0->Resize({1, 3, img_fp.rows, img_fp.cols});
+  auto *data0 = input_tensor0->mutable_data<float>();
+
+  std::vector<float> mean = {0.485f, 0.456f, 0.406f};
+  std::vector<float> scale = {1 / 0.229f, 1 / 0.224f, 1 / 0.225f};
+  const float *dimg = reinterpret_cast<const float *>(img_fp.data);
+  neon_mean_scale(dimg, data0, img_fp.rows * img_fp.cols, mean, scale);
+
+  // Run predictor
+  predictor->Run();
+
+  // Get output and post process
+  std::unique_ptr<const Tensor> output_tensor(
+      std::move(predictor->GetOutput(0)));
+  auto *outptr = output_tensor->data<float>();
+  auto shape_out = output_tensor->shape();
+
+  // Save output
+  float pred[shape_out[2] * shape_out[3]];
+  unsigned char cbuf[shape_out[2] * shape_out[3]];
+
+  for (int i = 0; i < int(shape_out[2] * shape_out[3]); i++) {
+    pred[i] = float(outptr[i]);
+    cbuf[i] = (unsigned char)((outptr[i]) * 255);
+  }
+
+  cv::Mat cbuf_map(shape_out[2], shape_out[3], CV_8UC1, (unsigned char *)cbuf);
+  cv::Mat pred_map(shape_out[2], shape_out[3], CV_32F, (float *)pred);
+
+  const double threshold = double(Config["det_db_thresh"]) * 255;
+  const double maxvalue = 255;
+  cv::Mat bit_map;
+  cv::threshold(cbuf_map, bit_map, threshold, maxvalue, cv::THRESH_BINARY);
+
+  auto boxes = BoxesFromBitmap(pred_map, bit_map, Config);
+
+  std::vector<std::vector<std::vector<int>>> filter_boxes =
+      FilterTagDetRes(boxes, ratio_hw[0], ratio_hw[1], srcimg);
+
+  return filter_boxes;
+}
+
+std::shared_ptr<PaddlePredictor> loadModel(std::string model_file) {
+  MobileConfig config;
+  config.set_model_from_file(model_file);
+
+  std::shared_ptr<PaddlePredictor> predictor =
+      CreatePaddlePredictor<MobileConfig>(config);
+  return predictor;
+}
+
+cv::Mat Visualization(cv::Mat srcimg,
+                      std::vector<std::vector<std::vector<int>>> boxes) {
+  cv::Point rook_points[boxes.size()][4];
+  for (int n = 0; n < boxes.size(); n++) {
+    for (int m = 0; m < boxes[0].size(); m++) {
+      rook_points[n][m] = cv::Point(int(boxes[n][m][0]), int(boxes[n][m][1]));
+    }
+  }
+  cv::Mat img_vis;
+  srcimg.copyTo(img_vis);
+  for (int n = 0; n < boxes.size(); n++) {
+    const cv::Point *ppt[1] = {rook_points[n]};
+    int npt[] = {4};
+    cv::polylines(img_vis, ppt, npt, 1, 1, CV_RGB(0, 255, 0), 2, 8, 0);
+  }
+
+  cv::imwrite("./vis.jpg", img_vis);
+  std::cout << "The detection visualized image saved in ./vis.jpg" << std::endl;
+  return img_vis;
+}
+
+std::vector<std::string> split(const std::string &str,
+                               const std::string &delim) {
+  std::vector<std::string> res;
+  if ("" == str)
+    return res;
+  char *strs = new char[str.length() + 1];
+  std::strcpy(strs, str.c_str());
+
+  char *d = new char[delim.length() + 1];
+  std::strcpy(d, delim.c_str());
+
+  char *p = std::strtok(strs, d);
+  while (p) {
+    string s = p;
+    res.push_back(s);
+    p = std::strtok(NULL, d);
+  }
+
+  return res;
+}
+
+std::map<std::string, double> LoadConfigTxt(std::string config_path) {
+  auto config = ReadDict(config_path);
+
+  std::map<std::string, double> dict;
+  for (int i = 0; i < config.size(); i++) {
+    std::vector<std::string> res = split(config[i], " ");
+    dict[res[0]] = stod(res[1]);
+  }
+  return dict;
+}
+
+int main(int argc, char **argv) {
+  if (argc < 5) {
+    std::cerr << "[ERROR] usage: " << argv[0]
+              << " det_model_file rec_model_file image_path\n";
+    exit(1);
+  }
+  std::string det_model_file = argv[1];
+  std::string rec_model_file = argv[2];
+  std::string img_path = argv[3];
+  std::string dict_path = argv[4];
+
+  //// load config from txt file
+  auto Config = LoadConfigTxt("./config.txt");
+
+  auto start = std::chrono::system_clock::now();
+
+  auto det_predictor = loadModel(det_model_file);
+  auto rec_predictor = loadModel(rec_model_file);
+
+  auto charactor_dict = ReadDict(dict_path);
+
+  cv::Mat srcimg = cv::imread(img_path, cv::IMREAD_COLOR);
+  auto boxes = RunDetModel(det_predictor, srcimg, Config);
+
+  std::vector<std::string> rec_text;
+  std::vector<float> rec_text_score;
+  RunRecModel(boxes, srcimg, rec_predictor, rec_text, rec_text_score,
+              charactor_dict);
+
+  auto end = std::chrono::system_clock::now();
+  auto duration =
+      std::chrono::duration_cast<std::chrono::microseconds>(end - start);
+
+  //// visualization
+  auto img_vis = Visualization(srcimg, boxes);
+
+  //// print recognized text
+  for (int i = 0; i < rec_text.size(); i++) {
+    std::cout << i << "\t" << rec_text[i] << "\t" << rec_text_score[i]
+              << std::endl;
+  }
+
+  std::cout << "花费了"
+            << double(duration.count()) *
+                   std::chrono::microseconds::period::num /
+                   std::chrono::microseconds::period::den
+            << "秒" << std::endl;
+
+  return 0;
+}

deploy/lite/readme.md

+# PaddleOCR 端侧模型部署
+
+本教程将介绍在移动端部署PaddleOCR超轻量中文检测、识别模型的详细步骤。
+
+
+## 1. 准备环境
+
+### 运行准备
+- 电脑（编译Paddle-Lite）
+- 安卓手机（armv7或armv8）
+
+### 1.1 准备交叉编译环境
+交叉编译环境用于编译[Paddle-Lite](https://github.com/PaddlePaddle/Paddle-Lite)和PaddleOCR的C++ demo。
+支持多种开发环境，不同开发环境的编译流程请参考对应文档。
+1. [Docker](https://paddle-lite.readthedocs.io/zh/latest/user_guides/source_compile.html#docker)
+2. [Linux](https://paddle-lite.readthedocs.io/zh/latest/user_guides/source_compile.html#android)
+3. [MAC OS](https://paddle-lite.readthedocs.io/zh/latest/user_guides/source_compile.html#id13)
+4. [Windows](https://paddle-lite.readthedocs.io/zh/latest/demo_guides/x86.html#windows)
+
+### 1.2 准备预测库
+
+预测库有两种获取方式：
+- 1. 直接下载，下载[链接](https://paddle-lite.readthedocs.io/zh/latest/user_guides/release_lib.html#android-toolchain-gcc).
+    注意选择`with_extra=ON，with_cv=ON`的下载链接。
+- 2. 编译Paddle-Lite得到，Paddle-Lite的编译方式如下：
+```
+git clone https://github.com/PaddlePaddle/Paddle-Lite.git
+cd Paddle-Lite
+git checkout 2.6.1
+./lite/tools/build_android.sh  --arch=armv8  --with_cv=ON --with_extra=ON
+```
+
+注意：编译Paddle-Lite获得预测库时，需要打开`--with_cv=ON --with_extra=ON`两个选项，`--arch`表示`arm`版本，这里指定为armv8，
+更多编译命令
+介绍请参考[链接](https://paddle-lite.readthedocs.io/zh/latest/user_guides/Compile/Android.html#id2)。
+
+直接下载预测库并解压后，可以得到`inference_lite_lib.android.armv8/`文件夹，通过编译Paddle-Lite得到的预测库位于
+`Paddle-Lite/build.lite.android.armv8.gcc/inference_lite_lib.android.armv8/`文件夹下。
+预测库的文件目录如下：
+```
+inference_lite_lib.android.armv8/
+|-- cxx                                        C++ 预测库和头文件
+|   |-- include                                C++ 头文件
+|   |   |-- paddle_api.h
+|   |   |-- paddle_image_preprocess.h
+|   |   |-- paddle_lite_factory_helper.h
+|   |   |-- paddle_place.h
+|   |   |-- paddle_use_kernels.h
+|   |   |-- paddle_use_ops.h
+|   |   `-- paddle_use_passes.h
+|   `-- lib                                           C++预测库
+|       |-- libpaddle_api_light_bundled.a             C++静态库
+|       `-- libpaddle_light_api_shared.so             C++动态库
+|-- java                                     Java预测库
+|   |-- jar
+|   |   `-- PaddlePredictor.jar
+|   |-- so
+|   |   `-- libpaddle_lite_jni.so
+|   `-- src
+|-- demo                                     C++和Java示例代码
+|   |-- cxx                                  C++  预测库demo
+|   `-- java                                 Java 预测库demo
+```
+
+## 2 开始运行
+
+### 2.1 模型优化
+
+Paddle-Lite 提供了多种策略来自动优化原始的模型，其中包括量化、子图融合、混合调度、Kernel优选等方法，使用Paddle-lite的opt工具可以自动
+对inference模型进行优化，优化后的模型更轻量，模型运行速度更快。
+
+下述表格中提供了优化好的超轻量中文模型：
+
+|模型简介|检测模型|识别模型|Paddle-Lite版本|
+|-|-|-|-|
+|超轻量级中文OCR opt优化模型|[下载地址](https://paddleocr.bj.bcebos.com/deploy/lite/ch_det_mv3_db_opt.nb)|[下载地址](https://paddleocr.bj.bcebos.com/deploy/lite/ch_rec_mv3_crnn_opt.nb)|2.6.1|
+
+如果直接使用上述表格中的模型进行部署，可略过下述步骤，直接阅读 [2.2节](###2.2与手机联调)。
+
+如果要部署的模型不在上述表格中，则需要按照如下步骤获得优化后的模型。
+
+模型优化需要Paddle-Lite的opt可执行文件，可以通过编译Paddle-Lite源码获得，编译步骤如下：
+```
+# 如果准备环境时已经clone了Paddle-Lite，则不用重新clone Paddle-Lite
+git clone https://github.com/PaddlePaddle/Paddle-Lite.git
+cd Paddle-Lite
+git checkout 2.6.1
+# 启动编译
+./lite/tools/build.sh build_optimize_tool
+```
+
+编译完成后，opt文件位于`build.opt/lite/api/`下，可通过如下方式查看opt的运行选项和使用方式；
+```
+cd build.opt/lite/api/
+./opt
+```
+
+|选项|说明|
+|-|-|
+|--model_dir|待优化的PaddlePaddle模型（非combined形式）的路径|
+|--model_file|待优化的PaddlePaddle模型（combined形式）的网络结构文件路径|
+|--param_file|待优化的PaddlePaddle模型（combined形式）的权重文件路径|
+|--optimize_out_type|输出模型类型，目前支持两种类型：protobuf和naive_buffer，其中naive_buffer是一种更轻量级的序列化/反序列化实现。若您需要在mobile端执行模型预测，请将此选项设置为naive_buffer。默认为protobuf|
+|--optimize_out|优化模型的输出路径|
+|--valid_targets|指定模型可执行的backend，默认为arm。目前可支持x86、arm、opencl、npu、xpu，可以同时指定多个backend(以空格分隔)，Model Optimize Tool将会自动选择最佳方式。如果需要支持华为NPU（Kirin 810/990 Soc搭载的达芬奇架构NPU），应当设置为npu, arm|
+|--record_tailoring_info|当使用 根据模型裁剪库文件 功能时，则设置该选项为true，以记录优化后模型含有的kernel和OP信息，默认为false|
+
+`--model_dir`适用于待优化的模型是非combined方式，PaddleOCR的inference模型是combined方式，即模型结构和模型参数使用单独一个文件存储。
+
+下面以PaddleOCR的超轻量中文模型为例，介绍使用编译好的opt文件完成inference模型到Paddle-Lite优化模型的转换。
+
+```
+# 下载PaddleOCR的超轻量文inference模型，并解压
+wget  https://paddleocr.bj.bcebos.com/ch_models/ch_det_mv3_db_infer.tar && tar xf ch_det_mv3_db_infer.tar
+wget  https://paddleocr.bj.bcebos.com/ch_models/ch_rec_mv3_crnn_infer.tar && tar xf ch_rec_mv3_crnn_infer.tar
+
+# 转换检测模型
+./opt --model_file=./ch_det_mv3_db/model --param_file=./ch_det_mv3_db/params --optimize_out_type=naive_buffer --optimize_out=./ch_det_mv3_db_opt --valid_targets=arm
+
+# 转换识别模型
+./opt --model_file=./ch_rec_mv3_crnn/model --param_file=./ch_rec_mv3_crnn/params --optimize_out_type=naive_buffer --optimize_out=./ch_rec_mv3_crnn_opt --valid_targets=arm
+```
+
+转换成功后，当前目录下会多出`ch_det_mv3_db_opt.nb`, `ch_rec_mv3_crnn_opt.nb`结尾的文件，即是转换成功的模型文件。
+
+注意：使用paddle-lite部署时，需要使用opt工具优化后的模型。 opt 转换的输入模型是paddle保存的inference模型
+
+### 2.2 与手机联调
+
+首先需要进行一些准备工作。
+ 1. 准备一台arm8的安卓手机，如果编译的预测库和opt文件是armv7，则需要arm7的手机，并修改Makefile中`ARM_ABI = arm7`。
+ 2. 打开手机的USB调试选项，选择文件传输模式，连接电脑。
+ 3. 电脑上安装adb工具，用于调试。 adb安装方式如下：
+
+    3.1. MAC电脑安装ADB:
+    ```
+    brew cask install android-platform-tools
+    ```
+    3.2. Linux安装ADB
+    ```
+    sudo apt update
+    sudo apt install -y wget adb
+    ```
+    3.3. Window安装ADB
+
+    win上安装需要去谷歌的安卓平台下载adb软件包进行安装：[链接](https://developer.android.com/studio)
+
+    打开终端，手机连接电脑，在终端中输入
+    ```
+    adb devices
+    ```
+    如果有device输出，则表示安装成功。
+    ```
+       List of devices attached
+       744be294    device
+    ```
+
+ 4. 准备优化后的模型、预测库文件、测试图像和使用的字典文件。
+ 在预测库`inference_lite_lib.android.armv8/demo/cxx/`下新建一个`ocr/`文件夹，
+ 将PaddleOCR repo中`PaddleOCR/deploy/lite/` 下的除`readme.md`所有文件放在新建的ocr文件夹下。在`ocr`文件夹下新建一个`debug`文件夹，
+ 将C++预测库so文件复制到debug文件夹下。
+  ```
+ # 进入OCR demo的工作目录
+ cd demo/cxx/ocr/
+ # 将C++预测动态库so文件复制到debug文件夹中
+ cp ../../../../cxx/lib/libpaddle_light_api_shared.so ./debug/
+ ```
+ 准备测试图像，以`PaddleOCR/doc/imgs/11.jpg`为例，将测试的图像复制到`demo/cxx/ocr/debug/`文件夹下。
+ 准备字典文件，中文超轻量模型的字典文件是`PaddleOCR/ppocr/utils/ppocr_keys_v1.txt`，将其复制到`demo/cxx/ocr/debug/`文件夹下。
+
+ 执行完成后，ocr文件夹下将有如下文件格式：
+
+```
+demo/cxx/ocr/
+|-- debug/  
+|   |--ch_det_mv3_db_opt.nb             优化后的检测模型文件
+|   |--ch_rec_mv3_crnn_opt.nb           优化后的识别模型文件
+|   |--11.jpg                           待测试图像
+|   |--ppocr_keys_v1.txt                字典文件
+|   |--libpaddle_light_api_shared.so    C++预测库文件
+|   |--config.txt                       DB-CRNN超参数配置
+|-- config.txt                  DB-CRNN超参数配置
+|-- crnn_process.cc             识别模型CRNN的预处理和后处理文件
+|-- crnn_process.h
+|-- db_post_process.cc          检测模型DB的后处理文件
+|-- db_post_process.h
+|-- Makefile                    编译文件
+|-- ocr_db_crnn.cc              C++预测源文件
+
+```
+
+ 5. 启动调试
+
+ 上述步骤完成后就可以使用adb将文件push到手机上运行，步骤如下：
+
+ ```
+ # 执行编译，得到可执行文件ocr_db_crnn
+ # ocr_db_crnn可执行文件的使用方式为:
+ # ./ocr_db_crnn  检测模型文件  识别模型文件  测试图像路径
+ make
+ # 将编译的可执行文件移动到debug文件夹中
+ mv ocr_db_crnn ./debug/
+ # 将debug文件夹push到手机上
+ adb push debug /data/local/tmp/
+ adb shell
+ cd /data/local/tmp/debug
+ export LD_LIBRARY_PATH=/data/local/tmp/debug:$LD_LIBRARY_PATH
+ ./ocr_db_crnn ch_det_mv3_db_opt.nb  ch_rec_mv3_crnn_opt.nb ./11.jpg  ppocr_keys_v1.txt
+ ```
+
+ 如果对代码做了修改，则需要重新编译并push到手机上。
+
+ 运行效果如下：
+ ![](..imgs/demo.png)