Merge branch 'dygraph' of https://github.com/PaddlePaddle/PaddleOCR into dygraph

deploy/android_demo/app/src/main/cpp/ocr_db_post_process.cpp

+// 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 "ocr_clipper.hpp"
+#include "opencv2/core.hpp"
+#include "opencv2/imgcodecs.hpp"
+#include "opencv2/imgproc.hpp"
+#include <iostream>
+#include <math.h>
+#include <vector>
+
+static void getcontourarea(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;
+}
+
+static cv::RotatedRect unclip(float **box) {
+  float unclip_ratio = 2.0;
+  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;
+}
+
+static float **Mat2Vec(cv::Mat mat) {
+  auto **array = new float *[mat.rows];
+  for (int i = 0; i < mat.rows; ++i) {
+    array[i] = new float[mat.cols];
+  }
+  for (int i = 0; i < mat.rows; ++i) {
+    for (int j = 0; j < mat.cols; ++j) {
+      array[i][j] = mat.at<float>(i, j);
+    }
+  }
+
+  return array;
+}
+
+static void 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);
+  }
+}
+
+static void 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);
+  }
+}
+
+static std::vector<std::vector<int>>
+order_points_clockwise(std::vector<std::vector<int>> pts) {
+  std::vector<std::vector<int>> box = pts;
+  quickSort_vector(box, 0, int(box.size() - 1), 0);
+  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;
+}
+
+static float **get_mini_boxes(cv::RotatedRect box, float &ssid) {
+  ssid = box.size.width >= box.size.height ? box.size.height : box.size.width;
+
+  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];
+  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;
+}
+
+template <class T> T clamp(T x, T min, T max) {
+  if (x > max) {
+    return max;
+  }
+  if (x < min) {
+    return min;
+  }
+  return x;
+}
+
+static float clampf(float x, float min, float max) {
+  if (x > max)
+    return max;
+  if (x < min)
+    return min;
+  return x;
+}
+
+float box_score_fast(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>>>
+boxes_from_bitmap(const cv::Mat &pred, const cv::Mat &bitmap) {
+  const int min_size = 3;
+  const int max_candidates = 1000;
+  const float box_thresh = 0.5;
+
+  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 = get_mini_boxes(box, ssid);
+
+    auto box_for_unclip = array;
+    // end get_mini_box
+
+    if (ssid < min_size) {
+      continue;
+    }
+
+    float score;
+    score = box_score_fast(array, pred);
+    // end box_score_fast
+    if (score < box_thresh) {
+      continue;
+    }
+
+    // start for unclip
+    cv::RotatedRect points = unclip(box_for_unclip);
+    // end for unclip
+
+    cv::RotatedRect clipbox = points;
+    auto cliparray = get_mini_boxes(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(clampf(roundf(cliparray[num_pt][0] / float(width) *
+                                           float(dest_width)),
+                                    0, float(dest_width))),
+                         int(clampf(roundf(cliparray[num_pt][1] /
+                                           float(height) * float(dest_height)),
+                                    0, float(dest_height)))};
+      intcliparray.emplace_back(std::move(a));
+    }
+    boxes.emplace_back(std::move(intcliparray));
+
+  } // end for
+  return boxes;
+}
+
+int _max(int a, int b) { return a >= b ? a : b; }
+
+int _min(int a, int b) { return a >= b ? b : a; }
+
+std::vector<std::vector<std::vector<int>>>
+filter_tag_det_res(const std::vector<std::vector<std::vector<int>>> &o_boxes,
+                   float ratio_h, float ratio_w, const cv::Mat &srcimg) {
+  int oriimg_h = srcimg.rows;
+  int oriimg_w = srcimg.cols;
+  std::vector<std::vector<std::vector<int>>> boxes{o_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]);
+    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(_min(_max(boxes[n][m][0], 0), oriimg_w - 1));
+      boxes[n][m][1] = int(_min(_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;
+}
\ No newline at end of file

deploy/android_demo/app/src/main/cpp/ocr_db_post_process.h

+//
+// Created by fujiayi on 2020/7/2.
+//
+#pragma once
+#include <opencv2/opencv.hpp>
+#include <vector>
+
+std::vector<std::vector<std::vector<int>>>
+boxes_from_bitmap(const cv::Mat &pred, const cv::Mat &bitmap);
+
+std::vector<std::vector<std::vector<int>>>
+filter_tag_det_res(const std::vector<std::vector<std::vector<int>>> &o_boxes,
+                   float ratio_h, float ratio_w, const cv::Mat &srcimg);
\ No newline at end of file

deploy/android_demo/app/src/main/cpp/ocr_ppredictor.cpp

+//
+// Created by fujiayi on 2020/7/1.
+//
+
+#include "ocr_ppredictor.h"
+#include "common.h"
+#include "ocr_cls_process.h"
+#include "ocr_crnn_process.h"
+#include "ocr_db_post_process.h"
+#include "preprocess.h"
+
+namespace ppredictor {
+
+OCR_PPredictor::OCR_PPredictor(const OCR_Config &config) : _config(config) {}
+
+int OCR_PPredictor::init(const std::string &det_model_content,
+                         const std::string &rec_model_content,
+                         const std::string &cls_model_content) {
+  _det_predictor = std::unique_ptr<PPredictor>(
+      new PPredictor{_config.thread_num, NET_OCR, _config.mode});
+  _det_predictor->init_nb(det_model_content);
+
+  _rec_predictor = std::unique_ptr<PPredictor>(
+      new PPredictor{_config.thread_num, NET_OCR_INTERNAL, _config.mode});
+  _rec_predictor->init_nb(rec_model_content);
+
+  _cls_predictor = std::unique_ptr<PPredictor>(
+      new PPredictor{_config.thread_num, NET_OCR_INTERNAL, _config.mode});
+  _cls_predictor->init_nb(cls_model_content);
+  return RETURN_OK;
+}
+
+int OCR_PPredictor::init_from_file(const std::string &det_model_path,
+                                   const std::string &rec_model_path,
+                                   const std::string &cls_model_path) {
+  _det_predictor = std::unique_ptr<PPredictor>(
+      new PPredictor{_config.thread_num, NET_OCR, _config.mode});
+  _det_predictor->init_from_file(det_model_path);
+
+  _rec_predictor = std::unique_ptr<PPredictor>(
+      new PPredictor{_config.thread_num, NET_OCR_INTERNAL, _config.mode});
+  _rec_predictor->init_from_file(rec_model_path);
+
+  _cls_predictor = std::unique_ptr<PPredictor>(
+      new PPredictor{_config.thread_num, NET_OCR_INTERNAL, _config.mode});
+  _cls_predictor->init_from_file(cls_model_path);
+  return RETURN_OK;
+}
+/**
+ * for debug use, show result of First Step
+ * @param filter_boxes
+ * @param boxes
+ * @param srcimg
+ */
+static void
+visual_img(const std::vector<std::vector<std::vector<int>>> &filter_boxes,
+           const std::vector<std::vector<std::vector<int>>> &boxes,
+           const cv::Mat &srcimg) {
+  // visualization
+  cv::Point rook_points[filter_boxes.size()][4];
+  for (int n = 0; n < filter_boxes.size(); n++) {
+    for (int m = 0; m < filter_boxes[0].size(); m++) {
+      rook_points[n][m] =
+          cv::Point(int(filter_boxes[n][m][0]), int(filter_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);
+  }
+  // 调试用，自行替换需要修改的路径
+  cv::imwrite("/sdcard/1/vis.png", img_vis);
+}
+
+std::vector<OCRPredictResult>
+OCR_PPredictor::infer_ocr(const std::vector<int64_t> &dims,
+                          const float *input_data, int input_len, int net_flag,
+                          cv::Mat &origin) {
+  PredictorInput input = _det_predictor->get_first_input();
+  input.set_dims(dims);
+  input.set_data(input_data, input_len);
+  std::vector<PredictorOutput> results = _det_predictor->infer();
+  PredictorOutput &res = results.at(0);
+  std::vector<std::vector<std::vector<int>>> filtered_box = calc_filtered_boxes(
+      res.get_float_data(), res.get_size(), (int)dims[2], (int)dims[3], origin);
+  LOGI("Filter_box size %ld", filtered_box.size());
+  return infer_rec(filtered_box, origin);
+}
+
+std::vector<OCRPredictResult> OCR_PPredictor::infer_rec(
+    const std::vector<std::vector<std::vector<int>>> &boxes,
+    const cv::Mat &origin_img) {
+  std::vector<float> mean = {0.5f, 0.5f, 0.5f};
+  std::vector<float> scale = {1 / 0.5f, 1 / 0.5f, 1 / 0.5f};
+  std::vector<int64_t> dims = {1, 3, 0, 0};
+  std::vector<OCRPredictResult> ocr_results;
+
+  PredictorInput input = _rec_predictor->get_first_input();
+  for (auto bp = boxes.crbegin(); bp != boxes.crend(); ++bp) {
+    const std::vector<std::vector<int>> &box = *bp;
+    cv::Mat crop_img = get_rotate_crop_image(origin_img, box);
+    crop_img = infer_cls(crop_img);
+
+    float wh_ratio = float(crop_img.cols) / float(crop_img.rows);
+    cv::Mat input_image = crnn_resize_img(crop_img, wh_ratio);
+    input_image.convertTo(input_image, CV_32FC3, 1 / 255.0f);
+    const float *dimg = reinterpret_cast<const float *>(input_image.data);
+    int input_size = input_image.rows * input_image.cols;
+
+    dims[2] = input_image.rows;
+    dims[3] = input_image.cols;
+    input.set_dims(dims);
+
+    neon_mean_scale(dimg, input.get_mutable_float_data(), input_size, mean,
+                    scale);
+
+    std::vector<PredictorOutput> results = _rec_predictor->infer();
+    const float *predict_batch = results.at(0).get_float_data();
+    const std::vector<int64_t> predict_shape = results.at(0).get_shape();
+
+    OCRPredictResult res;
+
+    // ctc decode
+    int argmax_idx;
+    int last_index = 0;
+    float score = 0.f;
+    int count = 0;
+    float max_value = 0.0f;
+
+    for (int n = 0; n < predict_shape[1]; n++) {
+      argmax_idx = int(argmax(&predict_batch[n * predict_shape[2]],
+                              &predict_batch[(n + 1) * predict_shape[2]]));
+      max_value =
+          float(*std::max_element(&predict_batch[n * predict_shape[2]],
+                                  &predict_batch[(n + 1) * predict_shape[2]]));
+      if (argmax_idx > 0 && (!(n > 0 && argmax_idx == last_index))) {
+        score += max_value;
+        count += 1;
+        res.word_index.push_back(argmax_idx);
+      }
+      last_index = argmax_idx;
+    }
+    score /= count;
+    if (res.word_index.empty()) {
+      continue;
+    }
+    res.score = score;
+    res.points = box;
+    ocr_results.emplace_back(std::move(res));
+  }
+  LOGI("ocr_results finished %lu", ocr_results.size());
+  return ocr_results;
+}
+
+cv::Mat OCR_PPredictor::infer_cls(const cv::Mat &img, float thresh) {
+  std::vector<float> mean = {0.5f, 0.5f, 0.5f};
+  std::vector<float> scale = {1 / 0.5f, 1 / 0.5f, 1 / 0.5f};
+  std::vector<int64_t> dims = {1, 3, 0, 0};
+  std::vector<OCRPredictResult> ocr_results;
+
+  PredictorInput input = _cls_predictor->get_first_input();
+
+  cv::Mat input_image = cls_resize_img(img);
+  input_image.convertTo(input_image, CV_32FC3, 1 / 255.0f);
+  const float *dimg = reinterpret_cast<const float *>(input_image.data);
+  int input_size = input_image.rows * input_image.cols;
+
+  dims[2] = input_image.rows;
+  dims[3] = input_image.cols;
+  input.set_dims(dims);
+
+  neon_mean_scale(dimg, input.get_mutable_float_data(), input_size, mean,
+                  scale);
+
+  std::vector<PredictorOutput> results = _cls_predictor->infer();
+
+  const float *scores = results.at(0).get_float_data();
+  float score = 0;
+  int label = 0;
+  for (int64_t i = 0; i < results.at(0).get_size(); i++) {
+    LOGI("output scores [%f]", scores[i]);
+    if (scores[i] > score) {
+      score = scores[i];
+      label = i;
+    }
+  }
+  cv::Mat srcimg;
+  img.copyTo(srcimg);
+  if (label % 2 == 1 && score > thresh) {
+    cv::rotate(srcimg, srcimg, 1);
+  }
+  return srcimg;
+}
+
+std::vector<std::vector<std::vector<int>>>
+OCR_PPredictor::calc_filtered_boxes(const float *pred, int pred_size,
+                                    int output_height, int output_width,
+                                    const cv::Mat &origin) {
+  const double threshold = 0.3;
+  const double maxvalue = 1;
+
+  cv::Mat pred_map = cv::Mat::zeros(output_height, output_width, CV_32F);
+  memcpy(pred_map.data, pred, pred_size * sizeof(float));
+  cv::Mat cbuf_map;
+  pred_map.convertTo(cbuf_map, CV_8UC1);
+
+  cv::Mat bit_map;
+  cv::threshold(cbuf_map, bit_map, threshold, maxvalue, cv::THRESH_BINARY);
+
+  std::vector<std::vector<std::vector<int>>> boxes =
+      boxes_from_bitmap(pred_map, bit_map);
+  float ratio_h = output_height * 1.0f / origin.rows;
+  float ratio_w = output_width * 1.0f / origin.cols;
+  std::vector<std::vector<std::vector<int>>> filter_boxes =
+      filter_tag_det_res(boxes, ratio_h, ratio_w, origin);
+  return filter_boxes;
+}
+
+std::vector<int>
+OCR_PPredictor::postprocess_rec_word_index(const PredictorOutput &res) {
+  const int *rec_idx = res.get_int_data();
+  const std::vector<std::vector<uint64_t>> rec_idx_lod = res.get_lod();
+
+  std::vector<int> pred_idx;
+  for (int n = int(rec_idx_lod[0][0]); n < int(rec_idx_lod[0][1] * 2); n += 2) {
+    pred_idx.emplace_back(rec_idx[n]);
+  }
+  return pred_idx;
+}
+
+float OCR_PPredictor::postprocess_rec_score(const PredictorOutput &res) {
+  const float *predict_batch = res.get_float_data();
+  const std::vector<int64_t> predict_shape = res.get_shape();
+  const std::vector<std::vector<uint64_t>> predict_lod = res.get_lod();
+  int blank = predict_shape[1];
+  float score = 0.f;
+  int count = 0;
+  for (int n = predict_lod[0][0]; n < predict_lod[0][1] - 1; n++) {
+    int argmax_idx = argmax(predict_batch + n * predict_shape[1],
+                            predict_batch + (n + 1) * predict_shape[1]);
+    float max_value = predict_batch[n * predict_shape[1] + argmax_idx];
+    if (blank - 1 - argmax_idx > 1e-5) {
+      score += max_value;
+      count += 1;
+    }
+  }
+  if (count == 0) {
+    LOGE("calc score count 0");
+  } else {
+    score /= count;
+  }
+  LOGI("calc score: %f", score);
+  return score;
+}
+
+NET_TYPE OCR_PPredictor::get_net_flag() const { return NET_OCR; }
+}
\ No newline at end of file

deploy/android_demo/app/src/main/cpp/ocr_ppredictor.h

+//
+// Created by fujiayi on 2020/7/1.
+//
+
+#pragma once
+
+#include "ppredictor.h"
+#include <opencv2/opencv.hpp>
+#include <paddle_api.h>
+#include <string>
+
+namespace ppredictor {
+
+/**
+ * Config
+ */
+struct OCR_Config {
+  int thread_num = 4; // Thread num
+  paddle::lite_api::PowerMode mode =
+      paddle::lite_api::LITE_POWER_HIGH; // PaddleLite Mode
+};
+
+/**
+ * PolyGone Result
+ */
+struct OCRPredictResult {
+  std::vector<int> word_index;
+  std::vector<std::vector<int>> points;
+  float score;
+};
+
+/**
+ * OCR there are 2 models
+ * 1. First model（det），select polygones to show where are the texts
+ * 2. crop from the origin images, use these polygones to infer
+ */
+class OCR_PPredictor : public PPredictor_Interface {
+public:
+  OCR_PPredictor(const OCR_Config &config);
+
+  virtual ~OCR_PPredictor() {}
+
+  /**
+   * 初始化二个模型的Predictor
+   * @param det_model_content
+   * @param rec_model_content
+   * @return
+   */
+  int init(const std::string &det_model_content,
+           const std::string &rec_model_content,
+           const std::string &cls_model_content);
+  int init_from_file(const std::string &det_model_path,
+                     const std::string &rec_model_path,
+                     const std::string &cls_model_path);
+  /**
+   * Return OCR result
+   * @param dims
+   * @param input_data
+   * @param input_len
+   * @param net_flag
+   * @param origin
+   * @return
+   */
+  virtual std::vector<OCRPredictResult>
+  infer_ocr(const std::vector<int64_t> &dims, const float *input_data,
+            int input_len, int net_flag, cv::Mat &origin);
+
+  virtual NET_TYPE get_net_flag() const;
+
+private:
+  /**
+   * calcul Polygone from the result image of first model
+   * @param pred
+   * @param output_height
+   * @param output_width
+   * @param origin
+   * @return
+   */
+  std::vector<std::vector<std::vector<int>>>
+  calc_filtered_boxes(const float *pred, int pred_size, int output_height,
+                      int output_width, const cv::Mat &origin);
+
+  /**
+   * infer for second model
+   *
+   * @param boxes
+   * @param origin
+   * @return
+   */
+  std::vector<OCRPredictResult>
+  infer_rec(const std::vector<std::vector<std::vector<int>>> &boxes,
+            const cv::Mat &origin);
+
+  /**
+  * infer for cls model
+  *
+  * @param boxes
+  * @param origin
+  * @return
+  */
+  cv::Mat infer_cls(const cv::Mat &origin, float thresh = 0.9);
+
+  /**
+   * Postprocess or sencod model to extract text
+   * @param res
+   * @return
+   */
+  std::vector<int> postprocess_rec_word_index(const PredictorOutput &res);
+
+  /**
+   * calculate confidence of second model text result
+   * @param res
+   * @return
+   */
+  float postprocess_rec_score(const PredictorOutput &res);
+
+  std::unique_ptr<PPredictor> _det_predictor;
+  std::unique_ptr<PPredictor> _rec_predictor;
+  std::unique_ptr<PPredictor> _cls_predictor;
+  OCR_Config _config;
+};
+}

deploy/android_demo/app/src/main/cpp/ppredictor.cpp

+#include "ppredictor.h"
+#include "common.h"
+
+namespace ppredictor {
+PPredictor::PPredictor(int thread_num, int net_flag,
+                       paddle::lite_api::PowerMode mode)
+    : _thread_num(thread_num), _net_flag(net_flag), _mode(mode) {}
+
+int PPredictor::init_nb(const std::string &model_content) {
+  paddle::lite_api::MobileConfig config;
+  config.set_model_from_buffer(model_content);
+  return _init(config);
+}
+
+int PPredictor::init_from_file(const std::string &model_content) {
+  paddle::lite_api::MobileConfig config;
+  config.set_model_from_file(model_content);
+  return _init(config);
+}
+
+template <typename ConfigT> int PPredictor::_init(ConfigT &config) {
+  config.set_threads(_thread_num);
+  config.set_power_mode(_mode);
+  _predictor = paddle::lite_api::CreatePaddlePredictor(config);
+  LOGI("paddle instance created");
+  return RETURN_OK;
+}
+
+PredictorInput PPredictor::get_input(int index) {
+  PredictorInput input{_predictor->GetInput(index), index, _net_flag};
+  _is_input_get = true;
+  return input;
+}
+
+std::vector<PredictorInput> PPredictor::get_inputs(int num) {
+  std::vector<PredictorInput> results;
+  for (int i = 0; i < num; i++) {
+    results.emplace_back(get_input(i));
+  }
+  return results;
+}
+
+PredictorInput PPredictor::get_first_input() { return get_input(0); }
+
+std::vector<PredictorOutput> PPredictor::infer() {
+  LOGI("infer Run start %d", _net_flag);
+  std::vector<PredictorOutput> results;
+  if (!_is_input_get) {
+    return results;
+  }
+  _predictor->Run();
+  LOGI("infer Run end");
+
+  for (int i = 0; i < _predictor->GetOutputNames().size(); i++) {
+    std::unique_ptr<const paddle::lite_api::Tensor> output_tensor =
+        _predictor->GetOutput(i);
+    LOGI("output tensor[%d] size %ld", i, product(output_tensor->shape()));
+    PredictorOutput result{std::move(output_tensor), i, _net_flag};
+    results.emplace_back(std::move(result));
+  }
+  return results;
+}
+
+NET_TYPE PPredictor::get_net_flag() const { return (NET_TYPE)_net_flag; }
+}
\ No newline at end of file

deploy/android_demo/app/src/main/cpp/ppredictor.h

+#pragma once
+
+#include "paddle_api.h"
+#include "predictor_input.h"
+#include "predictor_output.h"
+
+namespace ppredictor {
+
+/**
+ * PaddleLite Preditor Common Interface
+ */
+class PPredictor_Interface {
+public:
+  virtual ~PPredictor_Interface() {}
+
+  virtual NET_TYPE get_net_flag() const = 0;
+};
+
+/**
+ * Common Predictor
+ */
+class PPredictor : public PPredictor_Interface {
+public:
+  PPredictor(
+      int thread_num, int net_flag = 0,
+      paddle::lite_api::PowerMode mode = paddle::lite_api::LITE_POWER_HIGH);
+
+  virtual ~PPredictor() {}
+
+  /**
+   * init paddlitelite opt model，nb format ，or use ini_paddle
+   * @param model_content
+   * @return 0
+   */
+  virtual int init_nb(const std::string &model_content);
+
+  virtual int init_from_file(const std::string &model_content);
+
+  std::vector<PredictorOutput> infer();
+
+  std::shared_ptr<paddle::lite_api::PaddlePredictor> get_predictor() {
+    return _predictor;
+  }
+
+  virtual std::vector<PredictorInput> get_inputs(int num);
+
+  virtual PredictorInput get_input(int index);
+
+  virtual PredictorInput get_first_input();
+
+  virtual NET_TYPE get_net_flag() const;
+
+protected:
+  template <typename ConfigT> int _init(ConfigT &config);
+
+private:
+  int _thread_num;
+  paddle::lite_api::PowerMode _mode;
+  std::shared_ptr<paddle::lite_api::PaddlePredictor> _predictor;
+  bool _is_input_get = false;
+  int _net_flag;
+};
+}

deploy/android_demo/app/src/main/cpp/predictor_input.cpp

+#include "predictor_input.h"
+
+namespace ppredictor {
+
+void PredictorInput::set_dims(std::vector<int64_t> dims) {
+  // yolov3
+  if (_net_flag == 101 && _index == 1) {
+    _tensor->Resize({1, 2});
+    _tensor->mutable_data<int>()[0] = (int)dims.at(2);
+    _tensor->mutable_data<int>()[1] = (int)dims.at(3);
+  } else {
+    _tensor->Resize(dims);
+  }
+  _is_dims_set = true;
+}
+
+float *PredictorInput::get_mutable_float_data() {
+  if (!_is_dims_set) {
+    LOGE("PredictorInput::set_dims is not called");
+  }
+  return _tensor->mutable_data<float>();
+}
+
+void PredictorInput::set_data(const float *input_data, int input_float_len) {
+  float *input_raw_data = get_mutable_float_data();
+  memcpy(input_raw_data, input_data, input_float_len * sizeof(float));
+}
+}
\ No newline at end of file

deploy/android_demo/app/src/main/cpp/predictor_input.h

+#pragma once
+
+#include "common.h"
+#include <paddle_api.h>
+#include <vector>
+
+namespace ppredictor {
+class PredictorInput {
+public:
+  PredictorInput(std::unique_ptr<paddle::lite_api::Tensor> &&tensor, int index,
+                 int net_flag)
+      : _tensor(std::move(tensor)), _index(index), _net_flag(net_flag) {}
+
+  void set_dims(std::vector<int64_t> dims);
+
+  float *get_mutable_float_data();
+
+  void set_data(const float *input_data, int input_float_len);
+
+private:
+  std::unique_ptr<paddle::lite_api::Tensor> _tensor;
+  bool _is_dims_set = false;
+  int _index;
+  int _net_flag;
+};
+}

deploy/android_demo/app/src/main/cpp/predictor_output.cpp

+#include "predictor_output.h"
+namespace ppredictor {
+const float *PredictorOutput::get_float_data() const {
+  return _tensor->data<float>();
+}
+
+const int *PredictorOutput::get_int_data() const {
+  return _tensor->data<int>();
+}
+
+const std::vector<std::vector<uint64_t>> PredictorOutput::get_lod() const {
+  return _tensor->lod();
+}
+
+int64_t PredictorOutput::get_size() const {
+  if (_net_flag == NET_OCR) {
+    return _tensor->shape().at(2) * _tensor->shape().at(3);
+  } else {
+    return product(_tensor->shape());
+  }
+}
+
+const std::vector<int64_t> PredictorOutput::get_shape() const {
+  return _tensor->shape();
+}
+}
\ No newline at end of file

deploy/android_demo/app/src/main/cpp/predictor_output.h

+#pragma once
+
+#include "common.h"
+#include <paddle_api.h>
+#include <vector>
+
+namespace ppredictor {
+class PredictorOutput {
+public:
+  PredictorOutput() {}
+  PredictorOutput(std::unique_ptr<const paddle::lite_api::Tensor> &&tensor,
+                  int index, int net_flag)
+      : _tensor(std::move(tensor)), _index(index), _net_flag(net_flag) {}
+
+  const float *get_float_data() const;
+  const int *get_int_data() const;
+  int64_t get_size() const;
+  const std::vector<std::vector<uint64_t>> get_lod() const;
+  const std::vector<int64_t> get_shape() const;
+
+  std::vector<float> data;    // return float, or use data_int
+  std::vector<int> data_int;  // several layers return int ，or use data
+  std::vector<int64_t> shape; // PaddleLite output shape
+  std::vector<std::vector<uint64_t>> lod; // PaddleLite output lod
+
+private:
+  std::unique_ptr<const paddle::lite_api::Tensor> _tensor;
+  int _index;
+  int _net_flag;
+};
+}

deploy/android_demo/app/src/main/cpp/preprocess.cpp

+#include "preprocess.h"
+#include <android/bitmap.h>
+
+cv::Mat bitmap_to_cv_mat(JNIEnv *env, jobject bitmap) {
+  AndroidBitmapInfo info;
+  int result = AndroidBitmap_getInfo(env, bitmap, &info);
+  if (result != ANDROID_BITMAP_RESULT_SUCCESS) {
+    LOGE("AndroidBitmap_getInfo failed, result: %d", result);
+    return cv::Mat{};
+  }
+  if (info.format != ANDROID_BITMAP_FORMAT_RGBA_8888) {
+    LOGE("Bitmap format is not RGBA_8888 !");
+    return cv::Mat{};
+  }
+  unsigned char *srcData = NULL;
+  AndroidBitmap_lockPixels(env, bitmap, (void **)&srcData);
+  cv::Mat mat = cv::Mat::zeros(info.height, info.width, CV_8UC4);
+  memcpy(mat.data, srcData, info.height * info.width * 4);
+  AndroidBitmap_unlockPixels(env, bitmap);
+  cv::cvtColor(mat, mat, cv::COLOR_RGBA2BGR);
+  /**
+  if (!cv::imwrite("/sdcard/1/copy.jpg", mat)){
+      LOGE("Write image failed " );
+  }
+   */
+
+  return mat;
+}
+
+cv::Mat resize_img(const cv::Mat &img, int height, int width) {
+  if (img.rows == height && img.cols == width) {
+    return img;
+  }
+  cv::Mat new_img;
+  cv::resize(img, new_img, cv::Size(height, width));
+  return new_img;
+}
+
+// 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) {
+    LOGE("[ERROR] mean or scale size must equal to 3");
+    return;
+  }
+
+  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];
+  }
+}
\ No newline at end of file

deploy/android_demo/app/src/main/cpp/preprocess.h

+#pragma once
+
+#include "common.h"
+#include <jni.h>
+#include <opencv2/opencv.hpp>
+cv::Mat bitmap_to_cv_mat(JNIEnv *env, jobject bitmap);
+
+cv::Mat resize_img(const cv::Mat &img, int height, int width);
+
+void neon_mean_scale(const float *din, float *dout, int size,
+                     const std::vector<float> &mean,
+                     const std::vector<float> &scale);

deploy/android_demo/app/src/main/java/com/baidu/paddle/lite/demo/ocr/AppCompatPreferenceActivity.java

+/*
+ * Copyright (C) 2014 The Android Open Source Project
+ *
+ * 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.
+ */
+
+package com.baidu.paddle.lite.demo.ocr;
+
+import android.content.res.Configuration;
+import android.os.Bundle;
+import android.preference.PreferenceActivity;
+import android.view.MenuInflater;
+import android.view.View;
+import android.view.ViewGroup;
+
+import androidx.annotation.LayoutRes;
+import androidx.annotation.Nullable;
+import androidx.appcompat.app.ActionBar;
+import androidx.appcompat.app.AppCompatDelegate;
+import androidx.appcompat.widget.Toolbar;
+
+/**
+ * A {@link PreferenceActivity} which implements and proxies the necessary calls
+ * to be used with AppCompat.
+ * <p>
+ * This technique can be used with an {@link android.app.Activity} class, not just
+ * {@link PreferenceActivity}.
+ */
+public abstract class AppCompatPreferenceActivity extends PreferenceActivity {
+    private AppCompatDelegate mDelegate;
+
+    @Override
+    protected void onCreate(Bundle savedInstanceState) {
+        getDelegate().installViewFactory();
+        getDelegate().onCreate(savedInstanceState);
+        super.onCreate(savedInstanceState);
+    }
+
+    @Override
+    protected void onPostCreate(Bundle savedInstanceState) {
+        super.onPostCreate(savedInstanceState);
+        getDelegate().onPostCreate(savedInstanceState);
+    }
+
+    public ActionBar getSupportActionBar() {
+        return getDelegate().getSupportActionBar();
+    }
+
+    public void setSupportActionBar(@Nullable Toolbar toolbar) {
+        getDelegate().setSupportActionBar(toolbar);
+    }
+
+    @Override
+    public MenuInflater getMenuInflater() {
+        return getDelegate().getMenuInflater();
+    }
+
+    @Override
+    public void setContentView(@LayoutRes int layoutResID) {
+        getDelegate().setContentView(layoutResID);
+    }
+
+    @Override
+    public void setContentView(View view) {
+        getDelegate().setContentView(view);
+    }
+
+    @Override
+    public void setContentView(View view, ViewGroup.LayoutParams params) {
+        getDelegate().setContentView(view, params);
+    }
+
+    @Override
+    public void addContentView(View view, ViewGroup.LayoutParams params) {
+        getDelegate().addContentView(view, params);
+    }
+
+    @Override
+    protected void onPostResume() {
+        super.onPostResume();
+        getDelegate().onPostResume();
+    }
+
+    @Override
+    protected void onTitleChanged(CharSequence title, int color) {
+        super.onTitleChanged(title, color);
+        getDelegate().setTitle(title);
+    }
+
+    @Override
+    public void onConfigurationChanged(Configuration newConfig) {
+        super.onConfigurationChanged(newConfig);
+        getDelegate().onConfigurationChanged(newConfig);
+    }
+
+    @Override
+    protected void onStop() {
+        super.onStop();
+        getDelegate().onStop();
+    }
+
+    @Override
+    protected void onDestroy() {
+        super.onDestroy();
+        getDelegate().onDestroy();
+    }
+
+    public void invalidateOptionsMenu() {
+        getDelegate().invalidateOptionsMenu();
+    }
+
+    private AppCompatDelegate getDelegate() {
+        if (mDelegate == null) {
+            mDelegate = AppCompatDelegate.create(this, null);
+        }
+        return mDelegate;
+    }
+}

deploy/android_demo/app/src/main/java/com/baidu/paddle/lite/demo/ocr/MiniActivity.java

+package com.baidu.paddle.lite.demo.ocr;
+
+import android.graphics.Bitmap;
+import android.graphics.BitmapFactory;
+import android.os.Build;
+import android.os.Bundle;
+import android.os.Handler;
+import android.os.HandlerThread;
+import android.os.Message;
+import android.util.Log;
+import android.view.View;
+import android.widget.Button;
+import android.widget.ImageView;
+import android.widget.TextView;
+import android.widget.Toast;
+
+import androidx.appcompat.app.AppCompatActivity;
+
+import java.io.IOException;
+import java.io.InputStream;
+
+public class MiniActivity extends AppCompatActivity {
+
+
+    public static final int REQUEST_LOAD_MODEL = 0;
+    public static final int REQUEST_RUN_MODEL = 1;
+    public static final int REQUEST_UNLOAD_MODEL = 2;
+    public static final int RESPONSE_LOAD_MODEL_SUCCESSED = 0;
+    public static final int RESPONSE_LOAD_MODEL_FAILED = 1;
+    public static final int RESPONSE_RUN_MODEL_SUCCESSED = 2;
+    public static final int RESPONSE_RUN_MODEL_FAILED = 3;
+
+    private static final String TAG = "MiniActivity";
+
+    protected Handler receiver = null; // Receive messages from worker thread
+    protected Handler sender = null; // Send command to worker thread
+    protected HandlerThread worker = null; // Worker thread to load&run model
+    protected volatile Predictor predictor = null;
+
+    private String assetModelDirPath = "models/ocr_v2_for_cpu";
+    private String assetlabelFilePath = "labels/ppocr_keys_v1.txt";
+
+    private Button button;
+    private ImageView imageView; // image result
+    private TextView textView; // text result
+
+    @Override
+    protected void onCreate(Bundle savedInstanceState) {
+        super.onCreate(savedInstanceState);
+        setContentView(R.layout.activity_mini);
+
+        Log.i(TAG, "SHOW in Logcat");
+
+        // Prepare the worker thread for mode loading and inference
+        worker = new HandlerThread("Predictor Worker");
+        worker.start();
+        sender = new Handler(worker.getLooper()) {
+            public void handleMessage(Message msg) {
+                switch (msg.what) {
+                    case REQUEST_LOAD_MODEL:
+                        // Load model and reload test image
+                        if (!onLoadModel()) {
+                            runOnUiThread(new Runnable() {
+                                @Override
+                                public void run() {
+                                    Toast.makeText(MiniActivity.this, "Load model failed!", Toast.LENGTH_SHORT).show();
+                                }
+                            });
+                        }
+                        break;
+                    case REQUEST_RUN_MODEL:
+                        // Run model if model is loaded
+                        final boolean isSuccessed = onRunModel();
+                        runOnUiThread(new Runnable() {
+                            @Override
+                            public void run() {
+                                if (isSuccessed){
+                                    onRunModelSuccessed();
+                                }else{
+                                    Toast.makeText(MiniActivity.this, "Run model failed!", Toast.LENGTH_SHORT).show();
+                                }
+                            }
+                        });
+                        break;
+                }
+            }
+        };
+        sender.sendEmptyMessage(REQUEST_LOAD_MODEL); // corresponding to REQUEST_LOAD_MODEL， to call onLoadModel()
+
+        imageView = findViewById(R.id.imageView);
+        textView = findViewById(R.id.sample_text);
+        button = findViewById(R.id.button);
+        button.setOnClickListener(new View.OnClickListener() {
+            @Override
+            public void onClick(View v) {
+                sender.sendEmptyMessage(REQUEST_RUN_MODEL);
+            }
+        });
+
+
+    }
+
+    @Override
+    protected void onDestroy() {
+        onUnloadModel();
+        if (Build.VERSION.SDK_INT >= Build.VERSION_CODES.JELLY_BEAN_MR2) {
+            worker.quitSafely();
+        } else {
+            worker.quit();
+        }
+        super.onDestroy();
+    }
+
+    /**
+     * call in onCreate, model init
+     *
+     * @return
+     */
+    private boolean onLoadModel() {
+        if (predictor == null) {
+            predictor = new Predictor();
+        }
+        return predictor.init(this, assetModelDirPath, assetlabelFilePath);
+    }
+
+    /**
+     * init engine
+     * call in onCreate
+     *
+     * @return
+     */
+    private boolean onRunModel() {
+        try {
+            String assetImagePath = "images/0.jpg";
+            InputStream imageStream = getAssets().open(assetImagePath);
+            Bitmap image = BitmapFactory.decodeStream(imageStream);
+            // Input is Bitmap
+            predictor.setInputImage(image);
+            return predictor.isLoaded() && predictor.runModel();
+        } catch (IOException e) {
+            e.printStackTrace();
+            return false;
+        }
+    }
+
+    private void onRunModelSuccessed() {
+        Log.i(TAG, "onRunModelSuccessed");
+        textView.setText(predictor.outputResult);
+        imageView.setImageBitmap(predictor.outputImage);
+    }
+
+    private void onUnloadModel() {
+        if (predictor != null) {
+            predictor.releaseModel();
+        }
+    }
+}

deploy/android_demo/app/src/main/java/com/baidu/paddle/lite/demo/ocr/OCRPredictorNative.java

+package com.baidu.paddle.lite.demo.ocr;
+
+import android.graphics.Bitmap;
+import android.util.Log;
+
+import java.util.ArrayList;
+import java.util.concurrent.atomic.AtomicBoolean;
+
+public class OCRPredictorNative {
+
+    private static final AtomicBoolean isSOLoaded = new AtomicBoolean();
+
+    public static void loadLibrary() throws RuntimeException {
+        if (!isSOLoaded.get() && isSOLoaded.compareAndSet(false, true)) {
+            try {
+                System.loadLibrary("Native");
+            } catch (Throwable e) {
+                RuntimeException exception = new RuntimeException(
+                        "Load libNative.so failed, please check it exists in apk file.", e);
+                throw exception;
+            }
+        }
+    }
+
+    private Config config;
+
+    private long nativePointer = 0;
+
+    public OCRPredictorNative(Config config) {
+        this.config = config;
+        loadLibrary();
+        nativePointer = init(config.detModelFilename, config.recModelFilename,config.clsModelFilename,
+                config.cpuThreadNum, config.cpuPower);
+        Log.i("OCRPredictorNative", "load success " + nativePointer);
+
+    }
+
+
+    public ArrayList<OcrResultModel> runImage(float[] inputData, int width, int height, int channels, Bitmap originalImage) {
+        Log.i("OCRPredictorNative", "begin to run image " + inputData.length + " " + width + " " + height);
+        float[] dims = new float[]{1, channels, height, width};
+        float[] rawResults = forward(nativePointer, inputData, dims, originalImage);
+        ArrayList<OcrResultModel> results = postprocess(rawResults);
+        return results;
+    }
+
+    public static class Config {
+        public int cpuThreadNum;
+        public String cpuPower;
+        public String detModelFilename;
+        public String recModelFilename;
+        public String clsModelFilename;
+
+    }
+
+    public void destory(){
+        if (nativePointer > 0) {
+            release(nativePointer);
+            nativePointer = 0;
+        }
+    }
+
+    protected native long init(String detModelPath, String recModelPath,String clsModelPath, int threadNum, String cpuMode);
+
+    protected native float[] forward(long pointer, float[] buf, float[] ddims, Bitmap originalImage);
+
+    protected native void release(long pointer);
+
+    private ArrayList<OcrResultModel> postprocess(float[] raw) {
+        ArrayList<OcrResultModel> results = new ArrayList<OcrResultModel>();
+        int begin = 0;
+
+        while (begin < raw.length) {
+            int point_num = Math.round(raw[begin]);
+            int word_num = Math.round(raw[begin + 1]);
+            OcrResultModel model = parse(raw, begin + 2, point_num, word_num);
+            begin += 2 + 1 + point_num * 2 + word_num;
+            results.add(model);
+        }
+
+        return results;
+    }
+
+    private OcrResultModel parse(float[] raw, int begin, int pointNum, int wordNum) {
+        int current = begin;
+        OcrResultModel model = new OcrResultModel();
+        model.setConfidence(raw[current]);
+        current++;
+        for (int i = 0; i < pointNum; i++) {
+            model.addPoints(Math.round(raw[current + i * 2]), Math.round(raw[current + i * 2 + 1]));
+        }
+        current += (pointNum * 2);
+        for (int i = 0; i < wordNum; i++) {
+            int index = Math.round(raw[current + i]);
+            model.addWordIndex(index);
+        }
+        Log.i("OCRPredictorNative", "word finished " + wordNum);
+        return model;
+    }
+
+
+}

deploy/android_demo/app/src/main/java/com/baidu/paddle/lite/demo/ocr/OcrResultModel.java

+package com.baidu.paddle.lite.demo.ocr;
+
+import android.graphics.Point;
+
+import java.util.ArrayList;
+import java.util.List;
+
+public class OcrResultModel {
+    private List<Point> points;
+    private List<Integer> wordIndex;
+    private String label;
+    private float confidence;
+
+    public OcrResultModel() {
+        super();
+        points = new ArrayList<>();
+        wordIndex = new ArrayList<>();
+    }
+
+    public void addPoints(int x, int y) {
+        Point point = new Point(x, y);
+        points.add(point);
+    }
+
+    public void addWordIndex(int index) {
+        wordIndex.add(index);
+    }
+
+    public List<Point> getPoints() {
+        return points;
+    }
+
+    public List<Integer> getWordIndex() {
+        return wordIndex;
+    }
+
+    public String getLabel() {
+        return label;
+    }
+
+    public void setLabel(String label) {
+        this.label = label;
+    }
+
+    public float getConfidence() {
+        return confidence;
+    }
+
+    public void setConfidence(float confidence) {
+        this.confidence = confidence;
+    }
+}

deploy/android_demo/app/src/main/java/com/baidu/paddle/lite/demo/ocr/Predictor.java

+package com.baidu.paddle.lite.demo.ocr;
+
+import android.content.Context;
+import android.graphics.Bitmap;
+import android.graphics.Canvas;
+import android.graphics.Color;
+import android.graphics.Paint;
+import android.graphics.Path;
+import android.graphics.Point;
+import android.util.Log;
+
+import java.io.File;
+import java.io.InputStream;
+import java.util.ArrayList;
+import java.util.Date;
+import java.util.List;
+import java.util.Vector;
+
+import static android.graphics.Color.*;
+
+public class Predictor {
+    private static final String TAG = Predictor.class.getSimpleName();
+    public boolean isLoaded = false;
+    public int warmupIterNum = 1;
+    public int inferIterNum = 1;
+    public int cpuThreadNum = 4;
+    public String cpuPowerMode = "LITE_POWER_HIGH";
+    public String modelPath = "";
+    public String modelName = "";
+    protected OCRPredictorNative paddlePredictor = null;
+    protected float inferenceTime = 0;
+    // Only for object detection
+    protected Vector<String> wordLabels = new Vector<String>();
+    protected String inputColorFormat = "BGR";
+    protected long[] inputShape = new long[]{1, 3, 960};
+    protected float[] inputMean = new float[]{0.485f, 0.456f, 0.406f};
+    protected float[] inputStd = new float[]{1.0f / 0.229f, 1.0f / 0.224f, 1.0f / 0.225f};
+    protected float scoreThreshold = 0.1f;
+    protected Bitmap inputImage = null;
+    protected Bitmap outputImage = null;
+    protected volatile String outputResult = "";
+    protected float preprocessTime = 0;
+    protected float postprocessTime = 0;
+
+
+    public Predictor() {
+    }
+
+    public boolean init(Context appCtx, String modelPath, String labelPath) {
+        isLoaded = loadModel(appCtx, modelPath, cpuThreadNum, cpuPowerMode);
+        if (!isLoaded) {
+            return false;
+        }
+        isLoaded = loadLabel(appCtx, labelPath);
+        return isLoaded;
+    }
+
+
+    public boolean init(Context appCtx, String modelPath, String labelPath, int cpuThreadNum, String cpuPowerMode,
+                        String inputColorFormat,
+                        long[] inputShape, float[] inputMean,
+                        float[] inputStd, float scoreThreshold) {
+        if (inputShape.length != 3) {
+            Log.e(TAG, "Size of input shape should be: 3");
+            return false;
+        }
+        if (inputMean.length != inputShape[1]) {
+            Log.e(TAG, "Size of input mean should be: " + Long.toString(inputShape[1]));
+            return false;
+        }
+        if (inputStd.length != inputShape[1]) {
+            Log.e(TAG, "Size of input std should be: " + Long.toString(inputShape[1]));
+            return false;
+        }
+        if (inputShape[0] != 1) {
+            Log.e(TAG, "Only one batch is supported in the image classification demo, you can use any batch size in " +
+                    "your Apps!");
+            return false;
+        }
+        if (inputShape[1] != 1 && inputShape[1] != 3) {
+            Log.e(TAG, "Only one/three channels are supported in the image classification demo, you can use any " +
+                    "channel size in your Apps!");
+            return false;
+        }
+        if (!inputColorFormat.equalsIgnoreCase("BGR")) {
+            Log.e(TAG, "Only  BGR color format is supported.");
+            return false;
+        }
+        boolean isLoaded = init(appCtx, modelPath, labelPath);
+        if (!isLoaded) {
+            return false;
+        }
+        this.inputColorFormat = inputColorFormat;
+        this.inputShape = inputShape;
+        this.inputMean = inputMean;
+        this.inputStd = inputStd;
+        this.scoreThreshold = scoreThreshold;
+        return true;
+    }
+
+    protected boolean loadModel(Context appCtx, String modelPath, int cpuThreadNum, String cpuPowerMode) {
+        // Release model if exists
+        releaseModel();
+
+        // Load model
+        if (modelPath.isEmpty()) {
+            return false;
+        }
+        String realPath = modelPath;
+        if (!modelPath.substring(0, 1).equals("/")) {
+            // Read model files from custom path if the first character of mode path is '/'
+            // otherwise copy model to cache from assets
+            realPath = appCtx.getCacheDir() + "/" + modelPath;
+            Utils.copyDirectoryFromAssets(appCtx, modelPath, realPath);
+        }
+        if (realPath.isEmpty()) {
+            return false;
+        }
+
+        OCRPredictorNative.Config config = new OCRPredictorNative.Config();
+        config.cpuThreadNum = cpuThreadNum;
+        config.detModelFilename = realPath + File.separator + "ch_ppocr_mobile_v2.0_det_opt.nb";
+        config.recModelFilename = realPath + File.separator + "ch_ppocr_mobile_v2.0_rec_opt.nb";
+        config.clsModelFilename = realPath + File.separator + "ch_ppocr_mobile_v2.0_cls_opt.nb";
+        Log.e("Predictor", "model path" + config.detModelFilename + " ; " + config.recModelFilename + ";" + config.clsModelFilename);
+        config.cpuPower = cpuPowerMode;
+        paddlePredictor = new OCRPredictorNative(config);
+
+        this.cpuThreadNum = cpuThreadNum;
+        this.cpuPowerMode = cpuPowerMode;
+        this.modelPath = realPath;
+        this.modelName = realPath.substring(realPath.lastIndexOf("/") + 1);
+        return true;
+    }
+
+    public void releaseModel() {
+        if (paddlePredictor != null) {
+            paddlePredictor.destory();
+            paddlePredictor = null;
+        }
+        isLoaded = false;
+        cpuThreadNum = 1;
+        cpuPowerMode = "LITE_POWER_HIGH";
+        modelPath = "";
+        modelName = "";
+    }
+
+    protected boolean loadLabel(Context appCtx, String labelPath) {
+        wordLabels.clear();
+        wordLabels.add("black");
+        // Load word labels from file
+        try {
+            InputStream assetsInputStream = appCtx.getAssets().open(labelPath);
+            int available = assetsInputStream.available();
+            byte[] lines = new byte[available];
+            assetsInputStream.read(lines);
+            assetsInputStream.close();
+            String words = new String(lines);
+            String[] contents = words.split("\n");
+            for (String content : contents) {
+                wordLabels.add(content);
+            }
+            Log.i(TAG, "Word label size: " + wordLabels.size());
+        } catch (Exception e) {
+            Log.e(TAG, e.getMessage());
+            return false;
+        }
+        return true;
+    }
+
+
+    public boolean runModel() {
+        if (inputImage == null || !isLoaded()) {
+            return false;
+        }
+
+        // Pre-process image, and feed input tensor with pre-processed data
+
+        Bitmap scaleImage = Utils.resizeWithStep(inputImage, Long.valueOf(inputShape[2]).intValue(), 32);
+
+        Date start = new Date();
+        int channels = (int) inputShape[1];
+        int width = scaleImage.getWidth();
+        int height = scaleImage.getHeight();
+        float[] inputData = new float[channels * width * height];
+        if (channels == 3) {
+            int[] channelIdx = null;
+            if (inputColorFormat.equalsIgnoreCase("RGB")) {
+                channelIdx = new int[]{0, 1, 2};
+            } else if (inputColorFormat.equalsIgnoreCase("BGR")) {
+                channelIdx = new int[]{2, 1, 0};
+            } else {
+                Log.i(TAG, "Unknown color format " + inputColorFormat + ", only RGB and BGR color format is " +
+                        "supported!");
+                return false;
+            }
+            int[] channelStride = new int[]{width * height, width * height * 2};
+            int p = scaleImage.getPixel(scaleImage.getWidth() - 1, scaleImage.getHeight() - 1);
+            for (int y = 0; y < height; y++) {
+                for (int x = 0; x < width; x++) {
+                    int color = scaleImage.getPixel(x, y);
+                    float[] rgb = new float[]{(float) red(color) / 255.0f, (float) green(color) / 255.0f,
+                            (float) blue(color) / 255.0f};
+                    inputData[y * width + x] = (rgb[channelIdx[0]] - inputMean[0]) / inputStd[0];
+                    inputData[y * width + x + channelStride[0]] = (rgb[channelIdx[1]] - inputMean[1]) / inputStd[1];
+                    inputData[y * width + x + channelStride[1]] = (rgb[channelIdx[2]] - inputMean[2]) / inputStd[2];
+
+                }
+            }
+        } else if (channels == 1) {
+            for (int y = 0; y < height; y++) {
+                for (int x = 0; x < width; x++) {
+                    int color = inputImage.getPixel(x, y);
+                    float gray = (float) (red(color) + green(color) + blue(color)) / 3.0f / 255.0f;
+                    inputData[y * width + x] = (gray - inputMean[0]) / inputStd[0];
+                }
+            }
+        } else {
+            Log.i(TAG, "Unsupported channel size " + Integer.toString(channels) + ",  only channel 1 and 3 is " +
+                    "supported!");
+            return false;
+        }
+        float[] pixels = inputData;
+        Log.i(TAG, "pixels " + pixels[0] + " " + pixels[1] + " " + pixels[2] + " " + pixels[3]
+                + " " + pixels[pixels.length / 2] + " " + pixels[pixels.length / 2 + 1] + " " + pixels[pixels.length - 2] + " " + pixels[pixels.length - 1]);
+        Date end = new Date();
+        preprocessTime = (float) (end.getTime() - start.getTime());
+
+        // Warm up
+        for (int i = 0; i < warmupIterNum; i++) {
+            paddlePredictor.runImage(inputData, width, height, channels, inputImage);
+        }
+        warmupIterNum = 0; // do not need warm
+        // Run inference
+        start = new Date();
+        ArrayList<OcrResultModel> results = paddlePredictor.runImage(inputData, width, height, channels, inputImage);
+        end = new Date();
+        inferenceTime = (end.getTime() - start.getTime()) / (float) inferIterNum;
+
+        results = postprocess(results);
+        Log.i(TAG, "[stat] Preprocess Time: " + preprocessTime
+                + " ; Inference Time: " + inferenceTime + " ;Box Size " + results.size());
+        drawResults(results);
+
+        return true;
+    }
+
+
+    public boolean isLoaded() {
+        return paddlePredictor != null && isLoaded;
+    }
+
+    public String modelPath() {
+        return modelPath;
+    }
+
+    public String modelName() {
+        return modelName;
+    }
+
+    public int cpuThreadNum() {
+        return cpuThreadNum;
+    }
+
+    public String cpuPowerMode() {
+        return cpuPowerMode;
+    }
+
+    public float inferenceTime() {
+        return inferenceTime;
+    }
+
+    public Bitmap inputImage() {
+        return inputImage;
+    }
+
+    public Bitmap outputImage() {
+        return outputImage;
+    }
+
+    public String outputResult() {
+        return outputResult;
+    }
+
+    public float preprocessTime() {
+        return preprocessTime;
+    }
+
+    public float postprocessTime() {
+        return postprocessTime;
+    }
+
+
+    public void setInputImage(Bitmap image) {
+        if (image == null) {
+            return;
+        }
+        this.inputImage = image.copy(Bitmap.Config.ARGB_8888, true);
+    }
+
+    private ArrayList<OcrResultModel> postprocess(ArrayList<OcrResultModel> results) {
+        for (OcrResultModel r : results) {
+            StringBuffer word = new StringBuffer();
+            for (int index : r.getWordIndex()) {
+                if (index >= 0 && index < wordLabels.size()) {
+                    word.append(wordLabels.get(index));
+                } else {
+                    Log.e(TAG, "Word index is not in label list:" + index);
+                    word.append("×");
+                }
+            }
+            r.setLabel(word.toString());
+        }
+        return results;
+    }
+
+    private void drawResults(ArrayList<OcrResultModel> results) {
+        StringBuffer outputResultSb = new StringBuffer("");
+        for (int i = 0; i < results.size(); i++) {
+            OcrResultModel result = results.get(i);
+            StringBuilder sb = new StringBuilder("");
+            sb.append(result.getLabel());
+            sb.append(" ").append(result.getConfidence());
+            sb.append("; Points: ");
+            for (Point p : result.getPoints()) {
+                sb.append("(").append(p.x).append(",").append(p.y).append(") ");
+            }
+            Log.i(TAG, sb.toString()); // show LOG in Logcat panel
+            outputResultSb.append(i + 1).append(": ").append(result.getLabel()).append("\n");
+        }
+        outputResult = outputResultSb.toString();
+        outputImage = inputImage;
+        Canvas canvas = new Canvas(outputImage);
+        Paint paintFillAlpha = new Paint();
+        paintFillAlpha.setStyle(Paint.Style.FILL);
+        paintFillAlpha.setColor(Color.parseColor("#3B85F5"));
+        paintFillAlpha.setAlpha(50);
+
+        Paint paint = new Paint();
+        paint.setColor(Color.parseColor("#3B85F5"));
+        paint.setStrokeWidth(5);
+        paint.setStyle(Paint.Style.STROKE);
+
+        for (OcrResultModel result : results) {
+            Path path = new Path();
+            List<Point> points = result.getPoints();
+            path.moveTo(points.get(0).x, points.get(0).y);
+            for (int i = points.size() - 1; i >= 0; i--) {
+                Point p = points.get(i);
+                path.lineTo(p.x, p.y);
+            }
+            canvas.drawPath(path, paint);
+            canvas.drawPath(path, paintFillAlpha);
+        }
+    }
+
+}