tesseract/ccstruct/detlinefit.cpp

///////////////////////////////////////////////////////////////////////
// File:        detlinefit.cpp
// Description: Deterministic least median squares line fitting.
// Author:      Ray Smith
// Created:     Thu Feb 28 14:45:01 PDT 2008
//
// (C) Copyright 2008, Google Inc.
// 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 "detlinefit.h"
#include "statistc.h"
#include "ndminx.h"

namespace tesseract {

// The number of points to consider at each end.
const int kNumEndPoints = 3;

DetLineFit::DetLineFit() {
}

DetLineFit::~DetLineFit() {
}

// Delete all Added points.
void DetLineFit::Clear() {
  pt_list_.clear();
}

// Add a new point. Takes a copy - the pt doesn't need to stay in scope.
void DetLineFit::Add(const ICOORD& pt) {
  ICOORDELT_IT it = &pt_list_;
  ICOORDELT* new_pt = new ICOORDELT(pt);
  it.add_to_end(new_pt);
}

// Fit a line to the points, returning the fitted line as a pair of
// points, and the upper quartile error.
double DetLineFit::Fit(ICOORD* pt1, ICOORD* pt2) {
  ICOORDELT_IT it(&pt_list_);
  // Do something sensible with no points.
  if (pt_list_.empty()) {
    pt1->set_x(0);
    pt1->set_y(0);
    *pt2 = *pt1;
    return 0.0;
  }
  // Count the points and find the first and last kNumEndPoints.
  ICOORD* starts[kNumEndPoints];
  ICOORD* ends[kNumEndPoints];
  int pt_count = 0;
  for (it.mark_cycle_pt(); !it.cycled_list(); it.forward()) {
    if (pt_count < kNumEndPoints) {
      starts[pt_count] = it.data();
      ends[pt_count] = starts[pt_count];
    } else {
      for (int i = 1; i < kNumEndPoints; ++i)
        ends[i - 1] = ends[i];
      ends[kNumEndPoints - 1] = it.data();
    }
    ++pt_count;
  }
  // 1 or 2 points need special treatment.
  if (pt_count <= 2) {
    *pt1 = *starts[0];
    if (pt_count > 1)
      *pt2 = *starts[1];
    else
      *pt2 = *pt1;
    return 0.0;
  }
  int end_count = MIN(pt_count, kNumEndPoints);
  int* distances = new int[pt_count];
  double best_uq = -1.0;
  // Iterate each pair of points and find the best fitting line.
  for (int i = 0; i < end_count; ++i) {
    ICOORD* start = starts[i];
    for (int j = 0; j < end_count; ++j) {
      ICOORD* end = ends[j];
      if (start != end) {
        // Compute the upper quartile error from the line.
        double dist = ComputeErrors(*start, *end, distances);
        if (dist < best_uq || best_uq < 0.0) {
          best_uq = dist;
          *pt1 = *start;
          *pt2 = *end;
        }
      }
    }
  }
  delete [] distances;
  // Finally compute the square root to return the true distance.
  return best_uq > 0.0 ? sqrt(best_uq) : best_uq;
}

// Backwards compatible fit returning a gradient and constant.
// Deprecated. Prefer Fit(ICOORD*, ICOORD*) where possible, but use this
// function in preference to the LMS class.
double DetLineFit::Fit(float* m, float* c) {
  ICOORD start, end;
  double error = Fit(&start, &end);
  if (end.x() != start.x()) {
    *m = static_cast<float>(end.y() - start.y()) / (end.x() - start.x());
    *c = start.y() - *m * start.x();
  } else {
    *m = 0.0f;
    *c = 0.0f;
  }
  return error;
}

// Helper function to compute a fictitious end point that is on a line
// of a given gradient through the given start.
ICOORD ComputeEndFromGradient(const ICOORD& start, double m) {
  if (m > 1.0 || m < -1.0) {
    // dy dominates. Force it to have the opposite sign of start.y() and
    // compute dx based on dy being as large as possible
    int dx = static_cast<int>(floor(MAX_INT16 / m));
    if (dx < 0) ++dx;  // Truncate towards 0.
    if (start.y() > 0) dx = - dx;  // Force dy to be opposite to start.y().
    // Constrain dx so the result fits in an inT16.
    while (start.x() + dx > MAX_INT16 || start.x() + dx < -MAX_INT16)
      dx /= 2;
    if (-1 <= dx && dx <= 1) {
      return ICOORD(start.x(), start.y() + 1);  // Too steep for anything else.
    }
    int y = start.y() + static_cast<int>(floor(dx * m + 0.5));
    ASSERT_HOST(-MAX_INT16 <= y && y <= MAX_INT16);
    return ICOORD(start.x() + dx, y);
  } else {
    // dx dominates. Force it to have the opposite sign of start.x() and
    // compute dy based on dx being as large as possible.
    int dy = static_cast<int>(floor(MAX_INT16 * m));
    if (dy < 0) ++dy;  // Truncate towards 0.
    if (start.x() > 0) dy = - dy;  // Force dx to be opposite to start.x().
    // Constrain dy so the result fits in an inT16.
    while (start.y() + dy > MAX_INT16 || start.y() + dy < -MAX_INT16)
      dy /= 2;
    if (-1 <= dy && dy <= 1) {
      return ICOORD(start.x() + 1, start.y());  // Too flat for anything else.
    }
    int x = start.x() + static_cast<int>(floor(dy / m + 0.5));
    ASSERT_HOST(-MAX_INT16 <= x && x <= MAX_INT16);
    return ICOORD(x, start.y() + dy);
  }
}

// Backwards compatible constrained fit with a supplied gradient.
double DetLineFit::ConstrainedFit(double m, float* c) {
  ICOORDELT_IT it(&pt_list_);
  // Do something sensible with no points.
  if (pt_list_.empty()) {
    *c = 0.0f;
    return 0.0;
  }
  // Count the points and find the first and last kNumEndPoints.
  // Put the ends in a single array to make their use easier later.
  ICOORD* pts[kNumEndPoints * 2];
  int pt_count = 0;
  for (it.mark_cycle_pt(); !it.cycled_list(); it.forward()) {
    if (pt_count < kNumEndPoints) {
      pts[pt_count] = it.data();
      pts[kNumEndPoints + pt_count] = pts[pt_count];
    } else {
      for (int i = 1; i < kNumEndPoints; ++i)
        pts[kNumEndPoints + i - 1] = pts[kNumEndPoints + i];
      pts[kNumEndPoints * 2 - 1] = it.data();
    }
    ++pt_count;
  }
  while (pt_count < kNumEndPoints) {
    pts[pt_count] = NULL;
    pts[kNumEndPoints + pt_count++] = NULL;
  }
  int* distances = new int[pt_count];
  double best_uq = -1.0;
  // Iterate each pair of points and find the best fitting line.
  for (int i = 0; i < kNumEndPoints * 2; ++i) {
    ICOORD* start = pts[i];
    if (start == NULL) continue;
    ICOORD end = ComputeEndFromGradient(*start, m);
    // Compute the upper quartile error from the line.
    double dist = ComputeErrors(*start, end, distances);
    if (dist < best_uq || best_uq < 0.0) {
      best_uq = dist;
      *c = start->y() - start->x() * m;
    }
  }
  delete [] distances;
  // Finally compute the square root to return the true distance.
  return best_uq > 0.0 ? sqrt(best_uq) : best_uq;
}

// Comparator function used by the nth_item funtion.
static int CompareInts(const void *p1, const void *p2) {
  const int* i1 = reinterpret_cast<const int*>(p1);
  const int* i2 = reinterpret_cast<const int*>(p2);

  return *i1 - *i2;
}

// Compute all the cross product distances of the points from the line
// and return the true squared upper quartile distance.
double DetLineFit::ComputeErrors(const ICOORD start, const ICOORD end,
                                 int* distances) {
  ICOORDELT_IT it(&pt_list_);
  ICOORD line_vector = end;
  line_vector -= start;
  // Compute the distance of each point from the line.
  int pt_index = 0;
  for (it.mark_cycle_pt(); !it.cycled_list(); it.forward()) {
    ICOORD pt_vector = *it.data();
    pt_vector -= start;
    // Compute |line_vector||pt_vector|sin(angle between)
    int dist = line_vector * pt_vector;
    if (dist < 0)
      dist = -dist;
    distances[pt_index++] = dist;
  }
  // Now get the upper quartile distance.
  int index = choose_nth_item(3 * pt_index / 4, distances, pt_index,
                              sizeof(distances[0]), CompareInts);
  double dist = distances[index];
  // The true distance is the square root of the dist squared / the
  // squared length of line_vector (which is the dot product with itself)
  // Don't bother with the square root. Just return the square distance.
  return dist * dist / (line_vector % line_vector);
}

}  // namespace tesseract.