Jerusalem 1840-1949 Road Extraction and Alignment: Difference between revisions

From FDHwiki
Jump to navigation Jump to search
 
(164 intermediate revisions by 2 users not shown)
Line 1: Line 1:


==Introduction ==
==Motivation==
In this work, we present a semantic segmentation model based on neural networks for historical city maps. Based on the Jerusalem Old City corpora, we propose a new automatic map alignment method that surpasses the state of the art in terms of flexibility and performance.
[[File:Osm.png |thumb|Figure 1: Jerusalem In Openstreetmap ]]
[[File:Web 1.png|650px|thumb|Figure 2: Web interface screenshot(1)]]
 
 
The creation of large digital databases on urban development is a strategic challenge, which could lead to new discoveries in urban planning, environmental sciences, sociology, economics, and in a considerable number of scientific and social fields. Digital geohistorical data can also be used and valued by cultural institutions. These historical data could also be studied to better understand and optimize the construction of new infrastructures in cities nowadays, and provide humanities scientists with accurate variables that are essential to simulate and analyze urban ecosystems. Now there are many geographic information system platforms that can be directly applied, such as QGIS, ARCGIS, etc. How to digitize and standardize geo-historical data has become the focus of research. We hope to propose a model that can associate geographic historical data with today's digital maps, analyze and study them under the same geographic information platform, same coordinate projection, and the same scale. Eventually, it can eliminate errors caused by scaling, rotation, and the deformation of the map carrier that may exist in historical data and the entire process is automated and efficient.
 
The scale is restricted to Jerusalem in our project. Jerusalem is one of the oldest cities in the world and is considered holy to the three major Abrahamic religions—Judaism, Christianity, and Islam. We did georeferencing among Jerusalem’s historical maps from 1840 to 1949 and the modern map from OpenStreetMap so that the overlaid maps will reveal changes over time and enable map analysis and discovery. We focused on the wall of the Old City as the morphological feature to do georeferencing because the region outside the Old City has seen many new constructions while the Old City has not great changes and the closed polygon of the wall is relatively more consistent than other features like road networks. More specifically, we used dhSegment, a historical document segmentation tool, to extract the wall of the Old City of Jerusalem and proposed an alignment algorithm exploiting the geometrical features of the wall to align the maps.
 
==Deliverables==
- The latitudes and longitudes of the four vertices of the raw maps.
 
- Preprocessed (cropped and resized) patches of the raw maps with images information (including the position of the sub-image in the raw map) stored in a csv file.
 
- 42 1000 * 1000 annotations of the Old City area.


==Motivation==
- Trained segmentation model.
The creation of large digital databases on urban development is a strategic challenge, which could lead to new discoveries in urban planning, environmental sciences, sociology, economics, and in a considerable number of scientific and social fields. Digital geohistorical data can also be used and valued by cultural institutions. These historical data could also be studied to better understand and optimize the construction of new infrastructures in cities nowadays, and provide humanities scientists with accurate variables that are essential to simulate and analyze urban ecosystems. Now there are many geographic information system platforms that can be directly applied, such as QGIS, ARCGIS, etc. how to digitize and standardize geo-historical data has become the focus of research.
 
We hope to propose a model that can associate geographic historical data with today's digital maps, analyze and study them under the same geographic information platform, same coordinate projection, and the same scale. Eliminate errors caused by scaling, rotation, and the deformation of the map carrier that may exist in historical data and the entire process is automated and efficient.
- Code and instructions to reproduce the results on github.
The scale is restricted to Jerusalem in our project. It is one of the oldest cities in the world, and is considered holy to the three major Abrahamic religions—Judaism, Christianity, and Islam. We are going to do georeferencing among Jerusalem’s historical maps from 1840 to 1949 and the modern map from OpenStreetMap. The overlaid maps reveal changes over time and enable map analysis and discovery. We are going to use the wall of the Old City as the feature to georeferenced because the region outside the Old City has seen many new constructions while the Old City has not great changes and the shape of the wall is relatively more consistent than other features like road networks.


==Methodology==
==Methodology==
===Data collection===
===Dataset===
* 126 historical maps of Jerusalem from 1837 to 1938.


126 historical maps of Jerusalem from 1837 to 1938.
* Modern geographical data of Jerusalem from OpenStreetMap.


Modern geographical data of Jerusalem from OpenStreetMap.
===Wall Extraction===


We first used dhSegment to do image segmentation in order to extract the wall polygon from the maps. dhSegment [1] is a generic approach for Historical Document Processing, which relies on a Convolutional Neural Network to predict pixelwise characteristics. The whole process include preprocessing, training, predicting and postprocessing.


===Wall Extraction===
[[File:The network architecture is adapted from ResNet-50.jpg|550px|thumb|center|Figure 2: network architecture]]
dhSegment is a generic approach for Historical Document Processing. It relies on a Convolutional Neural Network to predict pixelwise characteristics.
[[File:The network architecture is adapted from ResNet-50.jpg|550px|thumb|center|Figure 1: network architecture]]
* '''Preprocessing'''
* '''Preprocessing'''
To make the image data fit into the neural networks and to make the old city region more dominant in the image, we cropped and scaled the original image using certain strategy to obtain our training, validation and testing set.
To make the image data fit into the neural networks and to make the old city region more dominant in the image, we cropped and scaled the original image by hand for subsequent uses.
 
* '''Map annotating'''
 
We used Procreate to pixel-wisely label 42 maps according to whether the pixel belong to the old city with different colors (RGB(0, 0, 0) for pixels inside the wall, RGB(255, 0, 255) for pixels outside the wall).  
 
* '''Training'''
* '''Training'''
For training set and validation set images, we pixel-wisely labelled them according to whether they belong to the old city with different colours.
 
There are 25 traing images and 5 validation images now, Learning rate: 5e-5, Batch size: 2, N_epochs: 30, Data augmentation (rotation, scaling and color): True
We divided the 42 annotated patches into 36 and 6 for training and validating respectively, and trained the model with learning rate 5e-5, batch size: 2, N_epochs: 40, and data augmentation like rotation, scaling and color.


{|class="wikitable" style="margin: 1em auto 1em auto;"
{|class="wikitable" style="margin: 1em auto 1em auto;"
|-
|-
|[[File:Error_distribution.png|300px|thumb|center|Figure 5: Original Image]]
|[[File:Original image before training.png|300px|thumb|center|Figure 3: Original Image]]
|[[File:Distplot_of_MSE.png|300px|thumb|center|Figure 6: Labelled Image]]
|[[File:Labelled Image.png|300px|thumb|center|Figure 4: Labelled Image]]
|}
|}


* '''Prediction'''
* '''Prediction'''
By doing prediction on the testing set, we got the predicted old city region and naturally got the contour of it, i.e. the wall.
By doing prediction on the testing set, we got the predicted old city region and naturally got the contour of it, i.e. the wall.
{|class="wikitable" style="margin: 1em auto 1em auto;"
{|class="wikitable" style="margin: 1em auto 1em auto;"
|-
|-
|[[File:Error_distribution.png|300px|thumb|center|Figure 5: Original Image]]
|[[File:Original image.png|300px|thumb|center|Figure 5: Original Image]]
|[[File:Distplot_of_MSE.png|300px|thumb|center|Figure 6: Predicted Binarized Mask]]
|[[File:1887res predict.jpg|300px|thumb|center|Figure 6: Predicted Binarized Mask]]
|}
|}


* '''Postprocessing'''
* '''Postprocessing'''
We remove the minor noises by retain the largest connected region and draw the contour of it. We put the contour on the original image and find that it generally fits well, although there are some minor deviations.
We removed the minor noises by retain the largest connected region and drew the contour of it. We put the contour on the original image and find that it generally fits well, although there are some minor deviations.
[[File:The network architecture is adapted from ResNet-50.jpg|550px|thumb|center|Figure 1: After removing noises and drawing contour]]
 
[[File:1887res contour.jpg|350px|thumb|center|Figure 7: After removing noises and drawing contour]]
 
===Wall Alignment===
 
After obtaining the wall polygons, we try to fit them to the the reference wall polygon from the OpenStreetMap and obtain the coordinates of the source polygon as well as the four vertices of the source image in the coordinate system of the reference image.
* '''Overview'''
 
The overall algorithm proposed is shown as follows. We do scaling, transformation and rotation iteratively for a given iteration numbers in order to get a better alignment result since the initial scaling isn't the accurate enough with the angle difference of the source polygon and reference polygon.
 
 
 
:For '''''i = 0'''''  to '''''iters ''''':


After obtaining the wall contour, we try to fit the wall anchor from the openstreetmap on it and obtain the transformation matrix. The algorithm is as follows.
::​Estimate and do scaling


Firstly, we do an approximate scaling to scale the walls into nearly the same size. Since at the first stage the walls are not in the same direction, the scaling got here is not accurate enough. So after doing the rotation steps, we will scale the size again. Of course we can do the translation, rotation, scaling many times iteratively, but here only once can get a good result.
::For each point pair ''''' P '''''  in selected point pair set:


The translation and rotation step is to find the best translation and rotation angle that maximize the overlapping area of the two walls. Since it’s computationally expensive to search the translation pixel by pixel, and the angle degree by degree, we here narrow down the search space. We limit the search space of translation by only considering the translations that can fit the key point pairs, (P00, P01), (P10, P11), (P20, P21), (P30, P31), together. Here we can also use more key points, like the centroids. For every translation, we fix the key point and rotate the wall every 5 degrees from -30 degrees to 30 degrees and calculate the overlapping area. To calculate the overlapping area, we use FloodFill algorithm to fill the area bounded by the walls respectively and use logical-and operation to obtain the overlapped region and then see the number of pixels that are not zero to be the area. After the iterations, we get the maximized overlapping area and the transformation matrix.
:::Calculate and do transformation


For the wall extraction step, the neural network gives good predictions. But still it does not work well for some maps. For some maps, there is some concavity and convexity. This may owe to underfitting or overfitting. For some maps, the model just cannot make a prediction. This may be because our training set lacks samples that have similar features to these maps.
:::For ''''' k = -6'''''  to ''''' 6 ''''':


We plot the overlapped area ratio w.r.t the year of the map and find that the alignment results of more recent maps are better.
::::Do rotation of angle = ''''' k(iters - i)'''''  around ''''' P '''''


We add the main road from OpenStreetMap to the wall and do alignment. The result also shows that the alignment results of more recent maps are better.
::::Calculate overlapping area


===Wall Alignment===
::::Update maximum overlapping area
* '''Scaling'''
* '''Scaling'''
Firstly, we do an approximate scaling to scale the walls into nearly the same size. Since at the first stage the walls are not in the same direction, the scaling got here is not accurate enough. So after doing the rotation steps, we will scale the size again. Of course we can do the translation, rotation, scaling many times iteratively, but here only once can get a good result.
 
[[File:Web 1.png|650px|thumb|center|Figure 2: An approximate scaling]]
The method to do scaling is shown in the image below.  
 
 
[[File:Scaling new.png|900px|thumb|center|Figure 8: An approximate scaling]]
 


* '''Translation & Rotation'''
* '''Translation & Rotation'''
We traverse the point-pair set to select the reference point-pair, perform translation and rotation to obtain maximized overlapping area of the two polygons.


[[File:Web 1.png|650px|thumb|center|Figure 2: Translation and rotation]]
The goal of the translation and rotation is to obtain optimal translation and rotation to maximize the overlapping area of the two polygons. Since it’s computationally expensive to search for the translation pixel by pixel, and the angle degree by degree, we here narrow down the search space to translations that can fit the key point pairs respectively together (up-left points, up-right points, bottom-left points, bottom-right points and the centroids) and rotation angle set, which starts with [-30, 30] of step 5 and whose range and step will get smaller after every iteration (it's intuitive since after one iteration, we will need more accurate rotation angles).
 
For every translation, we fix the key point together, rotate the wall and calculate the overlapping area. To calculate the overlapping area, we use the FloodFill algorithm to fill the area bounded by the source polygon and reference polygon respectively and use the logical-and operation to obtain the overlapping region. We then count the number of pixels that are not zero to be the overlapping area.
 
After the iterations, we get the maximized overlapping area and the coordinates of the source polygon as well as the four vertices of the source image in the coordinate system of the reference image.
 
 


The translation and rotation step is to find the best translation and rotation angle that maximize the overlapping area of the two walls. Since it’s computationally expensive to search the translation pixel by pixel, and the angle degree by degree, we here narrow down the search space. We limit the search space of translation by only considering the translations that can fit the key point pairs, (P00, P01), (P10, P11), (P20, P21), (P30, P31), together. Here we can also use more key points, like the centroids. For every translation, we fix the key point and rotate the wall every 5 degrees from -30 degrees to 30 degrees and calculate the overlapping area. To calculate the overlapping area, we use FloodFill algorithm to fill the area bounded by the walls respectively and use logical-and operation to obtain the overlapped region and then see the number of pixels that are not zero to be the area. After the iterations, we get the maximized overlapping area and the transformation matrix.
[[File:Translation and Rotation new.png|900px|thumb|center|Figure 9: Translation and rotation]]
 
===Coordinate System Transformation===
 
Since the final goal is to get the latitudes and longitudes in the geographic coordinate system of the four vertices of the raw maps, we finally transformed the coordinates of the image patches in the reference image coordinate system to the latitudes and longitudes of the raw maps through mathematical computations.
 
 
[[File:Coordinate system transformation.png|650px|thumb|center|Figure 10: Coordinate System Transformation]]


==Results==
==Results==
* '''Results – Wall Extraction'''
===Wall Extraction===
For the wall extraction step, the neural network gives good predictions. But still it does not work well for some maps. For some maps, there is some concavity and convexity. This may owe to underfitting or overfitting. For some maps, the model just cannot make a prediction. This may be because our training set lacks samples that have similar features to these maps.  
For the wall extraction step, the neural network generally gives good predictions.  
 
 
{|class="wikitable" style="margin: 1em auto 1em auto;"
|-
| [[File:1837 contour.jpg|250px|thumb|center|Figure 11: Predicted contour of 1837]]
| [[File:1845 contour.jpg|250px|thumb|center|Figure 12: Predicted contour of 1845]]
| [[File:1883 contour.jpg|250px|thumb|center|Figure 13: Predicted contour of 1883]]
| [[File:1921 contour.jpg|250px|thumb|center|Figure 14: Predicted contour of 1921]]
|}
 
 
But for some maps, there are some concavity and convexity in the predictions, but these flaws won't influence the alignment as shown in the next section.
 


{|class="wikitable" style="margin: 1em auto 1em auto;"
{|class="wikitable" style="margin: 1em auto 1em auto;"
|-
|-
| [[File:1625.jpg|300px|thumb|center|Figure 1:Good Predictions]]
| [[File:1853 3 contour.jpg|250px|thumb|center|Figure 15: Predicted contour of 1853]]
| [[File:1587.jpg|200px|thumb|center|Figure 2:Poor Predictions]]
| [[File:1855 contour.jpg|250px|thumb|center|Figure 16: Predicted contour of 1855]]
| [[File:1920 contour.jpg|250px|thumb|center|Figure 17: Predicted contour of 1920]]
| [[File:1929 contour.jpg|250px|thumb|center|Figure 18: Predicted contour of 1929]]
|}
 
===Wall Alignment===
 
For good predictions in the previous segmentation step, the alignment algorithm can give a satisfying alignment.
 
{|class="wikitable" style="margin: 1em auto 1em auto;"
|-
|-
| [[File:1613.jpg|300px|thumb|center|Figure 3:Good Predictions]]
| [[File:1883 contour.jpg|250px|thumb|center|Figure 19: Predicted contour of 1883]]
| [[File:1545.jpg|300px|thumb|center|Figure 4:Poor Predictions]]
| [[File:1921 contour.jpg|250px|thumb|center|Figure 20: Predicted contour of 1921]]
|-
|-
| [[File:583.jpg|300px|thumb|center|Figure 5:Good Predictions]]
| [[File:1883 predict.jpg|250px|thumb|center|Figure 21: Predicted contour of 1883]]
| [[File:573.jpg|300px|thumb|center|Figure 6:Poor Predictions]]
| [[File:1921 predict.jpg|250px|thumb|center|Figure 22: Predicted contour of 1921]]
|}
 
 
The bias generally is caused by the intrinsic difference of the wall in the original map and the reference wall. The maps whose wall polygons have large differences from the nowadays wall polygon were generally produced in the early 19 century. Take the map in 1837 and 1845 for example.
 
 
{|class="wikitable" style="margin: 1em auto 1em auto;"
|-
|-
| [[File:1031.jpg|300px|thumb|center|Figure 7:Good Predictions]]
| [[File:1837 contour.jpg|250px|thumb|center|Figure 23: Predicted contour of 1837]]
| [[File:1030.jpg|300px|thumb|center|Figure 8:Poor Predictions]]
| [[File:1845 contour.jpg|250px|thumb|center|Figure 24: Predicted contour of 1845]]
|-
| [[File:1837 predict.jpg|250px|thumb|center|Figure 25: Alignment result of 1837]]
| [[File:1845 predict.jpg|250px|thumb|center|Figure 26: Alignment result of 1845]]
|}
|}


* '''Results – Wall Alignment'''
We plot the overlapped area ratio w.r.t the year of the map and find that the alignment results of more recent maps are better.
[[File:-50.jpg|550px|thumb|center|Figure 1: the overlapped area ratio]]


We add the main road from OpenStreetMap to the wall and do alignment. The result also shows that the alignment results of more recent maps are better.  
We add the main road from OpenStreetMap to the reference map and align them to the historical maps. The result also shows that the alignment results of more recent maps are better.
 
{|class="wikitable" style="margin: 1em auto 1em auto;"
|-
|[[File:1845 2cropped.png|300px|thumb|center|Figure 27: Alignment result of 1845]]
|[[File:1887.png|300px|thumb|center|Figure 28: Alignment result of 1887]]
|[[File:1915.png|300px|thumb|center|Figure 29: Alignment result of 1915]]
|}
 
 
We plotted the overlapping area ratio with regard to the year of the map and find that the alignment result of early 19 century's maps are not as good as that of later maps, which is consistent with our previous observations.
 
 
[[File:Curve.png|550px|thumb|center|Figure 30: Overlapping area ratio w.r.t the year of map]]
 
 
 
Even for some maps with not that satisfying prediction results, the robustness of the alignment algorithms proposed make the polygons generally align together, thus the coordinates obtained are still reliable.
 
 
{|class="wikitable" style="margin: 1em auto 1em auto;"
{|class="wikitable" style="margin: 1em auto 1em auto;"
|-
|-
|[[File:Error_distribution.png|300px|thumb|center|Figure 5: 1887]]
| [[File:1853 3 contour.jpg|250px|thumb|center|Figure 31: 1853 contour]]
|[[File:Distplot_of_MSE.png|300px|thumb|center|Figure 6: 1915]]
| [[File:1855 contour.jpg|250px|thumb|center|Figure 32: 1855 contour]]
| [[File:1920 contour.jpg|250px|thumb|center|Figure 33: 1920 contour]]
| [[File:1929 contour.jpg|250px|thumb|center|Figure 34: 1929 contour]]
|-
| [[File:1853 3 predict.jpg|250px|thumb|center|Figure 35: 1853 predict]]
| [[File:1855 predict.jpg|250px|thumb|center|Figure 36: 1855 predict]]
| [[File:1920 predict.jpg|250px|thumb|center|Figure 37: 1920 predict]]
| [[File:1929 predict.jpg|250px|thumb|center|Figure 38: 1929 predict]]
|}
|}
===Coordinate System Transformation===
The results are stored in a csv file shown as follows.
[[File:Coordinate transformation.png|900px|thumb|center|Figure 39: Coordinates of the vertices of the raw map in geographic coordinate system]]
==Limitations==
-  Although the alignment algorithm is not sensitive to the minor flaws of the segmentation results, the alignment performance is still subject to the segmentation results. Specifically, when the segmentation gives very bad predictions, the alignment algorithm just can't work.
- The project can only do linear transformation, i.e. scaling, translation and rotation, but it can't do deformation.
- Only the closed polygon of the wall are utilized and segmented.  The morphological feature of road networks and buildings can also be exploited in future works.


==Project Plan and Milestones==
==Project Plan and Milestones==
===Milestones===
*Do survey on semantic segmentation and registration algorithms, get familiar with our dataset and decide what feature to be used for alignment.


*Preprocess the dataset.


*Annotate the maps and use dhSegment for segmentation.


*Design and implement the alignment algorithms.


*Do coordinate system transformation.
*Further improve the robustness of the alignment algorithm and connect the processing steps.
*Sort out the code and write the report.
===Plans===
{|class="wikitable"
{|class="wikitable"
! style="text-align:center;"|Date
! style="text-align:center;"|Date
Line 110: Line 236:
|By Week 4
|By Week 4
|
|
* Brainstorm project ideas, come up with at least one feasible innovative idea.
* Brainstorm project ideas.
* Prepare slides for initial project idea presentation.
* Prepare slides for initial project idea presentation.
| align="center" | ✓
| align="center" | ✓
Line 116: Line 242:
|By Week 6
|By Week 6
|
|
* Study related works about road extraction.
* Study relevant works about road extraction and alignment.
* Determine the methods to be used.
* Use Procreate to produce road-annotated images as the training dataset.
* Use Procreate to get road-tagged images as training dataset.
* Study and apply Projet JADIS to our dataset.
| align="center" | ✓
| align="center" | ✓
|-
|-
|By Week 8
|By Week 8
|
|
* Use Procreate to get wall-tagged images as training dataset.
* Filter and preprocess the dataset for training and testing.
*  
* Try various key point selection algorithms like SIFT and SURF.
*  
* Try various existing point cloud registration algorithms like ICP and TEASER-plusplus.
* Determine to use the wall for alignment.
* Design and develop wall alignment algorithms.
| align="center" | ✓
| align="center" | ✓
|-
|-
|By Week 10
|By Week 10
|
|
*  
* Use Procreate to produce wall-annotated images as the training dataset.
*  
* Train the dhSegment model from scratch and use it for wall extraction.
* Prepare slides for midterm presentation.
* Implement the image postprocessing methods.
| align="center" | --
* Further improve the performance of the alignment algorithm.
* Prepare for the midterm presentation.
| align="center" |
|-
|-
|By Week 11
|By Week 11
|
|
*  
* Produce more wall-annotated images.
*  
* Fine-tune the dhSegment model to get better segmentation results.
*
| align="center" |
| align="center" | --
|-
|-
|By Week 12
|By Week 12
|
|
*  
* Further improve the performance and robustness of the alignment algorithm to make it robust to poorly predicted data.
*  
* Transform the coordinates of the image patches in the reference image coordinate system to the latitudes and longitudes of the raw maps.
| align="center" | --
* Make all the processing steps a pipeline and user-friendly.
| align="center" |
|-
|-
|By Week 13
|By Week 13
|
|
* Sort out the codes and push them to GitHub repository.
* Sort out the codes and push them to GitHub repository.
* Write project report.
* Write the project report.
* Prepare slides for final presentation.
* Prepare slides for final presentation.
| align="center" | --
| align="center" |
|-
|-
|By Week 14
|By Week 14
Line 164: Line 294:


==Github Link==
==Github Link==
https://github.com/mzp0514/JerusalemGeoreferencing


==References==
==References==
[1] Oliveira, Sofia Ares, Benoit Seguin, and Frederic Kaplan. "dhSegment: A generic deep-learning approach for document segmentation." 2018 16th International Conference on Frontiers in Handwriting Recognition (ICFHR). IEEE, 2018.

Latest revision as of 21:42, 20 December 2021

Motivation

Figure 1: Jerusalem In Openstreetmap


The creation of large digital databases on urban development is a strategic challenge, which could lead to new discoveries in urban planning, environmental sciences, sociology, economics, and in a considerable number of scientific and social fields. Digital geohistorical data can also be used and valued by cultural institutions. These historical data could also be studied to better understand and optimize the construction of new infrastructures in cities nowadays, and provide humanities scientists with accurate variables that are essential to simulate and analyze urban ecosystems. Now there are many geographic information system platforms that can be directly applied, such as QGIS, ARCGIS, etc. How to digitize and standardize geo-historical data has become the focus of research. We hope to propose a model that can associate geographic historical data with today's digital maps, analyze and study them under the same geographic information platform, same coordinate projection, and the same scale. Eventually, it can eliminate errors caused by scaling, rotation, and the deformation of the map carrier that may exist in historical data and the entire process is automated and efficient.

The scale is restricted to Jerusalem in our project. Jerusalem is one of the oldest cities in the world and is considered holy to the three major Abrahamic religions—Judaism, Christianity, and Islam. We did georeferencing among Jerusalem’s historical maps from 1840 to 1949 and the modern map from OpenStreetMap so that the overlaid maps will reveal changes over time and enable map analysis and discovery. We focused on the wall of the Old City as the morphological feature to do georeferencing because the region outside the Old City has seen many new constructions while the Old City has not great changes and the closed polygon of the wall is relatively more consistent than other features like road networks. More specifically, we used dhSegment, a historical document segmentation tool, to extract the wall of the Old City of Jerusalem and proposed an alignment algorithm exploiting the geometrical features of the wall to align the maps.

Deliverables

- The latitudes and longitudes of the four vertices of the raw maps.

- Preprocessed (cropped and resized) patches of the raw maps with images information (including the position of the sub-image in the raw map) stored in a csv file.

- 42 1000 * 1000 annotations of the Old City area.

- Trained segmentation model.

- Code and instructions to reproduce the results on github.

Methodology

Dataset

  • 126 historical maps of Jerusalem from 1837 to 1938.
  • Modern geographical data of Jerusalem from OpenStreetMap.

Wall Extraction

We first used dhSegment to do image segmentation in order to extract the wall polygon from the maps. dhSegment [1] is a generic approach for Historical Document Processing, which relies on a Convolutional Neural Network to predict pixelwise characteristics. The whole process include preprocessing, training, predicting and postprocessing.

Figure 2: network architecture
  • Preprocessing

To make the image data fit into the neural networks and to make the old city region more dominant in the image, we cropped and scaled the original image by hand for subsequent uses.

  • Map annotating

We used Procreate to pixel-wisely label 42 maps according to whether the pixel belong to the old city with different colors (RGB(0, 0, 0) for pixels inside the wall, RGB(255, 0, 255) for pixels outside the wall).

  • Training

We divided the 42 annotated patches into 36 and 6 for training and validating respectively, and trained the model with learning rate 5e-5, batch size: 2, N_epochs: 40, and data augmentation like rotation, scaling and color.

Figure 3: Original Image
Figure 4: Labelled Image
  • Prediction

By doing prediction on the testing set, we got the predicted old city region and naturally got the contour of it, i.e. the wall.

Figure 5: Original Image
Figure 6: Predicted Binarized Mask
  • Postprocessing

We removed the minor noises by retain the largest connected region and drew the contour of it. We put the contour on the original image and find that it generally fits well, although there are some minor deviations.

Figure 7: After removing noises and drawing contour

Wall Alignment

After obtaining the wall polygons, we try to fit them to the the reference wall polygon from the OpenStreetMap and obtain the coordinates of the source polygon as well as the four vertices of the source image in the coordinate system of the reference image.

  • Overview

The overall algorithm proposed is shown as follows. We do scaling, transformation and rotation iteratively for a given iteration numbers in order to get a better alignment result since the initial scaling isn't the accurate enough with the angle difference of the source polygon and reference polygon.


For i = 0 to iters :
​Estimate and do scaling
For each point pair P in selected point pair set:
Calculate and do transformation
For k = -6 to 6 :
Do rotation of angle = k(iters - i) around P
Calculate overlapping area
Update maximum overlapping area
  • Scaling

The method to do scaling is shown in the image below.


Figure 8: An approximate scaling


  • Translation & Rotation

The goal of the translation and rotation is to obtain optimal translation and rotation to maximize the overlapping area of the two polygons. Since it’s computationally expensive to search for the translation pixel by pixel, and the angle degree by degree, we here narrow down the search space to translations that can fit the key point pairs respectively together (up-left points, up-right points, bottom-left points, bottom-right points and the centroids) and rotation angle set, which starts with [-30, 30] of step 5 and whose range and step will get smaller after every iteration (it's intuitive since after one iteration, we will need more accurate rotation angles).

For every translation, we fix the key point together, rotate the wall and calculate the overlapping area. To calculate the overlapping area, we use the FloodFill algorithm to fill the area bounded by the source polygon and reference polygon respectively and use the logical-and operation to obtain the overlapping region. We then count the number of pixels that are not zero to be the overlapping area.

After the iterations, we get the maximized overlapping area and the coordinates of the source polygon as well as the four vertices of the source image in the coordinate system of the reference image.


Figure 9: Translation and rotation

Coordinate System Transformation

Since the final goal is to get the latitudes and longitudes in the geographic coordinate system of the four vertices of the raw maps, we finally transformed the coordinates of the image patches in the reference image coordinate system to the latitudes and longitudes of the raw maps through mathematical computations.


Figure 10: Coordinate System Transformation

Results

Wall Extraction

For the wall extraction step, the neural network generally gives good predictions.


Figure 11: Predicted contour of 1837
Figure 12: Predicted contour of 1845
Figure 13: Predicted contour of 1883
Figure 14: Predicted contour of 1921


But for some maps, there are some concavity and convexity in the predictions, but these flaws won't influence the alignment as shown in the next section.


Figure 15: Predicted contour of 1853
Figure 16: Predicted contour of 1855
Figure 17: Predicted contour of 1920
Figure 18: Predicted contour of 1929

Wall Alignment

For good predictions in the previous segmentation step, the alignment algorithm can give a satisfying alignment.

Figure 19: Predicted contour of 1883
Figure 20: Predicted contour of 1921
Figure 21: Predicted contour of 1883
Figure 22: Predicted contour of 1921


The bias generally is caused by the intrinsic difference of the wall in the original map and the reference wall. The maps whose wall polygons have large differences from the nowadays wall polygon were generally produced in the early 19 century. Take the map in 1837 and 1845 for example.


Figure 23: Predicted contour of 1837
Figure 24: Predicted contour of 1845
Figure 25: Alignment result of 1837
Figure 26: Alignment result of 1845


We add the main road from OpenStreetMap to the reference map and align them to the historical maps. The result also shows that the alignment results of more recent maps are better.

Figure 27: Alignment result of 1845
Figure 28: Alignment result of 1887
Figure 29: Alignment result of 1915


We plotted the overlapping area ratio with regard to the year of the map and find that the alignment result of early 19 century's maps are not as good as that of later maps, which is consistent with our previous observations.


Figure 30: Overlapping area ratio w.r.t the year of map


Even for some maps with not that satisfying prediction results, the robustness of the alignment algorithms proposed make the polygons generally align together, thus the coordinates obtained are still reliable.


Figure 31: 1853 contour
Figure 32: 1855 contour
Figure 33: 1920 contour
Figure 34: 1929 contour
Figure 35: 1853 predict
Figure 36: 1855 predict
Figure 37: 1920 predict
Figure 38: 1929 predict

Coordinate System Transformation

The results are stored in a csv file shown as follows.

Figure 39: Coordinates of the vertices of the raw map in geographic coordinate system

Limitations

- Although the alignment algorithm is not sensitive to the minor flaws of the segmentation results, the alignment performance is still subject to the segmentation results. Specifically, when the segmentation gives very bad predictions, the alignment algorithm just can't work.

- The project can only do linear transformation, i.e. scaling, translation and rotation, but it can't do deformation.

- Only the closed polygon of the wall are utilized and segmented. The morphological feature of road networks and buildings can also be exploited in future works.

Project Plan and Milestones

Milestones

  • Do survey on semantic segmentation and registration algorithms, get familiar with our dataset and decide what feature to be used for alignment.
  • Preprocess the dataset.
  • Annotate the maps and use dhSegment for segmentation.
  • Design and implement the alignment algorithms.
  • Do coordinate system transformation.
  • Further improve the robustness of the alignment algorithm and connect the processing steps.
  • Sort out the code and write the report.

Plans

Date Task Completion
By Week 4
  • Brainstorm project ideas.
  • Prepare slides for initial project idea presentation.
By Week 6
  • Study relevant works about road extraction and alignment.
  • Use Procreate to produce road-annotated images as the training dataset.
  • Study and apply Projet JADIS to our dataset.
By Week 8
  • Filter and preprocess the dataset for training and testing.
  • Try various key point selection algorithms like SIFT and SURF.
  • Try various existing point cloud registration algorithms like ICP and TEASER-plusplus.
  • Determine to use the wall for alignment.
  • Design and develop wall alignment algorithms.
By Week 10
  • Use Procreate to produce wall-annotated images as the training dataset.
  • Train the dhSegment model from scratch and use it for wall extraction.
  • Implement the image postprocessing methods.
  • Further improve the performance of the alignment algorithm.
  • Prepare for the midterm presentation.
By Week 11
  • Produce more wall-annotated images.
  • Fine-tune the dhSegment model to get better segmentation results.
By Week 12
  • Further improve the performance and robustness of the alignment algorithm to make it robust to poorly predicted data.
  • Transform the coordinates of the image patches in the reference image coordinate system to the latitudes and longitudes of the raw maps.
  • Make all the processing steps a pipeline and user-friendly.
By Week 13
  • Sort out the codes and push them to GitHub repository.
  • Write the project report.
  • Prepare slides for final presentation.
By Week 14
  • Finish presentation slides and report writing.
  • Presentation rehearsal and final presentation.
--

Github Link

https://github.com/mzp0514/JerusalemGeoreferencing

References

[1] Oliveira, Sofia Ares, Benoit Seguin, and Frederic Kaplan. "dhSegment: A generic deep-learning approach for document segmentation." 2018 16th International Conference on Frontiers in Handwriting Recognition (ICFHR). IEEE, 2018.