Add kd tree and improve clustering

src/UglyToad.PdfPig.DocumentLayoutAnalysis/ClusteringAlgorithms.cs

@@ -1,11 +1,10 @@
 namespace UglyToad.PdfPig.DocumentLayoutAnalysis
 {
+    using Core;
     using System;
     using System.Collections.Generic;
     using System.Linq;
     using System.Threading.Tasks;
-    using Geometry;
-    using Core;
 
     /// <summary>
     /// Clustering Algorithms.
@@ -16,7 +15,7 @@
         /// Algorithm to group elements using nearest neighbours.
         /// </summary>
         /// <typeparam name="T">Letter, Word, TextLine, etc.</typeparam>
-        /// <param name="elements">List of elements to group.</param>
+        /// <param name="elements">Elements to group.</param>
         /// <param name="distMeasure">The distance measure between two points.</param>
         /// <param name="maxDistanceFunction">The function that determines the maximum distance between two points in the same cluster.</param>
         /// <param name="pivotPoint">The pivot's point to use for pairing, e.g. BottomLeft, TopLeft.</param>
@@ -26,7 +25,7 @@
         /// <param name="maxDegreeOfParallelism">Sets the maximum number of concurrent tasks enabled. 
         /// <para>A positive property value limits the number of concurrent operations to the set value. 
         /// If it is -1, there is no limit on the number of concurrently running operations.</para></param>
-        internal static IEnumerable<HashSet<int>> ClusterNearestNeighbours<T>(List<T> elements,
+        internal static IEnumerable<HashSet<int>> ClusterNearestNeighbours<T>(IReadOnlyList<T> elements,
             Func<PdfPoint, PdfPoint, double> distMeasure,
             Func<T, T, double> maxDistanceFunction,
             Func<T, PdfPoint> pivotPoint, Func<T, PdfPoint> candidatesPoint,
@@ -51,7 +50,7 @@
              *************************************************************************************/
 
             int[] indexes = Enumerable.Repeat(-1, elements.Count).ToArray();
-            var candidatesPoints = elements.Select(candidatesPoint).ToList();
+            KdTree<T> kdTree = new KdTree<T>(elements, candidatesPoint);
 
             ParallelOptions parallelOptions = new ParallelOptions() { MaxDegreeOfParallelism = maxDegreeOfParallelism };
 
@@ -62,92 +61,17 @@
 
                 if (filterPivot(pivot))
                 {
-                    int index = pivot.FindIndexNearest(elements, candidatesPoint, pivotPoint, distMeasure, out double dist);
+                    var paired = kdTree.FindNearestNeighbours(pivot, pivotPoint, distMeasure, out int index, out double dist);
 
-                    if (index != -1)
-                    {
-                        var paired = elements[index];
                     if (filterFinal(pivot, paired) && dist < maxDistanceFunction(pivot, paired))
                     {
                         indexes[e] = index;
                     }
                 }
-                }
             });
 
             // 2. Group indexes
-            var groupedIndexes = GroupIndexes(indexes);
+            return GroupIndexes(indexes);
-
-            return groupedIndexes;
-        }
-
-        /// <summary>
-        /// Algorithm to group elements using nearest neighbours.
-        /// </summary>
-        /// <typeparam name="T">Letter, Word, TextLine, etc.</typeparam>
-        /// <param name="elements">Array of elements to group.</param>
-        /// <param name="distMeasure">The distance measure between two points.</param>
-        /// <param name="maxDistanceFunction">The function that determines the maximum distance between two points in the same cluster.</param>
-        /// <param name="pivotPoint">The pivot's point to use for pairing, e.g. BottomLeft, TopLeft.</param>
-        /// <param name="candidatesPoint">The candidates' point to use for pairing, e.g. BottomLeft, TopLeft.</param>
-        /// <param name="filterPivot">Filter to apply to the pivot point. If false, point will not be paired at all, e.g. is white space.</param>
-        /// <param name="filterFinal">Filter to apply to both the pivot and the paired point. If false, point will not be paired at all, e.g. pivot and paired point have same font.</param>
-        /// <param name="maxDegreeOfParallelism">Sets the maximum number of concurrent tasks enabled. 
-        /// <para>A positive property value limits the number of concurrent operations to the set value. 
-        /// If it is -1, there is no limit on the number of concurrently running operations.</para></param>
-        internal static IEnumerable<HashSet<int>> ClusterNearestNeighbours<T>(T[] elements,
-            Func<PdfPoint, PdfPoint, double> distMeasure,
-            Func<T, T, double> maxDistanceFunction,
-            Func<T, PdfPoint> pivotPoint, Func<T, PdfPoint> candidatesPoint,
-            Func<T, bool> filterPivot, Func<T, T, bool> filterFinal,
-            int maxDegreeOfParallelism)
-        {
-            /*************************************************************************************
-             * Algorithm steps
-             * 1. Find nearest neighbours indexes (done in parallel)
-             *  Iterate every point (pivot) and put its nearest neighbour's index in an array
-             *  e.g. if nearest neighbour of point i is point j, then indexes[i] = j.
-             *  Only conciders a neighbour if it is within the maximum distance. 
-             *  If not within the maximum distance, index will be set to -1.
-             *  Each element has only one connected neighbour.
-             *  NB: Given the possible asymmetry in the relationship, it is possible 
-             *  that if indexes[i] = j then indexes[j] != i.
-             *  
-             * 2. Group indexes
-             *  Group indexes if share neighbours in common - Depth-first search
-             *  e.g. if we have indexes[i] = j, indexes[j] = k, indexes[m] = n and indexes[n] = -1
-             *  (i,j,k) will form a group and (m,n) will form another group.
-             *************************************************************************************/
-
-            int[] indexes = Enumerable.Repeat(-1, elements.Length).ToArray();
-            var candidatesPoints = elements.Select(candidatesPoint).ToList();
-
-            ParallelOptions parallelOptions = new ParallelOptions() { MaxDegreeOfParallelism = maxDegreeOfParallelism };
-
-            // 1. Find nearest neighbours indexes
-            Parallel.For(0, elements.Length, parallelOptions, e =>
-            {
-                var pivot = elements[e];
-
-                if (filterPivot(pivot))
-                {
-                    int index = pivot.FindIndexNearest(elements, candidatesPoint, pivotPoint, distMeasure, out double dist);
-
-                    if (index != -1)
-                    {
-                        var paired = elements[index];
-                        if (filterFinal(pivot, paired) && dist < maxDistanceFunction(pivot, paired))
-                        {
-                            indexes[e] = index;
-                        }
-                    }
-                }
-            });
-
-            // 2. Group indexes
-            var groupedIndexes = GroupIndexes(indexes);
-
-            return groupedIndexes;
         }
 
         /// <summary>
@@ -189,7 +113,6 @@
              *************************************************************************************/
 
             int[] indexes = Enumerable.Repeat(-1, elements.Length).ToArray();
-            var candidatesLines = elements.Select(x => candidatesLine(x)).ToList();
 
             ParallelOptions parallelOptions = new ParallelOptions() { MaxDegreeOfParallelism = maxDegreeOfParallelism };
 
@@ -200,7 +123,7 @@
 
                 if (filterPivot(pivot))
                 {
-                    int index = pivot.FindIndexNearest(elements, candidatesLine, pivotLine, distMeasure, out double dist);
+                    int index = Distances.FindIndexNearest(pivot, elements, candidatesLine, pivotLine, distMeasure, out double dist);
 
                     if (index != -1)
                     {
@@ -214,9 +137,7 @@
             });
 
             // 2. Group indexes
-            var groupedIndexes = GroupIndexes(indexes);
+            return GroupIndexes(indexes);
-
-            return groupedIndexes;
         }
 
         /// <summary>

src/UglyToad.PdfPig.DocumentLayoutAnalysis/Distances.cs

@@ -130,7 +130,7 @@
         /// <param name="pivotPoint"></param>
         /// <param name="distanceMeasure">The distance measure to use.</param>
         /// <param name="distance">The distance between reference point, and its nearest neighbour.</param>
-        internal static int FindIndexNearest<T>(this T element, IReadOnlyList<T> candidates,
+        internal static int FindIndexNearest<T>(T element, IReadOnlyList<T> candidates,
             Func<T, PdfPoint> candidatesPoint, Func<T, PdfPoint> pivotPoint,
             Func<PdfPoint, PdfPoint, double> distanceMeasure, out double distance)
         {
@@ -172,7 +172,7 @@
         /// <param name="pivotLine"></param>
         /// <param name="distanceMeasure">The distance measure between two lines to use.</param>
         /// <param name="distance">The distance between reference line, and its nearest neighbour.</param>
-        internal static int FindIndexNearest<T>(this T element, IReadOnlyList<T> candidates,
+        internal static int FindIndexNearest<T>(T element, IReadOnlyList<T> candidates,
             Func<T, PdfLine> candidatesLine, Func<T, PdfLine> pivotLine,
             Func<PdfLine, PdfLine, double> distanceMeasure, out double distance)
         {

src/UglyToad.PdfPig.DocumentLayoutAnalysis/KdTree.cs

@@ -0,0 +1,230 @@
+namespace UglyToad.PdfPig.DocumentLayoutAnalysis
+{
+    using System;
+    using System.Collections.Generic;
+    using System.Linq;
+    using UglyToad.PdfPig.Core;
+
+    // for kd-tree with line segments, see https://stackoverflow.com/questions/14376679/how-to-represent-line-segments-in-kd-tree 
+
+    internal class KdTree : KdTree<PdfPoint>
+    {
+        public KdTree(PdfPoint[] candidates) : base(candidates, p => p)
+        { }
+
+        public PdfPoint FindNearestNeighbours(PdfPoint pivot, Func<PdfPoint, PdfPoint, double> distanceMeasure, out int index, out double distance)
+        {
+            return FindNearestNeighbours(pivot, p => p, distanceMeasure, out index, out distance);
+        }
+    }
+
+    internal class KdTree<T>
+    {
+        private KdTreeNode<T> Root;
+
+        public KdTree(IReadOnlyList<T> candidates, Func<T, PdfPoint> candidatesPointFunc)
+        {
+            var pointsIndex = Enumerable.Range(0, candidates.Count).Zip(candidates, (e, p) => (e, candidatesPointFunc(p), p)).ToList();
+            if (candidates != null && candidates.Count > 0)
+            {
+                Root = BuildTree(pointsIndex, 0);
+            }
+        }
+
+        private KdTreeNode<T> BuildTree(IReadOnlyList<(int, PdfPoint, T)> P, int depth)
+        {
+            var median = P.Count / 2;
+            if (depth % 2 == 0) // depth is even
+            {
+                P = P.OrderBy(p => p.Item2.X).ToArray();
+            }
+            else
+            {
+                P = P.OrderBy(p => p.Item2.Y).ToArray();
+            }
+
+            // left side
+            var P1 = P.Take(median).ToArray();
+            KdTreeNode<T> vLeft = null;
+            if (P1.Length == 1)
+            {
+                var item = P1[0];
+                vLeft = new KdTreeLeaf<T>(item.Item2, item.Item3, depth, item.Item1);
+            }
+            else if (P1.Length > 1)
+            {
+                vLeft = BuildTree(P1, depth + 1);
+            }
+
+            // right side
+            var P2 = P.Skip(median + 1).ToArray();
+            KdTreeNode<T> vRight = null;
+            if (P2.Length == 1)
+            {
+                var item = P2[0];
+                vRight = new KdTreeLeaf<T>(item.Item2, item.Item3, depth, item.Item1);
+            }
+            else if (P2.Length > 1)
+            {
+                vRight = BuildTree(P2, depth + 1);
+            }
+
+            var medianItem = P[median];
+            return new KdTreeNode<T>(vLeft, vRight, medianItem.Item2, medianItem.Item3, depth, medianItem.Item1);
+        }
+
+        #region NN
+        public T FindNearestNeighbours(T pivot, Func<T, PdfPoint> pivotPointFunc, Func<PdfPoint, PdfPoint, double> distanceMeasure, out int index, out double distance)
+        {
+            var result = FindNearestNeighbours(Root, pivot, pivotPointFunc, distanceMeasure);
+            index = result.Item1.Index;
+            distance = result.Item2.Value;
+            return result.Item1.Element;
+        }
+
+        private static (KdTreeNode<T>, double?) FindNearestNeighbours(KdTreeNode<T> node, T pivot, Func<T, PdfPoint> pivotPointFunc, Func<PdfPoint, PdfPoint, double> distance)
+        {
+            if (node == null)
+            {
+                return (null, null);
+            }
+            else if (node.IsLeaf)
+            {
+                if (node.Element.Equals(pivot))
+                {
+                    return (null, null);
+                }
+                return (node, distance(node.Value, pivotPointFunc(pivot)));
+            }
+            else
+            {
+                var point = pivotPointFunc(pivot);
+                var currentNearestNode = node;
+                var currentDistance = distance(node.Value, point);
+
+                KdTreeNode<T> newNode = null;
+                double? newDist = null;
+
+                var pointValue = node.Depth == 0 ? point.X : point.Y;
+
+                if (pointValue < node.L)
+                {
+                    // start left
+                    (newNode, newDist) = FindNearestNeighbours(node.LeftChild, pivot, pivotPointFunc, distance);
+
+                    if (newDist.HasValue && newDist <= currentDistance && !newNode.Element.Equals(pivot))
+                    {
+                        currentDistance = newDist.Value;
+                        currentNearestNode = newNode;
+                    }
+
+                    if (node.RightChild != null && pointValue + currentDistance >= node.L)
+                    {
+                        (newNode, newDist) = FindNearestNeighbours(node.RightChild, pivot, pivotPointFunc, distance);
+                    }
+                }
+                else
+                {
+                    // start right
+                    (newNode, newDist) = FindNearestNeighbours(node.RightChild, pivot, pivotPointFunc, distance);
+
+                    if (newDist.HasValue && newDist <= currentDistance && !newNode.Element.Equals(pivot))
+                    {
+                        currentDistance = newDist.Value;
+                        currentNearestNode = newNode;
+                    }
+
+                    if (node.LeftChild != null && pointValue - currentDistance <= node.L)
+                    {
+                        (newNode, newDist) = FindNearestNeighbours(node.LeftChild, pivot, pivotPointFunc, distance);
+                    }
+                }
+
+                if (newDist.HasValue && newDist <= currentDistance && !newNode.Element.Equals(pivot))
+                {
+                    currentDistance = newDist.Value;
+                    currentNearestNode = newNode;
+                }
+
+                return (currentNearestNode, currentDistance);
+            }
+        }
+        #endregion
+
+        private class KdTreeLeaf<Q> : KdTreeNode<Q>
+        {
+            public override bool IsLeaf => true;
+
+            public KdTreeLeaf(PdfPoint l, Q element, int depth, int index)
+                : base(null, null, l, element, depth, index)
+            { }
+
+            public override string ToString()
+            {
+                return "Leaf->" + Value.ToString();
+            }
+        }
+
+        private class KdTreeNode<Q>
+        {
+            /// <summary>
+            /// Split value.
+            /// </summary>
+            public double L => Depth == 0 ? Value.X : Value.Y;
+
+            public PdfPoint Value { get; }
+
+            public KdTreeNode<Q> LeftChild { get; internal set; }
+
+            public KdTreeNode<Q> RightChild { get; internal set; }
+
+            public Q Element { get; }
+
+            /// <summary>
+            /// 0 is even (x), 1 is odd (y).
+            /// </summary>
+            public int Depth { get; }
+
+            public virtual bool IsLeaf => false;
+
+            public int Index { get; }
+
+            public KdTreeNode(KdTreeNode<Q> leftChild, KdTreeNode<Q> rightChild, PdfPoint l, Q element, int depth, int index)
+            {
+                LeftChild = leftChild;
+                RightChild = rightChild;
+                Value = l;
+                Element = element;
+                Depth = depth % 2;
+                Index = index;
+            }
+
+            public IEnumerable<KdTreeLeaf<Q>> GetLeaves()
+            {
+                var leafs = new List<KdTreeLeaf<Q>>();
+                RecursiveGetLeaves(LeftChild, ref leafs);
+                RecursiveGetLeaves(RightChild, ref leafs);
+                return leafs;
+            }
+
+            private void RecursiveGetLeaves(KdTreeNode<Q> leaf, ref List<KdTreeLeaf<Q>> leafs)
+            {
+                if (leaf == null) return;
+                if (leaf is KdTreeLeaf<Q> lLeaf)
+                {
+                    leafs.Add(lLeaf);
+                }
+                else
+                {
+                    RecursiveGetLeaves(leaf.LeftChild, ref leafs);
+                    RecursiveGetLeaves(leaf.RightChild, ref leafs);
+                }
+            }
+
+            public override string ToString()
+            {
+                return "Node->" + Value.ToString();
+            }
+        }
+    }
+}

src/UglyToad.PdfPig.DocumentLayoutAnalysis/PageSegmenter/DocstrumBoundingBoxes.cs

@@ -230,7 +230,7 @@
                 return null;
             }
 
-            var closestWordIndex = pointR.FindIndexNearest(wordsWithinAngleBoundDistancePoints, p => p,
+            var closestWordIndex = Distances.FindIndexNearest(pointR, wordsWithinAngleBoundDistancePoints, p => p,
                 p => p, Distances.Euclidean, out _);
 
             if (closestWordIndex < 0 || closestWordIndex >= wordsWithinAngleBoundDistancePoints.Count)