Merge pull request #1 from PaddlePaddle/dygraph

Dygraph

deploy/cpp_infer/src/postprocess_op.cpp

@@ -159,6 +159,53 @@ std::vector<std::vector<float>> PostProcessor::GetMiniBoxes(cv::RotatedRect box,
   return array;
 }
 
+float PostProcessor::PolygonScoreAcc(std::vector<cv::Point> contour,
+                                     cv::Mat pred) {
+  int width = pred.cols;
+  int height = pred.rows;
+  std::vector<float> box_x;
+  std::vector<float> box_y;
+  for (int i = 0; i < contour.size(); ++i) {
+    box_x.push_back(contour[i].x);
+    box_y.push_back(contour[i].y);
+  }
+
+  int xmin =
+      clamp(int(std::floor(*(std::min_element(box_x.begin(), box_x.end())))), 0,
+            width - 1);
+  int xmax =
+      clamp(int(std::ceil(*(std::max_element(box_x.begin(), box_x.end())))), 0,
+            width - 1);
+  int ymin =
+      clamp(int(std::floor(*(std::min_element(box_y.begin(), box_y.end())))), 0,
+            height - 1);
+  int ymax =
+      clamp(int(std::ceil(*(std::max_element(box_y.begin(), box_y.end())))), 0,
+            height - 1);
+
+  cv::Mat mask;
+  mask = cv::Mat::zeros(ymax - ymin + 1, xmax - xmin + 1, CV_8UC1);
+
+
+  cv::Point* rook_point = new cv::Point[contour.size()];
+   
+  for (int i = 0; i < contour.size(); ++i) {
+    rook_point[i] = cv::Point(int(box_x[i]) - xmin, int(box_y[i]) - ymin);
+  }
+  const cv::Point *ppt[1] = {rook_point};
+  int npt[] = {int(contour.size())};
+
+
+  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);
+  float score = cv::mean(croppedImg, mask)[0];
+
+  delete []rook_point;
+  return score;
+}
+
 float PostProcessor::BoxScoreFast(std::vector<std::vector<float>> box_array,
                                   cv::Mat pred) {
   auto array = box_array;
@@ -197,10 +244,9 @@ float PostProcessor::BoxScoreFast(std::vector<std::vector<float>> box_array,
   return score;
 }
 
-std::vector<std::vector<std::vector<int>>>
-PostProcessor::BoxesFromBitmap(const cv::Mat pred, const cv::Mat bitmap,
-                               const float &box_thresh,
-                               const float &det_db_unclip_ratio) {
+std::vector<std::vector<std::vector<int>>> PostProcessor::BoxesFromBitmap(
+    const cv::Mat pred, const cv::Mat bitmap, const float &box_thresh,
+    const float &det_db_unclip_ratio, const bool &use_polygon_score) {
   const int min_size = 3;
   const int max_candidates = 1000;
 
@@ -234,7 +280,12 @@ PostProcessor::BoxesFromBitmap(const cv::Mat pred, const cv::Mat bitmap,
     }
 
     float score;
-    score = BoxScoreFast(array, pred);
+    if (use_polygon_score)
+      /* compute using polygon*/
+      score = PolygonScoreAcc(contours[_i], pred);
+    else
+      score = BoxScoreFast(array, pred);
+
     if (score < box_thresh)
       continue;
 

deploy/cpp_infer/src/preprocess_op.cpp

@@ -77,19 +77,10 @@ void ResizeImgType0::Run(const cv::Mat &img, cv::Mat &resize_img,
 
   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) * 32;
+  
+  resize_h = max(int(round(float(resize_h) / 32) * 32), 32);
+  resize_w = max(int(round(float(resize_w) / 32) * 32), 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) * 32;
   if (!use_tensorrt) {
     cv::resize(img, resize_img, cv::Size(resize_w, resize_h));
     ratio_h = float(resize_h) / float(h);

deploy/cpp_infer/tools/config.txt

@@ -10,6 +10,7 @@ max_side_len  960
 det_db_thresh  0.3
 det_db_box_thresh  0.5
 det_db_unclip_ratio  1.6
+use_polygon_score 1
 det_model_dir  ./inference/ch_ppocr_mobile_v2.0_det_infer/
 
 # cls config

deploy/hubserving/ocr_cls/module.py

@@ -6,6 +6,7 @@ from __future__ import print_function
 import os
 import sys
 sys.path.insert(0, ".")
+import copy
 
 from paddlehub.common.logger import logger
 from paddlehub.module.module import moduleinfo, runnable, serving
@@ -14,6 +15,8 @@ import paddlehub as hub
 
 from tools.infer.utility import base64_to_cv2
 from tools.infer.predict_cls import TextClassifier
+from tools.infer.utility import parse_args
+from deploy.hubserving.ocr_cls.params import read_params
 
 
 @moduleinfo(
@@ -28,8 +31,7 @@ class OCRCls(hub.Module):
         """
         initialize with the necessary elements
         """
-        from ocr_cls.params import read_params
-        cfg = read_params()
+        cfg = self.merge_configs()
 
         cfg.use_gpu = use_gpu
         if use_gpu:
@@ -48,6 +50,20 @@ class OCRCls(hub.Module):
 
         self.text_classifier = TextClassifier(cfg)
 
+    def merge_configs(self, ):
+        # deafult cfg
+        backup_argv = copy.deepcopy(sys.argv)
+        sys.argv = sys.argv[:1]
+        cfg = parse_args()
+
+        update_cfg_map = vars(read_params())
+
+        for key in update_cfg_map:
+            cfg.__setattr__(key, update_cfg_map[key])
+
+        sys.argv = copy.deepcopy(backup_argv)
+        return cfg
+
     def read_images(self, paths=[]):
         images = []
         for img_path in paths:

deploy/hubserving/ocr_det/module.py

@@ -7,6 +7,8 @@ import os
 import sys
 sys.path.insert(0, ".")
 
+import copy
+
 from paddlehub.common.logger import logger
 from paddlehub.module.module import moduleinfo, runnable, serving
 import cv2
@@ -15,6 +17,8 @@ import paddlehub as hub
 
 from tools.infer.utility import base64_to_cv2
 from tools.infer.predict_det import TextDetector
+from tools.infer.utility import parse_args
+from deploy.hubserving.ocr_system.params import read_params
 
 
 @moduleinfo(
@@ -29,8 +33,7 @@ class OCRDet(hub.Module):
         """
         initialize with the necessary elements
         """
-        from ocr_det.params import read_params
-        cfg = read_params()
+        cfg = self.merge_configs()
 
         cfg.use_gpu = use_gpu
         if use_gpu:
@@ -49,6 +52,20 @@ class OCRDet(hub.Module):
 
         self.text_detector = TextDetector(cfg)
 
+    def merge_configs(self, ):
+        # deafult cfg
+        backup_argv = copy.deepcopy(sys.argv)
+        sys.argv = sys.argv[:1]
+        cfg = parse_args()
+
+        update_cfg_map = vars(read_params())
+
+        for key in update_cfg_map:
+            cfg.__setattr__(key, update_cfg_map[key])
+
+        sys.argv = copy.deepcopy(backup_argv)
+        return cfg
+
     def read_images(self, paths=[]):
         images = []
         for img_path in paths:

deploy/hubserving/ocr_det/params.py

@@ -22,6 +22,7 @@ def read_params():
     cfg.det_db_box_thresh = 0.5
     cfg.det_db_unclip_ratio = 1.6
     cfg.use_dilation = False
+    cfg.det_db_score_mode = "fast"
 
     # #EAST parmas
     # cfg.det_east_score_thresh = 0.8

deploy/hubserving/ocr_rec/module.py

@@ -6,6 +6,7 @@ from __future__ import print_function
 import os
 import sys
 sys.path.insert(0, ".")
+import copy
 
 from paddlehub.common.logger import logger
 from paddlehub.module.module import moduleinfo, runnable, serving
@@ -14,6 +15,8 @@ import paddlehub as hub
 
 from tools.infer.utility import base64_to_cv2
 from tools.infer.predict_rec import TextRecognizer
+from tools.infer.utility import parse_args
+from deploy.hubserving.ocr_rec.params import read_params
 
 
 @moduleinfo(
@@ -28,8 +31,7 @@ class OCRRec(hub.Module):
         """
         initialize with the necessary elements
         """
-        from ocr_rec.params import read_params
-        cfg = read_params()
+        cfg = self.merge_configs()
 
         cfg.use_gpu = use_gpu
         if use_gpu:
@@ -48,6 +50,20 @@ class OCRRec(hub.Module):
 
         self.text_recognizer = TextRecognizer(cfg)
 
+    def merge_configs(self, ):
+        # deafult cfg
+        backup_argv = copy.deepcopy(sys.argv)
+        sys.argv = sys.argv[:1]
+        cfg = parse_args()
+
+        update_cfg_map = vars(read_params())
+
+        for key in update_cfg_map:
+            cfg.__setattr__(key, update_cfg_map[key])
+
+        sys.argv = copy.deepcopy(backup_argv)
+        return cfg
+
     def read_images(self, paths=[]):
         images = []
         for img_path in paths:

deploy/hubserving/ocr_system/module.py

@@ -6,6 +6,7 @@ from __future__ import print_function
 import os
 import sys
 sys.path.insert(0, ".")
+import copy
 
 import time
 
@@ -17,6 +18,8 @@ import paddlehub as hub
 
 from tools.infer.utility import base64_to_cv2
 from tools.infer.predict_system import TextSystem
+from tools.infer.utility import parse_args
+from deploy.hubserving.ocr_system.params import read_params
 
 
 @moduleinfo(
@@ -31,8 +34,7 @@ class OCRSystem(hub.Module):
         """
         initialize with the necessary elements
         """
-        from ocr_system.params import read_params
-        cfg = read_params()
+        cfg = self.merge_configs()
 
         cfg.use_gpu = use_gpu
         if use_gpu:
@@ -51,6 +53,20 @@ class OCRSystem(hub.Module):
 
         self.text_sys = TextSystem(cfg)
 
+    def merge_configs(self, ):
+        # deafult cfg
+        backup_argv = copy.deepcopy(sys.argv)
+        sys.argv = sys.argv[:1]
+        cfg = parse_args()
+
+        update_cfg_map = vars(read_params())
+
+        for key in update_cfg_map:
+            cfg.__setattr__(key, update_cfg_map[key])
+
+        sys.argv = copy.deepcopy(backup_argv)
+        return cfg
+
     def read_images(self, paths=[]):
         images = []
         for img_path in paths:

deploy/hubserving/ocr_system/params.py

@@ -22,6 +22,7 @@ def read_params():
     cfg.det_db_box_thresh = 0.5
     cfg.det_db_unclip_ratio = 1.6
     cfg.use_dilation = False
+    cfg.det_db_score_mode = "fast"
 
     #EAST parmas
     cfg.det_east_score_thresh = 0.8

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 cls_process.o
+	$(CC) $(SYSROOT_LINK) $(CXXFLAGS_LINK) ocr_db_crnn.o crnn_process.o db_post_process.o clipper.o cls_process.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
+
+cls_process.o: fetch_opencv cls_process.cc
+	$(CC) $(SYSROOT_COMPLILE) $(CXX_DEFINES) $(CXX_INCLUDES) $(CXX_FLAGS) -o cls_process.o -c cls_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 cls_process.o
+	rm -f ocr_db_crnn

deploy/lite/cls_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 "cls_process.h" //NOLINT
+#include <algorithm>
+#include <memory>
+#include <string>
+
+const std::vector<int> rec_image_shape{3, 48, 192};
+
+cv::Mat ClsResizeImg(cv::Mat img) {
+  int imgC, imgH, imgW;
+  imgC = rec_image_shape[0];
+  imgH = rec_image_shape[1];
+  imgW = rec_image_shape[2];
+
+  float ratio = static_cast<float>(img.cols) / static_cast<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);
+  if (resize_w < imgW) {
+    cv::copyMakeBorder(resize_img, resize_img, 0, 0, 0, imgW - resize_w,
+                       cv::BORDER_CONSTANT, cv::Scalar(0, 0, 0));
+  }
+  return resize_img;
+}
\ No newline at end of file

deploy/lite/cls_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 ClsResizeImg(cv::Mat img);
\ No newline at end of file

deploy/lite/config.txt

+max_side_len  960
+det_db_thresh  0.3
+det_db_box_thresh  0.5
+det_db_unclip_ratio  1.6
+det_db_use_dilate 0
+det_use_polygon_score 1
+use_direction_classify  1
\ 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 = static_cast<float>(img.cols) / static_cast<float>(img.rows);
+  int resize_w, resize_h;
+  if (ceilf(imgH * ratio) > imgW)
+    resize_w = imgW;
+  else
+    resize_w = static_cast<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 =
+      static_cast<int>(sqrt(pow(points[0][0] - points[1][0], 2) +
+                            pow(points[0][1] - points[1][1], 2)));
+  int img_crop_height =
+      static_cast<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);
+
+  const float ratio = 1.5;
+  if (static_cast<float>(dst_img.rows) >=
+      static_cast<float>(dst_img.cols) * ratio) {
+    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(static_cast<int>(box[0][0]),
+                            static_cast<int>(box[0][1]))
+    << ClipperLib::IntPoint(static_cast<int>(box[1][0]),
+                            static_cast<int>(box[1][1]))
+    << ClipperLib::IntPoint(static_cast<int>(box[2][0]),
+                            static_cast<int>(box[2][1]))
+    << ClipperLib::IntPoint(static_cast<int>(box[3][0]),
+                            static_cast<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::min(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(
+      static_cast<int>(std::floorf(*(std::min_element(box_x, box_x + 4)))), 0,
+      width - 1);
+  int xmax =
+      clamp(static_cast<int>(std::ceilf(*(std::max_element(box_x, box_x + 4)))),
+            0, width - 1);
+  int ymin = clamp(
+      static_cast<int>(std::floorf(*(std::min_element(box_y, box_y + 4)))), 0,
+      height - 1);
+  int ymax =
+      clamp(static_cast<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(static_cast<int>(array[0][0]) - xmin,
+                            static_cast<int>(array[0][1]) - ymin);
+  root_point[1] = cv::Point(static_cast<int>(array[1][0]) - xmin,
+                            static_cast<int>(array[1][1]) - ymin);
+  root_point[2] = cv::Point(static_cast<int>(array[2][0]) - xmin,
+                            static_cast<int>(array[2][1]) - ymin);
+  root_point[3] = cv::Point(static_cast<int>(array[3][0]) - xmin,
+                            static_cast<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;
+}
+
+float PolygonScoreAcc(std::vector<cv::Point> contour, cv::Mat pred) {
+  int width = pred.cols;
+  int height = pred.rows;
+  std::vector<float> box_x;
+  std::vector<float> box_y;
+  for (int i = 0; i < contour.size(); ++i) {
+    box_x.push_back(contour[i].x);
+    box_y.push_back(contour[i].y);
+  }
+
+  int xmin =
+      clamp(int(std::floor(*(std::min_element(box_x.begin(), box_x.end())))), 0,
+            width - 1);
+  int xmax =
+      clamp(int(std::ceil(*(std::max_element(box_x.begin(), box_x.end())))), 0,
+            width - 1);
+  int ymin =
+      clamp(int(std::floor(*(std::min_element(box_y.begin(), box_y.end())))), 0,
+            height - 1);
+  int ymax =
+      clamp(int(std::ceil(*(std::max_element(box_y.begin(), box_y.end())))), 0,
+            height - 1);
+
+  cv::Mat mask;
+  mask = cv::Mat::zeros(ymax - ymin + 1, xmax - xmin + 1, CV_8UC1);
+
+  cv::Point *rook_point = new cv::Point[contour.size()];
+
+  for (int i = 0; i < contour.size(); ++i) {
+    rook_point[i] = cv::Point(int(box_x[i]) - xmin, int(box_y[i]) - ymin);
+  }
+  const cv::Point *ppt[1] = {rook_point};
+  int npt[] = {int(contour.size())};
+
+  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);
+  float score = cv::mean(croppedImg, mask)[0];
+
+  delete[] rook_point;
+  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 = static_cast<float>(Config["det_db_box_thresh"]);
+  const float unclip_ratio = static_cast<float>(Config["det_db_unclip_ratio"]);
+  const int det_use_polygon_score = int(Config["det_use_polygon_score"]);
+
+  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;
+    if (contours[i].size() <= 2)
+      continue;
+
+    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;
+    if (det_use_polygon_score) {
+      score = PolygonScoreAcc(contours[i], pred);
+    } else {
+      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);
+    if (points.size.height < 1.001 && points.size.width < 1.001)
+      continue;
+    // 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{
+          static_cast<int>(clamp(
+              roundf(cliparray[num_pt][0] / float(width) * float(dest_width)),
+              float(0), float(dest_width))),
+          static_cast<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 < static_cast<int>(boxes.size()); n++) {
+    boxes[n] = OrderPointsClockwise(boxes[n]);
+    for (int m = 0; m < static_cast<int>(boxes[0].size()); m++) {
+      boxes[n][m][0] /= ratio_w;
+      boxes[n][m][1] /= ratio_h;
+
+      boxes[n][m][0] =
+          static_cast<int>(std::min(std::max(boxes[n][m][0], 0), oriimg_w - 1));
+      boxes[n][m][1] =
+          static_cast<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 =
+        static_cast<int>(sqrt(pow(boxes[n][0][0] - boxes[n][1][0], 2) +
+                              pow(boxes[n][0][1] - boxes[n][1][1], 2)));
+    rect_height =
+        static_cast<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 <= 4 || rect_height <= 4)
+      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 "cls_process.h"
+#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 NeonMeanScale(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 = static_cast<float>(max_size_len) / static_cast<float>(h);
+    } else {
+      ratio = static_cast<float>(max_size_len) / static_cast<float>(w);
+    }
+  }
+
+  int resize_h = static_cast<int>(float(h) * ratio);
+  int resize_w = static_cast<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(static_cast<float>(resize_h) / static_cast<float>(h));
+  ratio_hw.push_back(static_cast<float>(resize_w) / static_cast<float>(w));
+  return resize_img;
+}
+
+cv::Mat RunClsModel(cv::Mat img, std::shared_ptr<PaddlePredictor> predictor_cls,
+                    const float thresh = 0.9) {
+  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;
+  img.copyTo(crop_img);
+  cv::Mat resize_img;
+
+  int index = 0;
+  float wh_ratio =
+      static_cast<float>(crop_img.cols) / static_cast<float>(crop_img.rows);
+
+  resize_img = ClsResizeImg(crop_img);
+  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_cls->GetInput(0)));
+  input_tensor0->Resize({1, 3, resize_img.rows, resize_img.cols});
+  auto *data0 = input_tensor0->mutable_data<float>();
+
+  NeonMeanScale(dimg, data0, resize_img.rows * resize_img.cols, mean, scale);
+  // Run CLS predictor
+  predictor_cls->Run();
+
+  // Get output and run postprocess
+  std::unique_ptr<const Tensor> softmax_out(
+      std::move(predictor_cls->GetOutput(0)));
+  auto *softmax_scores = softmax_out->mutable_data<float>();
+  auto softmax_out_shape = softmax_out->shape();
+  float score = 0;
+  int label = 0;
+  for (int i = 0; i < softmax_out_shape[1]; i++) {
+    if (softmax_scores[i] > score) {
+      score = softmax_scores[i];
+      label = i;
+    }
+  }
+  if (label % 2 == 1 && score > thresh) {
+    cv::rotate(srcimg, srcimg, 1);
+  }
+  return srcimg;
+}
+
+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::shared_ptr<PaddlePredictor> predictor_cls,
+                 int use_direction_classify) {
+  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]);
+    if (use_direction_classify >= 1) {
+      crop_img = RunClsModel(crop_img, predictor_cls);
+    }
+    float wh_ratio =
+        static_cast<float>(crop_img.cols) / static_cast<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>();
+
+    NeonMeanScale(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 *predict_batch = output_tensor0->data<float>();
+    auto predict_shape = output_tensor0->shape();
+
+    // ctc decode
+    std::string str_res;
+    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;
+        str_res += charactor_dict[argmax_idx];
+      }
+      last_index = argmax_idx;
+    }
+    score /= count;
+    rec_text.push_back(str_res);
+    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"]);
+  int det_db_use_dilate = int(Config["det_db_use_dilate"]);
+
+  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);
+  NeonMeanScale(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] = static_cast<float>(outptr[i]);
+    cbuf[i] = static_cast<unsigned char>((outptr[i]) * 255);
+  }
+
+  cv::Mat cbuf_map(shape_out[2], shape_out[3], CV_8UC1,
+                   reinterpret_cast<unsigned char *>(cbuf));
+  cv::Mat pred_map(shape_out[2], shape_out[3], CV_32F,
+                   reinterpret_cast<float *>(pred));
+
+  const double threshold = double(Config["det_db_thresh"]) * 255;
+  const double max_value = 255;
+  cv::Mat bit_map;
+  cv::threshold(cbuf_map, bit_map, threshold, max_value, cv::THRESH_BINARY);
+  if (det_db_use_dilate == 1) {
+    cv::Mat dilation_map;
+    cv::Mat dila_ele =
+        cv::getStructuringElement(cv::MORPH_RECT, cv::Size(2, 2));
+    cv::dilate(bit_map, dilation_map, dila_ele);
+    bit_map = dilation_map;
+  }
+  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(static_cast<int>(boxes[n][m][0]),
+                                    static_cast<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 cls_model_file rec_model_file image_path "
+                 "charactor_dict\n";
+    exit(1);
+  }
+  std::string det_model_file = argv[1];
+  std::string rec_model_file = argv[2];
+  std::string cls_model_file = argv[3];
+  std::string img_path = argv[4];
+  std::string dict_path = argv[5];
+
+  //// load config from txt file
+  auto Config = LoadConfigTxt("./config.txt");
+  int use_direction_classify = int(Config["use_direction_classify"]);
+
+  auto start = std::chrono::system_clock::now();
+
+  auto det_predictor = loadModel(det_model_file);
+  auto rec_predictor = loadModel(rec_model_file);
+  auto cls_predictor = loadModel(cls_model_file);
+
+  auto charactor_dict = ReadDict(dict_path);
+  charactor_dict.insert(charactor_dict.begin(), "#"); // blank char for ctc
+  charactor_dict.push_back(" ");
+
+  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, cls_predictor, use_direction_classify);
+
+  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;
+}
\ No newline at end of file

deploy/lite/prepare.sh

+#!/bin/bash
+
+mkdir -p  $1/demo/cxx/ocr/debug/
+cp  ../../ppocr/utils/ppocr_keys_v1.txt  $1/demo/cxx/ocr/debug/
+cp -r  ./*   $1/demo/cxx/ocr/
+cp ./config.txt  $1/demo/cxx/ocr/debug/
+cp ../../doc/imgs/11.jpg  $1/demo/cxx/ocr/debug/
+
+echo "Prepare Done"