Page detection.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "import cv2\n",
    "import numpy as np\n",
    "import sys\n",
    "\n",
    "image = cv2.cvtColor(cv2.imread('files/form8.jpg'), cv2.COLOR_BGR2RGB)\n",
    "\n",
    "def resize(img, height=800):\n",
    "    \"\"\" Resize image to given height \"\"\"\n",
    "    rat = height / img.shape[0]\n",
    "    return cv2.resize(img, (int(rat * img.shape[1]), height))\n",
    "\n",
    "# Resize and convert to grayscale\n",
    "img = cv2.cvtColor(resize(image), cv2.COLOR_BGR2GRAY)\n",
    "\n",
    "# Bilateral filter preserv edges\n",
    "img = cv2.bilateralFilter(img, 9, 75, 75)\n",
    "\n",
    "# Create black and white image based on adaptive threshold\n",
    "img = cv2.adaptiveThreshold(img, 255, cv2.ADAPTIVE_THRESH_GAUSSIAN_C, cv2.THRESH_BINARY, 115, 4)\n",
    "\n",
    "# Median filter clears small details\n",
    "img = cv2.medianBlur(img, 11)\n",
    "\n",
    "# Add black border in case that page is touching an image border\n",
    "img = cv2.copyMakeBorder(img, 5, 5, 5, 5, cv2.BORDER_CONSTANT, value=[0, 0, 0])\n",
    "\n",
    "edges = cv2.Canny(img, 200, 250)\n",
    "\n",
    "# Getting contours  \n",
    "im2, contours, hierarchy = cv2.findContours(edges, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)\n",
    "\n",
    "# Finding contour of biggest rectangle\n",
    "# Otherwise return corners of original image\n",
    "# Don't forget on our 5px border!\n",
    "height = edges.shape[0]\n",
    "width = edges.shape[1]\n",
    "MAX_COUNTOUR_AREA = (width - 10) * (height - 10)\n",
    "\n",
    "# Page fill at least half of image, then saving max area found\n",
    "maxAreaFound = MAX_COUNTOUR_AREA * 0.5\n",
    "\n",
    "# Saving page contour\n",
    "pageContour = np.array([[5, 5], [5, height-5], [width-5, height-5], [width-5, 5]])\n",
    "\n",
    "# Go through all contours\n",
    "for cnt in contours:\n",
    "    # Simplify contour\n",
    "    perimeter = cv2.arcLength(cnt, True)\n",
    "    approx = cv2.approxPolyDP(cnt, 0.03 * perimeter, True)\n",
    "\n",
    "    # Page has 4 corners and it is convex\n",
    "    # Page area must be bigger than maxAreaFound \n",
    "    if (len(approx) == 4 and\n",
    "            cv2.isContourConvex(approx) and\n",
    "            maxAreaFound < cv2.contourArea(approx) < MAX_COUNTOUR_AREA):\n",
    "\n",
    "        maxAreaFound = cv2.contourArea(approx)\n",
    "        pageContour = approx\n",
    "\n",
    "# Result in pageConoutr (numpy array of 4 points):\n",
    "\n",
    "\n",
    "def fourCornersSort(pts):\n",
    "    \"\"\" Sort corners: top-left, bot-left, bot-right, top-right \"\"\"\n",
    "    # Difference and sum of x and y value\n",
    "    # Inspired by http://www.pyimagesearch.com\n",
    "    diff = np.diff(pts, axis=1)\n",
    "    summ = pts.sum(axis=1)\n",
    "    \n",
    "    # Top-left point has smallest sum...\n",
    "    # np.argmin() returns INDEX of min\n",
    "    return np.array([pts[np.argmin(summ)],\n",
    "                     pts[np.argmax(diff)],\n",
    "                     pts[np.argmax(summ)],\n",
    "                     pts[np.argmin(diff)]])\n",
    "\n",
    "\n",
    "def contourOffset(cnt, offset):\n",
    "    \"\"\" Offset contour, by 5px border \"\"\"\n",
    "    # Matrix addition\n",
    "    cnt += offset\n",
    "    \n",
    "    # if value < 0 => replace it by 0\n",
    "    cnt[cnt < 0] = 0\n",
    "    return cnt\n",
    "\n",
    "\n",
    "# Sort and offset corners\n",
    "pageContour = fourCornersSort(pageContour[:, 0])\n",
    "pageContour = contourOffset(pageContour, (-5, -5))\n",
    "\n",
    "# Recalculate to original scale - start Points\n",
    "sPoints = pageContour.dot(image.shape[0] / 800)\n",
    "  \n",
    "# Using Euclidean distance\n",
    "# Calculate maximum height (maximal length of vertical edges) and width\n",
    "height = max(np.linalg.norm(sPoints[0] - sPoints[1]),\n",
    "             np.linalg.norm(sPoints[2] - sPoints[3]))\n",
    "width = max(np.linalg.norm(sPoints[1] - sPoints[2]),\n",
    "             np.linalg.norm(sPoints[3] - sPoints[0]))\n",
    "\n",
    "# Create target points\n",
    "tPoints = np.array([[0, 0],\n",
    "                    [0, height],\n",
    "                    [width, height],\n",
    "                    [width, 0]], np.float32)\n",
    "\n",
    "# getPerspectiveTransform() needs float32\n",
    "if sPoints.dtype != np.float32:\n",
    "    sPoints = sPoints.astype(np.float32)\n",
    "\n",
    "# Wraping perspective\n",
    "M = cv2.getPerspectiveTransform(sPoints, tPoints) \n",
    "newImage = cv2.warpPerspective(image, M, (int(width), int(height)))\n",
    "\n",
    "# Saving the result. Yay! (don't forget to convert colors bact to BGR)\n",
    "cv2.imwrite(\"files/result-page-detection.jpg\", newImage)"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.6.3"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}