first commit

deploy/cpp_infer/include/args.h

@@ -0,0 +1,68 @@
+// 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 <gflags/gflags.h>
+
+// common args
+DECLARE_bool(use_gpu);
+DECLARE_bool(use_tensorrt);
+DECLARE_int32(gpu_id);
+DECLARE_int32(gpu_mem);
+DECLARE_int32(cpu_threads);
+DECLARE_bool(enable_mkldnn);
+DECLARE_string(precision);
+DECLARE_bool(benchmark);
+DECLARE_string(output);
+DECLARE_string(image_dir);
+DECLARE_string(type);
+// detection related
+DECLARE_string(det_model_dir);
+DECLARE_string(limit_type);
+DECLARE_int32(limit_side_len);
+DECLARE_double(det_db_thresh);
+DECLARE_double(det_db_box_thresh);
+DECLARE_double(det_db_unclip_ratio);
+DECLARE_bool(use_dilation);
+DECLARE_string(det_db_score_mode);
+DECLARE_bool(visualize);
+// classification related
+DECLARE_bool(use_angle_cls);
+DECLARE_string(cls_model_dir);
+DECLARE_double(cls_thresh);
+DECLARE_int32(cls_batch_num);
+// recognition related
+DECLARE_string(rec_model_dir);
+DECLARE_int32(rec_batch_num);
+DECLARE_string(rec_char_dict_path);
+DECLARE_int32(rec_img_h);
+DECLARE_int32(rec_img_w);
+// layout model related
+DECLARE_string(layout_model_dir);
+DECLARE_string(layout_dict_path);
+DECLARE_double(layout_score_threshold);
+DECLARE_double(layout_nms_threshold);
+// structure model related
+DECLARE_string(table_model_dir);
+DECLARE_int32(table_max_len);
+DECLARE_int32(table_batch_num);
+DECLARE_string(table_char_dict_path);
+DECLARE_bool(merge_no_span_structure);
+// forward related
+DECLARE_bool(det);
+DECLARE_bool(rec);
+DECLARE_bool(cls);
+DECLARE_bool(table);
+DECLARE_bool(layout);

deploy/cpp_infer/include/clipper.h

@@ -0,0 +1,435 @@
+/*******************************************************************************
+ *                                                                              *
+ * Author    :  Angus Johnson * Version   :  6.4.2 * Date      :  27 February
+ *2017                                                * Website   :
+ *http://www.angusj.com                                           * Copyright :
+ *Angus Johnson 2010-2017                                         *
+ *                                                                              *
+ * License: * Use, modification & distribution is subject to Boost Software
+ *License Ver 1. * http://www.boost.org/LICENSE_1_0.txt *
+ *                                                                              *
+ * Attributions: * The code in this library is an extension of Bala Vatti's
+ *clipping algorithm: * "A generic solution to polygon clipping" *
+ * Communications of the ACM, Vol 35, Issue 7 (July 1992) pp 56-63. *
+ * http://portal.acm.org/citation.cfm?id=129906 *
+ *                                                                              *
+ * Computer graphics and geometric modeling: implementation and algorithms * By
+ *Max K. Agoston                                                            *
+ * Springer; 1 edition (January 4, 2005) *
+ * http://books.google.com/books?q=vatti+clipping+agoston *
+ *                                                                              *
+ * See also: * "Polygon Offsetting by Computing Winding Numbers" * Paper no.
+ *DETC2005-85513 pp. 565-575                                         * ASME 2005
+ *International Design Engineering Technical Conferences             * and
+ *Computers and Information in Engineering Conference (IDETC/CIE2005)      *
+ * September 24-28, 2005 , Long Beach, California, USA *
+ * http://www.me.berkeley.edu/~mcmains/pubs/DAC05OffsetPolygon.pdf *
+ *                                                                              *
+ *******************************************************************************/
+
+#pragma once
+
+#ifndef clipper_hpp
+#define clipper_hpp
+
+#define CLIPPER_VERSION "6.4.2"
+
+// use_int32: When enabled 32bit ints are used instead of 64bit ints. This
+// improve performance but coordinate values are limited to the range +/- 46340
+//#define use_int32
+
+// use_xyz: adds a Z member to IntPoint. Adds a minor cost to performance.
+//#define use_xyz
+
+// use_lines: Enables line clipping. Adds a very minor cost to performance.
+#define use_lines
+
+// use_deprecated: Enables temporary support for the obsolete functions
+//#define use_deprecated
+
+#include <list>
+#include <queue>
+#include <string>
+#include <vector>
+
+namespace ClipperLib {
+
+enum ClipType { ctIntersection, ctUnion, ctDifference, ctXor };
+enum PolyType { ptSubject, ptClip };
+// By far the most widely used winding rules for polygon filling are
+// EvenOdd & NonZero (GDI, GDI+, XLib, OpenGL, Cairo, AGG, Quartz, SVG, Gr32)
+// Others rules include Positive, Negative and ABS_GTR_EQ_TWO (only in OpenGL)
+// see http://glprogramming.com/red/chapter11.html
+enum PolyFillType { pftEvenOdd, pftNonZero, pftPositive, pftNegative };
+
+#ifdef use_int32
+typedef int cInt;
+static cInt const loRange = 0x7FFF;
+static cInt const hiRange = 0x7FFF;
+#else
+typedef signed long long cInt;
+static cInt const loRange = 0x3FFFFFFF;
+static cInt const hiRange = 0x3FFFFFFFFFFFFFFFLL;
+typedef signed long long long64; // used by Int128 class
+typedef unsigned long long ulong64;
+
+#endif
+
+struct IntPoint {
+  cInt X;
+  cInt Y;
+#ifdef use_xyz
+  cInt Z;
+  IntPoint(cInt x = 0, cInt y = 0, cInt z = 0) noexcept : X(x), Y(y), Z(z) {}
+  IntPoint(IntPoint const &ip) noexcept : X(ip.X), Y(ip.Y), Z(ip.Z) {}
+#else
+  IntPoint(cInt x = 0, cInt y = 0) noexcept : X(x), Y(y) {}
+  IntPoint(IntPoint const &ip) noexcept : X(ip.X), Y(ip.Y) {}
+#endif
+
+  inline void reset(cInt x = 0, cInt y = 0) noexcept {
+    X = x;
+    Y = y;
+  }
+
+  friend inline bool operator==(const IntPoint &a, const IntPoint &b) noexcept {
+    return a.X == b.X && a.Y == b.Y;
+  }
+  friend inline bool operator!=(const IntPoint &a, const IntPoint &b) noexcept {
+    return a.X != b.X || a.Y != b.Y;
+  }
+};
+//------------------------------------------------------------------------------
+
+typedef std::vector<IntPoint> Path;
+typedef std::vector<Path> Paths;
+
+inline Path &operator<<(Path &poly, IntPoint &&p) noexcept {
+  poly.emplace_back(std::forward<IntPoint>(p));
+  return poly;
+}
+inline Paths &operator<<(Paths &polys, Path &&p) noexcept {
+  polys.emplace_back(std::forward<Path>(p));
+  return polys;
+}
+
+std::ostream &operator<<(std::ostream &s, const IntPoint &p) noexcept;
+std::ostream &operator<<(std::ostream &s, const Path &p) noexcept;
+std::ostream &operator<<(std::ostream &s, const Paths &p) noexcept;
+
+struct DoublePoint {
+  double X;
+  double Y;
+  DoublePoint(double x = 0, double y = 0) noexcept : X(x), Y(y) {}
+  DoublePoint(IntPoint const &ip) noexcept : X((double)ip.X), Y((double)ip.Y) {}
+  inline void reset(double x = 0, double y = 0) noexcept {
+    X = x;
+    Y = y;
+  }
+};
+//------------------------------------------------------------------------------
+
+#ifdef use_xyz
+typedef void (*ZFillCallback)(IntPoint &e1bot, IntPoint &e1top, IntPoint &e2bot,
+                              IntPoint &e2top, IntPoint &pt);
+#endif
+
+enum InitOptions {
+  ioReverseSolution = 1,
+  ioStrictlySimple = 2,
+  ioPreserveCollinear = 4
+};
+enum JoinType { jtSquare, jtRound, jtMiter };
+enum EndType {
+  etClosedPolygon,
+  etClosedLine,
+  etOpenButt,
+  etOpenSquare,
+  etOpenRound
+};
+
+class PolyNode;
+typedef std::vector<PolyNode *> PolyNodes;
+
+class PolyNode {
+public:
+  PolyNode() noexcept;
+  virtual ~PolyNode() {}
+  Path Contour;
+  PolyNodes Children;
+  PolyNode *Parent;
+  PolyNode *GetNext() const noexcept;
+  bool IsHole() const noexcept;
+  bool IsOpen() const noexcept;
+  int ChildCount() const noexcept;
+
+private:
+  // PolyNode& operator =(PolyNode& other);
+  unsigned Index; // node index in Parent.Children
+  bool m_IsOpen;
+  JoinType m_jointype;
+  EndType m_endtype;
+  PolyNode *GetNextSiblingUp() const noexcept;
+  void AddChild(PolyNode &child) noexcept;
+  friend class Clipper; // to access Index
+  friend class ClipperOffset;
+};
+
+class PolyTree : public PolyNode {
+public:
+  ~PolyTree() { Clear(); }
+  PolyNode *GetFirst() const noexcept;
+  void Clear() noexcept;
+  int Total() const noexcept;
+
+private:
+  // PolyTree& operator =(PolyTree& other);
+  PolyNodes AllNodes;
+  friend class Clipper; // to access AllNodes
+};
+
+bool Orientation(const Path &poly) noexcept;
+double Area(const Path &poly) noexcept;
+int PointInPolygon(const IntPoint &pt, const Path &path) noexcept;
+
+#if 0
+void SimplifyPolygon(const Path &in_poly, Paths &out_polys,
+                     PolyFillType fillType = pftEvenOdd);
+void SimplifyPolygons(const Paths &in_polys, Paths &out_polys,
+                      PolyFillType fillType = pftEvenOdd);
+void SimplifyPolygons(Paths &polys, PolyFillType fillType = pftEvenOdd);
+#endif
+
+void CleanPolygon(const Path &in_poly, Path &out_poly,
+                  double distance = 1.415) noexcept;
+void CleanPolygon(Path &poly, double distance = 1.415) noexcept;
+void CleanPolygons(const Paths &in_polys, Paths &out_polys,
+                   double distance = 1.415) noexcept;
+void CleanPolygons(Paths &polys, double distance = 1.415) noexcept;
+
+#if 0
+void MinkowskiSum(const Path &pattern, const Path &path, Paths &solution,
+                  bool pathIsClosed);
+void MinkowskiSum(const Path &pattern, const Paths &paths, Paths &solution,
+                  bool pathIsClosed);
+
+void MinkowskiDiff(const Path &poly1, const Path &poly2, Paths &solution);
+#endif
+
+void PolyTreeToPaths(const PolyTree &polytree, Paths &paths) noexcept;
+void ClosedPathsFromPolyTree(const PolyTree &polytree, Paths &paths) noexcept;
+void OpenPathsFromPolyTree(PolyTree &polytree, Paths &paths) noexcept;
+
+void ReversePath(Path &p) noexcept;
+void ReversePaths(Paths &p) noexcept;
+
+struct IntRect {
+  cInt left;
+  cInt top;
+  cInt right;
+  cInt bottom;
+};
+
+// enums that are used internally ...
+enum EdgeSide { esLeft = 1, esRight = 2 };
+
+// forward declarations (for stuff used internally) ...
+struct TEdge;
+struct IntersectNode;
+struct LocalMinimum;
+struct OutPt;
+struct OutRec;
+struct Join;
+
+typedef std::vector<OutRec *> PolyOutList;
+typedef std::vector<TEdge *> EdgeList;
+typedef std::vector<Join *> JoinList;
+typedef std::vector<IntersectNode *> IntersectList;
+
+//------------------------------------------------------------------------------
+
+// ClipperBase is the ancestor to the Clipper class. It should not be
+// instantiated directly. This class simply abstracts the conversion of sets of
+// polygon coordinates into edge objects that are stored in a LocalMinima list.
+class ClipperBase {
+public:
+  ClipperBase() noexcept;
+  virtual ~ClipperBase();
+  virtual bool AddPath(const Path &pg, PolyType PolyTyp, bool Closed);
+  bool AddPaths(const Paths &ppg, PolyType PolyTyp, bool Closed);
+  virtual void Clear() noexcept;
+  IntRect GetBounds() noexcept;
+  bool PreserveCollinear() const noexcept { return m_PreserveCollinear; }
+  void PreserveCollinear(bool value) noexcept { m_PreserveCollinear = value; }
+
+protected:
+  void DisposeLocalMinimaList() noexcept;
+  TEdge *AddBoundsToLML(TEdge *e, bool IsClosed) noexcept;
+  virtual void Reset() noexcept;
+  TEdge *ProcessBound(TEdge *E, bool IsClockwise) noexcept;
+  void InsertScanbeam(const cInt Y) noexcept;
+  bool PopScanbeam(cInt &Y) noexcept;
+  bool LocalMinimaPending() noexcept;
+  bool PopLocalMinima(cInt Y, const LocalMinimum *&locMin) noexcept;
+  OutRec *CreateOutRec() noexcept;
+  void DisposeAllOutRecs() noexcept;
+  void DisposeOutRec(PolyOutList::size_type index) noexcept;
+  void SwapPositionsInAEL(TEdge *edge1, TEdge *edge2) noexcept;
+  void DeleteFromAEL(TEdge *e) noexcept;
+  void UpdateEdgeIntoAEL(TEdge *&e);
+
+  typedef std::vector<LocalMinimum> MinimaList;
+  MinimaList::iterator m_CurrentLM;
+  MinimaList m_MinimaList;
+
+  bool m_UseFullRange;
+  EdgeList m_edges;
+  bool m_PreserveCollinear;
+  bool m_HasOpenPaths;
+  PolyOutList m_PolyOuts;
+  TEdge *m_ActiveEdges;
+
+  typedef std::priority_queue<cInt> ScanbeamList;
+  ScanbeamList m_Scanbeam;
+};
+//------------------------------------------------------------------------------
+
+class Clipper : public virtual ClipperBase {
+public:
+  Clipper(int initOptions = 0) noexcept;
+  bool Execute(ClipType clipType, Paths &solution,
+               PolyFillType fillType = pftEvenOdd);
+  bool Execute(ClipType clipType, Paths &solution, PolyFillType subjFillType,
+               PolyFillType clipFillType);
+  bool Execute(ClipType clipType, PolyTree &polytree,
+               PolyFillType fillType = pftEvenOdd) noexcept;
+  bool Execute(ClipType clipType, PolyTree &polytree, PolyFillType subjFillType,
+               PolyFillType clipFillType) noexcept;
+  bool ReverseSolution() const noexcept { return m_ReverseOutput; }
+  void ReverseSolution(bool value) noexcept { m_ReverseOutput = value; }
+  bool StrictlySimple() const noexcept { return m_StrictSimple; }
+  void StrictlySimple(bool value) noexcept { m_StrictSimple = value; }
+// set the callback function for z value filling on intersections (otherwise Z
+// is 0)
+#ifdef use_xyz
+  void ZFillFunction(ZFillCallback zFillFunc) noexcept;
+#endif
+protected:
+  virtual bool ExecuteInternal() noexcept;
+
+private:
+  JoinList m_Joins;
+  JoinList m_GhostJoins;
+  IntersectList m_IntersectList;
+  ClipType m_ClipType;
+  typedef std::list<cInt> MaximaList;
+  MaximaList m_Maxima;
+  TEdge *m_SortedEdges;
+  bool m_ExecuteLocked;
+  PolyFillType m_ClipFillType;
+  PolyFillType m_SubjFillType;
+  bool m_ReverseOutput;
+  bool m_UsingPolyTree;
+  bool m_StrictSimple;
+#ifdef use_xyz
+  ZFillCallback m_ZFill; // custom callback
+#endif
+  void SetWindingCount(TEdge &edge) noexcept;
+  bool IsEvenOddFillType(const TEdge &edge) const noexcept;
+  bool IsEvenOddAltFillType(const TEdge &edge) const noexcept;
+  void InsertLocalMinimaIntoAEL(const cInt botY) noexcept;
+  void InsertEdgeIntoAEL(TEdge *edge, TEdge *startEdge) noexcept;
+  void AddEdgeToSEL(TEdge *edge) noexcept;
+  bool PopEdgeFromSEL(TEdge *&edge) noexcept;
+  void CopyAELToSEL() noexcept;
+  void DeleteFromSEL(TEdge *e) noexcept;
+  void SwapPositionsInSEL(TEdge *edge1, TEdge *edge2) noexcept;
+  bool IsContributing(const TEdge &edge) const noexcept;
+  bool IsTopHorz(const cInt XPos) noexcept;
+  void DoMaxima(TEdge *e);
+  void ProcessHorizontals() noexcept;
+  void ProcessHorizontal(TEdge *horzEdge) noexcept;
+  void AddLocalMaxPoly(TEdge *e1, TEdge *e2, const IntPoint &pt) noexcept;
+  OutPt *AddLocalMinPoly(TEdge *e1, TEdge *e2, const IntPoint &pt) noexcept;
+  OutRec *GetOutRec(int idx) noexcept;
+  void AppendPolygon(TEdge *e1, TEdge *e2) noexcept;
+  void IntersectEdges(TEdge *e1, TEdge *e2, IntPoint &pt) noexcept;
+  OutPt *AddOutPt(TEdge *e, const IntPoint &pt) noexcept;
+  OutPt *GetLastOutPt(TEdge *e) noexcept;
+  bool ProcessIntersections(const cInt topY);
+  void BuildIntersectList(const cInt topY) noexcept;
+  void ProcessIntersectList() noexcept;
+  void ProcessEdgesAtTopOfScanbeam(const cInt topY);
+  void BuildResult(Paths &polys) noexcept;
+  void BuildResult2(PolyTree &polytree) noexcept;
+  void SetHoleState(TEdge *e, OutRec *outrec) noexcept;
+  void DisposeIntersectNodes() noexcept;
+  bool FixupIntersectionOrder() noexcept;
+  void FixupOutPolygon(OutRec &outrec) noexcept;
+  void FixupOutPolyline(OutRec &outrec) noexcept;
+  bool IsHole(TEdge *e) noexcept;
+  bool FindOwnerFromSplitRecs(OutRec &outRec, OutRec *&currOrfl) noexcept;
+  void FixHoleLinkage(OutRec &outrec) noexcept;
+  void AddJoin(OutPt *op1, OutPt *op2, const IntPoint offPt) noexcept;
+  void ClearJoins() noexcept;
+  void ClearGhostJoins() noexcept;
+  void AddGhostJoin(OutPt *op, const IntPoint offPt) noexcept;
+  bool JoinPoints(Join *j, OutRec *outRec1, OutRec *outRec2) noexcept;
+  void JoinCommonEdges() noexcept;
+  void DoSimplePolygons() noexcept;
+  void FixupFirstLefts1(OutRec *OldOutRec, OutRec *NewOutRec) noexcept;
+  void FixupFirstLefts2(OutRec *InnerOutRec, OutRec *OuterOutRec) noexcept;
+  void FixupFirstLefts3(OutRec *OldOutRec, OutRec *NewOutRec) noexcept;
+#ifdef use_xyz
+  void SetZ(IntPoint &pt, TEdge &e1, TEdge &e2) noexcept;
+#endif
+};
+//------------------------------------------------------------------------------
+
+class ClipperOffset {
+public:
+  ClipperOffset(double miterLimit = 2.0, double roundPrecision = 0.25) noexcept;
+  ~ClipperOffset();
+  void AddPath(const Path &path, JoinType joinType, EndType endType) noexcept;
+  void AddPaths(const Paths &paths, JoinType joinType,
+                EndType endType) noexcept;
+  bool Execute(Paths &solution, double delta) noexcept;
+  bool Execute(PolyTree &solution, double delta) noexcept;
+  void Clear() noexcept;
+
+private:
+  double MiterLimit;
+  double ArcTolerance;
+
+  Paths m_destPolys;
+  Path m_srcPoly;
+  Path m_destPoly;
+  std::vector<DoublePoint> m_normals;
+  double m_delta, m_sinA, m_sin, m_cos;
+  double m_miterLim, m_StepsPerRad;
+  IntPoint m_lowest;
+  PolyNode m_polyNodes;
+
+  void FixOrientations() noexcept;
+  void DoOffset(double delta) noexcept;
+  void OffsetPoint(int j, int &k, JoinType jointype) noexcept;
+  void DoSquare(int j, int k) noexcept;
+  void DoMiter(int j, int k, double r) noexcept;
+  void DoRound(int j, int k) noexcept;
+};
+//------------------------------------------------------------------------------
+
+class clipperException : public std::exception {
+public:
+  clipperException(const char *description) noexcept : m_descr(description) {}
+  ~clipperException() {}
+  virtual const char *what() const noexcept { return m_descr.c_str(); }
+
+private:
+  std::string m_descr;
+};
+//------------------------------------------------------------------------------
+
+} // namespace ClipperLib
+
+#endif // clipper_hpp

deploy/cpp_infer/include/ocr_cls.h

@@ -0,0 +1,102 @@
+// 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 <fstream>
+#include <include/preprocess_op.h>
+#include <include/utility.h>
+#include <iostream>
+#include <memory>
+#include <yaml-cpp/yaml.h>
+
+namespace paddle_infer {
+class Predictor;
+}
+
+namespace PaddleOCR {
+
+class Classifier {
+public:
+  explicit Classifier(const std::string &model_dir, const bool &use_gpu,
+                      const int &gpu_id, const int &gpu_mem,
+                      const int &cpu_math_library_num_threads,
+                      const bool &use_mkldnn, const double &cls_thresh,
+                      const bool &use_tensorrt, const std::string &precision,
+                      const int &cls_batch_num) noexcept {
+    this->use_gpu_ = use_gpu;
+    this->gpu_id_ = gpu_id;
+    this->gpu_mem_ = gpu_mem;
+    this->cpu_math_library_num_threads_ = cpu_math_library_num_threads;
+    this->use_mkldnn_ = use_mkldnn;
+
+    this->cls_thresh = cls_thresh;
+    this->use_tensorrt_ = use_tensorrt;
+    this->precision_ = precision;
+    this->cls_batch_num_ = cls_batch_num;
+
+    std::string yaml_file_path = model_dir + "/inference.yml";
+    std::ifstream yaml_file(yaml_file_path);
+    if (yaml_file.is_open()) {
+      std::string model_name;
+      try {
+        YAML::Node config = YAML::LoadFile(yaml_file_path);
+        if (config["Global"] && config["Global"]["model_name"]) {
+          model_name = config["Global"]["model_name"].as<std::string>();
+        }
+        if (!model_name.empty() &&
+            model_name != "PP-LCNet_x0_25_textline_ori" &&
+            model_name != "PP-LCNet_x1_0_textline_ori") {
+          std::cerr << "Error: " << model_name << " is currently not supported."
+                    << std::endl;
+          std::exit(EXIT_FAILURE);
+        }
+      } catch (const YAML::Exception &e) {
+        std::cerr << "Failed to load YAML file: " << e.what() << std::endl;
+      }
+    }
+
+    LoadModel(model_dir);
+  }
+  double cls_thresh = 0.9;
+
+  // Load Paddle inference model
+  void LoadModel(const std::string &model_dir) noexcept;
+
+  void Run(const std::vector<cv::Mat> &img_list, std::vector<int> &cls_labels,
+           std::vector<float> &cls_scores, std::vector<double> &times) noexcept;
+
+private:
+  std::shared_ptr<paddle_infer::Predictor> predictor_;
+
+  bool use_gpu_ = false;
+  int gpu_id_ = 0;
+  int gpu_mem_ = 4000;
+  int cpu_math_library_num_threads_ = 4;
+  bool use_mkldnn_ = false;
+
+  std::vector<float> mean_ = {0.5f, 0.5f, 0.5f};
+  std::vector<float> scale_ = {1 / 0.5f, 1 / 0.5f, 1 / 0.5f};
+  bool is_scale_ = true;
+  bool use_tensorrt_ = false;
+  std::string precision_ = "fp32";
+  int cls_batch_num_ = 1;
+  // pre-process
+  ClsResizeImg resize_op_;
+  Normalize normalize_op_;
+  PermuteBatch permute_op_;
+
+}; // class Classifier
+
+} // namespace PaddleOCR

deploy/cpp_infer/include/ocr_det.h

@@ -0,0 +1,126 @@
+// 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 <fstream>
+#include <include/postprocess_op.h>
+#include <include/preprocess_op.h>
+#include <iostream>
+#include <memory>
+#include <yaml-cpp/yaml.h>
+
+namespace paddle_infer {
+class Predictor;
+}
+
+namespace PaddleOCR {
+
+class DBDetector {
+public:
+  explicit DBDetector(const std::string &model_dir, const bool &use_gpu,
+                      const int &gpu_id, const int &gpu_mem,
+                      const int &cpu_math_library_num_threads,
+                      const bool &use_mkldnn, const std::string &limit_type,
+                      const int &limit_side_len, const double &det_db_thresh,
+                      const double &det_db_box_thresh,
+                      const double &det_db_unclip_ratio,
+                      const std::string &det_db_score_mode,
+                      const bool &use_dilation, const bool &use_tensorrt,
+                      const std::string &precision) noexcept {
+    this->use_gpu_ = use_gpu;
+    this->gpu_id_ = gpu_id;
+    this->gpu_mem_ = gpu_mem;
+    this->cpu_math_library_num_threads_ = cpu_math_library_num_threads;
+    this->use_mkldnn_ = use_mkldnn;
+
+    this->limit_type_ = limit_type;
+    this->limit_side_len_ = limit_side_len;
+
+    this->det_db_thresh_ = det_db_thresh;
+    this->det_db_box_thresh_ = det_db_box_thresh;
+    this->det_db_unclip_ratio_ = det_db_unclip_ratio;
+    this->det_db_score_mode_ = det_db_score_mode;
+    this->use_dilation_ = use_dilation;
+
+    this->use_tensorrt_ = use_tensorrt;
+    this->precision_ = precision;
+
+    std::string yaml_file_path = model_dir + "/inference.yml";
+    std::ifstream yaml_file(yaml_file_path);
+    if (yaml_file.is_open()) {
+      std::string model_name;
+      try {
+        YAML::Node config = YAML::LoadFile(yaml_file_path);
+        if (config["Global"] && config["Global"]["model_name"]) {
+          model_name = config["Global"]["model_name"].as<std::string>();
+        }
+        if (!model_name.empty() && model_name != "PP-OCRv5_mobile_det" &&
+            model_name != "PP-OCRv5_server_det") {
+          std::cerr << "Error: " << model_name << " is currently not supported."
+                    << std::endl;
+          std::exit(EXIT_FAILURE);
+        }
+      } catch (const YAML::Exception &e) {
+        std::cerr << "Failed to load YAML file: " << e.what() << std::endl;
+      }
+    }
+
+    LoadModel(model_dir);
+  }
+
+  // Load Paddle inference model
+  void LoadModel(const std::string &model_dir) noexcept;
+
+  // Run predictor
+  void Run(const cv::Mat &img,
+           std::vector<std::vector<std::vector<int>>> &boxes,
+           std::vector<double> &times) noexcept;
+
+private:
+  std::shared_ptr<paddle_infer::Predictor> predictor_;
+
+  bool use_gpu_ = false;
+  int gpu_id_ = 0;
+  int gpu_mem_ = 4000;
+  int cpu_math_library_num_threads_ = 4;
+  bool use_mkldnn_ = false;
+
+  std::string limit_type_ = "max";
+  int limit_side_len_ = 960;
+
+  double det_db_thresh_ = 0.3;
+  double det_db_box_thresh_ = 0.5;
+  double det_db_unclip_ratio_ = 2.0;
+  std::string det_db_score_mode_ = "slow";
+  bool use_dilation_ = false;
+
+  bool visualize_ = true;
+  bool use_tensorrt_ = false;
+  std::string precision_ = "fp32";
+
+  std::vector<float> mean_ = {0.485f, 0.456f, 0.406f};
+  std::vector<float> scale_ = {1 / 0.229f, 1 / 0.224f, 1 / 0.225f};
+  bool is_scale_ = true;
+
+  // pre-process
+  ResizeImgType0 resize_op_;
+  Normalize normalize_op_;
+  Permute permute_op_;
+
+  // post-process
+  DBPostProcessor post_processor_;
+};
+
+} // namespace PaddleOCR

deploy/cpp_infer/include/ocr_rec.h

@@ -0,0 +1,133 @@
+// 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 <fstream>
+#include <include/preprocess_op.h>
+#include <include/utility.h>
+#include <iostream>
+#include <memory>
+#include <yaml-cpp/yaml.h>
+
+namespace paddle_infer {
+class Predictor;
+}
+
+namespace PaddleOCR {
+
+class CRNNRecognizer {
+public:
+  explicit CRNNRecognizer(const std::string &model_dir, const bool &use_gpu,
+                          const int &gpu_id, const int &gpu_mem,
+                          const int &cpu_math_library_num_threads,
+                          const bool &use_mkldnn, const std::string &label_path,
+                          const bool &use_tensorrt,
+                          const std::string &precision,
+                          const int &rec_batch_num, const int &rec_img_h,
+                          const int &rec_img_w) noexcept {
+    this->use_gpu_ = use_gpu;
+    this->gpu_id_ = gpu_id;
+    this->gpu_mem_ = gpu_mem;
+    this->cpu_math_library_num_threads_ = cpu_math_library_num_threads;
+    this->use_mkldnn_ = use_mkldnn;
+    this->use_tensorrt_ = use_tensorrt;
+    this->precision_ = precision;
+    this->rec_batch_num_ = rec_batch_num;
+    this->rec_img_h_ = rec_img_h;
+    this->rec_img_w_ = rec_img_w;
+    std::vector<int> rec_image_shape = {3, rec_img_h, rec_img_w};
+    this->rec_image_shape_ = rec_image_shape;
+
+    std::string new_label_path = label_path;
+    std::string yaml_file_path = model_dir + "/inference.yml";
+    std::ifstream yaml_file(yaml_file_path);
+    if (yaml_file.is_open()) {
+      std::string model_name;
+      std::vector<std::string> rec_char_list;
+      try {
+        YAML::Node config = YAML::LoadFile(yaml_file_path);
+        if (config["Global"] && config["Global"]["model_name"]) {
+          model_name = config["Global"]["model_name"].as<std::string>();
+        }
+        if (!model_name.empty() && model_name != "PP-OCRv5_mobile_rec" &&
+            model_name != "PP-OCRv5_server_rec") {
+          std::cerr << "Error: " << model_name << " is currently not supported."
+                    << std::endl;
+          std::exit(EXIT_FAILURE);
+        }
+        if (config["PostProcess"] && config["PostProcess"]["character_dict"]) {
+          rec_char_list = config["PostProcess"]["character_dict"]
+                              .as<std::vector<std::string>>();
+        }
+      } catch (const YAML::Exception &e) {
+        std::cerr << "Failed to load YAML file: " << e.what() << std::endl;
+      }
+      if (label_path == "../../ppocr/utils/ppocr_keys_v1.txt" &&
+          !rec_char_list.empty()) {
+        std::string new_rec_char_dict_path = model_dir + "/ppocr_keys.txt";
+        std::ofstream new_file(new_rec_char_dict_path);
+        if (new_file.is_open()) {
+          for (const auto &character : rec_char_list) {
+            new_file << character << '\n';
+          }
+          new_label_path = new_rec_char_dict_path;
+        }
+      }
+    }
+
+    this->label_list_ = Utility::ReadDict(new_label_path);
+    this->label_list_.emplace(this->label_list_.begin(),
+                              "#"); // blank char for ctc
+    this->label_list_.emplace_back(" ");
+
+    LoadModel(model_dir);
+  }
+
+  // Load Paddle inference model
+  void LoadModel(const std::string &model_dir) noexcept;
+
+  void Run(const std::vector<cv::Mat> &img_list,
+           std::vector<std::string> &rec_texts,
+           std::vector<float> &rec_text_scores,
+           std::vector<double> &times) noexcept;
+
+private:
+  std::shared_ptr<paddle_infer::Predictor> predictor_;
+
+  bool use_gpu_ = false;
+  int gpu_id_ = 0;
+  int gpu_mem_ = 4000;
+  int cpu_math_library_num_threads_ = 4;
+  bool use_mkldnn_ = false;
+
+  std::vector<std::string> label_list_;
+
+  std::vector<float> mean_ = {0.5f, 0.5f, 0.5f};
+  std::vector<float> scale_ = {1 / 0.5f, 1 / 0.5f, 1 / 0.5f};
+  bool is_scale_ = true;
+  bool use_tensorrt_ = false;
+  std::string precision_ = "fp32";
+  int rec_batch_num_ = 6;
+  int rec_img_h_ = 32;
+  int rec_img_w_ = 320;
+  std::vector<int> rec_image_shape_ = {3, rec_img_h_, rec_img_w_};
+  // pre-process
+  CrnnResizeImg resize_op_;
+  Normalize normalize_op_;
+  PermuteBatch permute_op_;
+
+}; // class CrnnRecognizer
+
+} // namespace PaddleOCR

deploy/cpp_infer/include/paddleocr.h

@@ -0,0 +1,52 @@
+// 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 <include/utility.h>
+
+namespace PaddleOCR {
+
+class PPOCR {
+public:
+  explicit PPOCR() noexcept;
+  virtual ~PPOCR();
+
+  std::vector<std::vector<OCRPredictResult>>
+  ocr(const std::vector<cv::Mat> &img_list, bool det = true, bool rec = true,
+      bool cls = true) noexcept;
+  std::vector<OCRPredictResult> ocr(const cv::Mat &img, bool det = true,
+                                    bool rec = true, bool cls = true) noexcept;
+
+  void reset_timer() noexcept;
+  void benchmark_log(int img_num) noexcept;
+
+protected:
+  std::vector<double> time_info_det = {0, 0, 0};
+  std::vector<double> time_info_rec = {0, 0, 0};
+  std::vector<double> time_info_cls = {0, 0, 0};
+
+  void det(const cv::Mat &img,
+           std::vector<OCRPredictResult> &ocr_results) noexcept;
+  void rec(const std::vector<cv::Mat> &img_list,
+           std::vector<OCRPredictResult> &ocr_results) noexcept;
+  void cls(const std::vector<cv::Mat> &img_list,
+           std::vector<OCRPredictResult> &ocr_results) noexcept;
+
+private:
+  struct PPOCR_PRIVATE;
+  PPOCR_PRIVATE *pri_;
+};
+
+} // namespace PaddleOCR

deploy/cpp_infer/include/paddlestructure.h

@@ -0,0 +1,66 @@
+// Copyright (c) 2022 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 <include/paddleocr.h>
+
+namespace PaddleOCR {
+
+class PaddleStructure : public PPOCR {
+public:
+  explicit PaddleStructure() noexcept;
+  ~PaddleStructure();
+
+  std::vector<StructurePredictResult> structure(const cv::Mat &img,
+                                                bool layout = false,
+                                                bool table = true,
+                                                bool ocr = false) noexcept;
+
+  void reset_timer() noexcept;
+  void benchmark_log(int img_num) noexcept;
+
+private:
+  struct STRUCTURE_PRIVATE;
+  STRUCTURE_PRIVATE *pri_;
+
+  std::vector<double> time_info_table = {0, 0, 0};
+  std::vector<double> time_info_layout = {0, 0, 0};
+
+  void layout(const cv::Mat &img,
+              std::vector<StructurePredictResult> &structure_result) noexcept;
+
+  void table(const cv::Mat &img,
+             StructurePredictResult &structure_result) noexcept;
+
+  std::string rebuild_table(const std::vector<std::string> &rec_html_tags,
+                            const std::vector<std::vector<int>> &rec_boxes,
+                            std::vector<OCRPredictResult> &ocr_result) noexcept;
+
+  float dis(const std::vector<int> &box1,
+            const std::vector<int> &box2) noexcept;
+
+  static bool comparison_dis(const std::vector<float> &dis1,
+                             const std::vector<float> &dis2) noexcept {
+    if (dis1[1] < dis2[1]) {
+      return true;
+    } else if (dis1[1] == dis2[1]) {
+      return dis1[0] < dis2[0];
+    } else {
+      return false;
+    }
+  }
+};
+
+} // namespace PaddleOCR

deploy/cpp_infer/include/postprocess_op.h

@@ -0,0 +1,129 @@
+// 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 <include/utility.h>
+
+namespace PaddleOCR {
+
+class DBPostProcessor {
+public:
+  void GetContourArea(const std::vector<std::vector<float>> &box,
+                      float unclip_ratio, float &distance) noexcept;
+
+  cv::RotatedRect UnClip(const std::vector<std::vector<float>> &box,
+                         const float &unclip_ratio) noexcept;
+
+  float **Mat2Vec(const cv::Mat &mat) noexcept;
+
+  std::vector<std::vector<int>>
+  OrderPointsClockwise(const std::vector<std::vector<int>> &pts) noexcept;
+
+  std::vector<std::vector<float>> GetMiniBoxes(const cv::RotatedRect &box,
+                                               float &ssid) noexcept;
+
+  float BoxScoreFast(const std::vector<std::vector<float>> &box_array,
+                     const cv::Mat &pred) noexcept;
+  float PolygonScoreAcc(const std::vector<cv::Point> &contour,
+                        const cv::Mat &pred) noexcept;
+
+  std::vector<std::vector<std::vector<int>>>
+  BoxesFromBitmap(const cv::Mat &pred, const cv::Mat &bitmap,
+                  const float &box_thresh, const float &det_db_unclip_ratio,
+                  const std::string &det_db_score_mode) noexcept;
+
+  void FilterTagDetRes(std::vector<std::vector<std::vector<int>>> &boxes,
+                       float ratio_h, float ratio_w,
+                       const cv::Mat &srcimg) noexcept;
+
+private:
+  static bool XsortInt(const std::vector<int> &a,
+                       const std::vector<int> &b) noexcept;
+
+  static bool XsortFp32(const std::vector<float> &a,
+                        const std::vector<float> &b) noexcept;
+
+  std::vector<std::vector<float>> Mat2Vector(const cv::Mat &mat) noexcept;
+
+  inline int _max(int a, int b) const noexcept { return a >= b ? a : b; }
+
+  inline int _min(int a, int b) const noexcept { return a >= b ? b : a; }
+
+  template <class T> inline T clamp(T x, T min, T max) const noexcept {
+    if (x > max)
+      return max;
+    if (x < min)
+      return min;
+    return x;
+  }
+
+  inline float clampf(float x, float min, float max) const noexcept {
+    if (x > max)
+      return max;
+    if (x < min)
+      return min;
+    return x;
+  }
+};
+
+class TablePostProcessor {
+public:
+  void init(const std::string &label_path,
+            bool merge_no_span_structure = true) noexcept;
+  void Run(const std::vector<float> &loc_preds,
+           const std::vector<float> &structure_probs,
+           std::vector<float> &rec_scores,
+           const std::vector<int> &loc_preds_shape,
+           const std::vector<int> &structure_probs_shape,
+           std::vector<std::vector<std::string>> &rec_html_tag_batch,
+           std::vector<std::vector<std::vector<int>>> &rec_boxes_batch,
+           const std::vector<int> &width_list,
+           const std::vector<int> &height_list) noexcept;
+
+private:
+  std::vector<std::string> label_list_;
+  const std::string end = "eos";
+  const std::string beg = "sos";
+};
+
+class PicodetPostProcessor {
+public:
+  void init(const std::string &label_path, const double score_threshold = 0.4,
+            const double nms_threshold = 0.5,
+            const std::vector<int> &fpn_stride = {8, 16, 32, 64}) noexcept;
+  void Run(std::vector<StructurePredictResult> &results,
+           const std::vector<std::vector<float>> &outs,
+           const std::vector<int> &ori_shape,
+           const std::vector<int> &resize_shape, int eg_max) noexcept;
+  inline size_t fpn_stride_size() const noexcept { return fpn_stride_.size(); }
+
+private:
+  StructurePredictResult disPred2Bbox(const std::vector<float> &bbox_pred,
+                                      int label, float score, int x, int y,
+                                      int stride,
+                                      const std::vector<int> &im_shape,
+                                      int reg_max) noexcept;
+  void nms(std::vector<StructurePredictResult> &input_boxes,
+           float nms_threshold) noexcept;
+
+  std::vector<int> fpn_stride_ = {8, 16, 32, 64};
+
+  std::vector<std::string> label_list_;
+  double score_threshold_ = 0.4;
+  double nms_threshold_ = 0.5;
+  int num_class_ = 5;
+};
+
+} // namespace PaddleOCR

deploy/cpp_infer/include/preprocess_op.h

@@ -0,0 +1,79 @@
+// 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 <opencv2/imgproc.hpp>
+
+namespace PaddleOCR {
+
+class Normalize {
+public:
+  virtual void Run(cv::Mat &im, const std::vector<float> &mean,
+                   const std::vector<float> &scale,
+                   const bool is_scale = true) noexcept;
+};
+
+// RGB -> CHW
+class Permute {
+public:
+  virtual void Run(const cv::Mat &im, float *data) noexcept;
+};
+
+class PermuteBatch {
+public:
+  virtual void Run(const std::vector<cv::Mat> &imgs, float *data) noexcept;
+};
+
+class ResizeImgType0 {
+public:
+  virtual void Run(const cv::Mat &img, cv::Mat &resize_img,
+                   const std::string &limit_type, int limit_side_len,
+                   float &ratio_h, float &ratio_w, bool use_tensorrt) noexcept;
+};
+
+class CrnnResizeImg {
+public:
+  virtual void Run(const cv::Mat &img, cv::Mat &resize_img, float wh_ratio,
+                   bool use_tensorrt = false,
+                   const std::vector<int> &rec_image_shape = {3, 32,
+                                                              320}) noexcept;
+};
+
+class ClsResizeImg {
+public:
+  virtual void
+  Run(const cv::Mat &img, cv::Mat &resize_img, bool use_tensorrt = false,
+      const std::vector<int> &rec_image_shape = {3, 48, 192}) noexcept;
+};
+
+class TableResizeImg {
+public:
+  virtual void Run(const cv::Mat &img, cv::Mat &resize_img,
+                   const int max_len = 488) noexcept;
+};
+
+class TablePadImg {
+public:
+  virtual void Run(const cv::Mat &img, cv::Mat &resize_img,
+                   const int max_len = 488) noexcept;
+};
+
+class Resize {
+public:
+  virtual void Run(const cv::Mat &img, cv::Mat &resize_img, const int h,
+                   const int w) noexcept;
+};
+
+} // namespace PaddleOCR

deploy/cpp_infer/include/structure_layout.h

@@ -0,0 +1,119 @@
+// 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 <fstream>
+#include <include/postprocess_op.h>
+#include <include/preprocess_op.h>
+#include <iostream>
+#include <memory>
+#include <yaml-cpp/yaml.h>
+
+namespace paddle_infer {
+class Predictor;
+}
+
+namespace PaddleOCR {
+
+class StructureLayoutRecognizer {
+public:
+  explicit StructureLayoutRecognizer(
+      const std::string &model_dir, const bool &use_gpu, const int &gpu_id,
+      const int &gpu_mem, const int &cpu_math_library_num_threads,
+      const bool &use_mkldnn, const std::string &label_path,
+      const bool &use_tensorrt, const std::string &precision,
+      const double &layout_score_threshold,
+      const double &layout_nms_threshold) noexcept {
+    this->use_gpu_ = use_gpu;
+    this->gpu_id_ = gpu_id;
+    this->gpu_mem_ = gpu_mem;
+    this->cpu_math_library_num_threads_ = cpu_math_library_num_threads;
+    this->use_mkldnn_ = use_mkldnn;
+    this->use_tensorrt_ = use_tensorrt;
+    this->precision_ = precision;
+
+    std::string new_label_path = label_path;
+    std::string yaml_file_path = model_dir + "/inference.yml";
+    std::ifstream yaml_file(yaml_file_path);
+    if (yaml_file.is_open()) {
+      std::string model_name;
+      std::vector<std::string> rec_char_list;
+      try {
+        YAML::Node config = YAML::LoadFile(yaml_file_path);
+        if (config["Global"] && config["Global"]["model_name"]) {
+          model_name = config["Global"]["model_name"].as<std::string>();
+        }
+        if (!model_name.empty()) {
+          std::cerr << "Error: " << model_name << " is currently not supported."
+                    << std::endl;
+          std::exit(EXIT_FAILURE);
+        }
+        if (config["PostProcess"] && config["PostProcess"]["character_dict"]) {
+          rec_char_list = config["PostProcess"]["character_dict"]
+                              .as<std::vector<std::string>>();
+        }
+      } catch (const YAML::Exception &e) {
+        std::cerr << "Failed to load YAML file: " << e.what() << std::endl;
+      }
+      if (label_path == "../../ppocr/utils/ppocr_keys_v1.txt" &&
+          !rec_char_list.empty()) {
+        std::string new_rec_char_dict_path = model_dir + "/ppocr_keys.txt";
+        std::ofstream new_file(new_rec_char_dict_path);
+        if (new_file.is_open()) {
+          for (const auto &character : rec_char_list) {
+            new_file << character << '\n';
+          }
+          new_label_path = new_rec_char_dict_path;
+        }
+      }
+    }
+
+    this->post_processor_.init(new_label_path, layout_score_threshold,
+                               layout_nms_threshold);
+    LoadModel(model_dir);
+  }
+
+  // Load Paddle inference model
+  void LoadModel(const std::string &model_dir) noexcept;
+
+  void Run(const cv::Mat &img, std::vector<StructurePredictResult> &result,
+           std::vector<double> &times) noexcept;
+
+private:
+  std::shared_ptr<paddle_infer::Predictor> predictor_;
+
+  bool use_gpu_ = false;
+  int gpu_id_ = 0;
+  int gpu_mem_ = 4000;
+  int cpu_math_library_num_threads_ = 4;
+  bool use_mkldnn_ = false;
+
+  std::vector<float> mean_ = {0.485f, 0.456f, 0.406f};
+  std::vector<float> scale_ = {1 / 0.229f, 1 / 0.224f, 1 / 0.225f};
+  bool is_scale_ = true;
+
+  bool use_tensorrt_ = false;
+  std::string precision_ = "fp32";
+
+  // pre-process
+  Resize resize_op_;
+  Normalize normalize_op_;
+  Permute permute_op_;
+
+  // post-process
+  PicodetPostProcessor post_processor_;
+};
+
+} // namespace PaddleOCR

deploy/cpp_infer/include/structure_table.h

@@ -0,0 +1,127 @@
+// 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 <fstream>
+#include <include/postprocess_op.h>
+#include <include/preprocess_op.h>
+#include <iostream>
+#include <memory>
+#include <yaml-cpp/yaml.h>
+
+namespace paddle_infer {
+class Predictor;
+}
+
+namespace PaddleOCR {
+
+class StructureTableRecognizer {
+public:
+  explicit StructureTableRecognizer(
+      const std::string &model_dir, const bool &use_gpu, const int &gpu_id,
+      const int &gpu_mem, const int &cpu_math_library_num_threads,
+      const bool &use_mkldnn, const std::string &label_path,
+      const bool &use_tensorrt, const std::string &precision,
+      const int &table_batch_num, const int &table_max_len,
+      const bool &merge_no_span_structure) noexcept {
+    this->use_gpu_ = use_gpu;
+    this->gpu_id_ = gpu_id;
+    this->gpu_mem_ = gpu_mem;
+    this->cpu_math_library_num_threads_ = cpu_math_library_num_threads;
+    this->use_mkldnn_ = use_mkldnn;
+    this->use_tensorrt_ = use_tensorrt;
+    this->precision_ = precision;
+    this->table_batch_num_ = table_batch_num;
+    this->table_max_len_ = table_max_len;
+
+    std::string new_label_path = label_path;
+    std::string yaml_file_path = model_dir + "/inference.yml";
+    std::ifstream yaml_file(yaml_file_path);
+    if (yaml_file.is_open()) {
+      std::string model_name;
+      std::vector<std::string> rec_char_list;
+      try {
+        YAML::Node config = YAML::LoadFile(yaml_file_path);
+        if (config["Global"] && config["Global"]["model_name"]) {
+          model_name = config["Global"]["model_name"].as<std::string>();
+        }
+        if (!model_name.empty()) {
+          std::cerr << "Error: " << model_name << " is currently not supported."
+                    << std::endl;
+          std::exit(EXIT_FAILURE);
+        }
+        if (config["PostProcess"] && config["PostProcess"]["character_dict"]) {
+          rec_char_list = config["PostProcess"]["character_dict"]
+                              .as<std::vector<std::string>>();
+        }
+      } catch (const YAML::Exception &e) {
+        std::cerr << "Failed to load YAML file: " << e.what() << std::endl;
+      }
+      if (label_path == "../../ppocr/utils/ppocr_keys_v1.txt" &&
+          !rec_char_list.empty()) {
+        std::string new_rec_char_dict_path = model_dir + "/ppocr_keys.txt";
+        std::ofstream new_file(new_rec_char_dict_path);
+        if (new_file.is_open()) {
+          for (const auto &character : rec_char_list) {
+            new_file << character << '\n';
+          }
+          new_label_path = new_rec_char_dict_path;
+        }
+      }
+    }
+
+    this->post_processor_.init(new_label_path, merge_no_span_structure);
+    LoadModel(model_dir);
+  }
+
+  // Load Paddle inference model
+  void LoadModel(const std::string &model_dir) noexcept;
+
+  void Run(const std::vector<cv::Mat> &img_list,
+           std::vector<std::vector<std::string>> &rec_html_tags,
+           std::vector<float> &rec_scores,
+           std::vector<std::vector<std::vector<int>>> &rec_boxes,
+           std::vector<double> &times) noexcept;
+
+private:
+  std::shared_ptr<paddle_infer::Predictor> predictor_;
+
+  bool use_gpu_ = false;
+  int gpu_id_ = 0;
+  int gpu_mem_ = 4000;
+  int cpu_math_library_num_threads_ = 4;
+  bool use_mkldnn_ = false;
+  int table_max_len_ = 488;
+
+  std::vector<float> mean_ = {0.485f, 0.456f, 0.406f};
+  std::vector<float> scale_ = {1 / 0.229f, 1 / 0.224f, 1 / 0.225f};
+  bool is_scale_ = true;
+
+  bool use_tensorrt_ = false;
+  std::string precision_ = "fp32";
+  int table_batch_num_ = 1;
+
+  // pre-process
+  TableResizeImg resize_op_;
+  Normalize normalize_op_;
+  PermuteBatch permute_op_;
+  TablePadImg pad_op_;
+
+  // post-process
+  TablePostProcessor post_processor_;
+
+}; // class StructureTableRecognizer
+
+} // namespace PaddleOCR

deploy/cpp_infer/include/utility.h

@@ -0,0 +1,113 @@
+// 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 <opencv2/imgproc.hpp>
+
+namespace PaddleOCR {
+
+struct OCRPredictResult {
+  std::vector<std::vector<int>> box;
+  std::string text;
+  float score = -1.0;
+  float cls_score;
+  int cls_label = -1;
+};
+
+struct StructurePredictResult {
+  std::vector<float> box;
+  std::vector<std::vector<int>> cell_box;
+  std::string type;
+  std::vector<OCRPredictResult> text_res;
+  std::string html;
+  float html_score = -1;
+  float confidence;
+};
+
+class Utility {
+public:
+  static std::vector<std::string> ReadDict(const std::string &path) noexcept;
+
+  static void VisualizeBboxes(const cv::Mat &srcimg,
+                              const std::vector<OCRPredictResult> &ocr_result,
+                              const std::string &save_path) noexcept;
+
+  static void VisualizeBboxes(const cv::Mat &srcimg,
+                              const StructurePredictResult &structure_result,
+                              const std::string &save_path) noexcept;
+
+  template <class ForwardIterator>
+  inline static size_t argmax(ForwardIterator first,
+                              ForwardIterator last) noexcept {
+    return std::distance(first, std::max_element(first, last));
+  }
+
+  static void GetAllFiles(const char *dir_name,
+                          std::vector<std::string> &all_inputs) noexcept;
+
+  static cv::Mat
+  GetRotateCropImage(const cv::Mat &srcimage,
+                     const std::vector<std::vector<int>> &box) noexcept;
+
+  static std::vector<size_t> argsort(const std::vector<float> &array) noexcept;
+
+  static std::string basename(const std::string &filename) noexcept;
+
+  static bool PathExists(const char *path) noexcept;
+  static inline bool PathExists(const std::string &path) noexcept {
+    return PathExists(path.c_str());
+  }
+
+  static void CreateDir(const char *path) noexcept;
+  static inline void CreateDir(const std::string &path) noexcept {
+    CreateDir(path.c_str());
+  }
+
+  static void
+  print_result(const std::vector<OCRPredictResult> &ocr_result) noexcept;
+
+  static cv::Mat crop_image(const cv::Mat &img,
+                            const std::vector<int> &area) noexcept;
+  static cv::Mat crop_image(const cv::Mat &img,
+                            const std::vector<float> &area) noexcept;
+
+  static void sort_boxes(std::vector<OCRPredictResult> &ocr_result) noexcept;
+
+  static std::vector<int>
+  xyxyxyxy2xyxy(const std::vector<std::vector<int>> &box) noexcept;
+  static std::vector<int> xyxyxyxy2xyxy(const std::vector<int> &box) noexcept;
+
+  static float fast_exp(float x) noexcept;
+  static std::vector<float>
+  activation_function_softmax(const std::vector<float> &src) noexcept;
+  static float iou(const std::vector<int> &box1,
+                   const std::vector<int> &box2) noexcept;
+  static float iou(const std::vector<float> &box1,
+                   const std::vector<float> &box2) noexcept;
+
+private:
+  static bool comparison_box(const OCRPredictResult &result1,
+                             const OCRPredictResult &result2) noexcept {
+    if (result1.box[0][1] < result2.box[0][1]) {
+      return true;
+    } else if (result1.box[0][1] == result2.box[0][1]) {
+      return result1.box[0][0] < result2.box[0][0];
+    } else {
+      return false;
+    }
+  }
+};
+
+} // namespace PaddleOCR