Joseph Chazalon

ICDAR 2021 competition on historical map segmentation

Abstract

This paper presents the final results of the ICDAR 2021 Competition on Historical Map Segmentation (MapSeg), encouraging research on a series of historical atlases of Paris, France, drawn at 1/5000 scale between 1894 and 1937. The competition featured three tasks, awarded separately. Task 1 consists in detecting building blocks and was won by the L3IRIS team using a DenseNet-121 network trained in a weakly supervised fashion. This task is evaluated on 3 large images containing hundreds of shapes to detect. Task 2 consists in segmenting map content from the larger map sheet, and was won by the UWB team using a U-Net-like FCN combined with a binarization method to increase detection edge accuracy. Task 3 consists in locating intersection points of geo-referencing lines, and was also won by the UWB team who used a dedicated pipeline combining binarization, line detection with Hough transform, candidate filtering, and template matching for intersection refinement. Tasks 2 and 3 are evaluated on 95 map sheets with complex content. Dataset, evaluation tools and results are available under permissive licensing at https://icdar21-mapseg.github.io/.

Continue reading

Revisiting the Coco panoptic metric to enable visual and qualitative analysis of historical map instance segmentation

Abstract

Segmentation is an important task. It is so important that there exist tens of metrics trying to score and rank segmentation systems. It is so important that each topic has its own metric because their problem is too specific. Does it? What are the fundamental differences with the ZoneMap metric used for page segmentation, the COCO Panoptic metric used in computer vision and metrics used to rank hierarchical segmentations? In this paper, while assessing segmentation accuracy for historical maps, we explain, compare and demystify some the most used segmentation evaluation protocols. In particular, we focus on an alternative view of the COCO Panoptic metric as a classification evaluation; we show its soundness and propose extensions with more “shape-oriented” metrics. Beyond a quantitative metric, this paper aims also at providing qualitative measures through precision-recall maps that enable visualizing the success and the failures of a segmentation method.

Continue reading

Vectorization of historical maps using deep edge filtering and closed shape extraction

Abstract

Maps have been a unique source of knowledge for centuries. Such historical documents provide invaluable information for analyzing the complex spatial transformation of landscapes over important time frames. This is particularly true for urban areas that encompass multiple interleaved research domains (social sciences, economy, etc.). The large amount and significant diversity of map sources call for automatic image processing techniques in order to extract the relevant objects under a vectorial shape. The complexity of maps (text, noise, digitization artifacts, etc.) has hindered the capacity of proposing a versatile and efficient raster-to-vector approaches for decades. We propose a learnable, reproducible, and reusable solution for the automatic transformation of raster maps into vector objects (building blocks, streets, rivers). It is built upon the complementary strength of mathematical morphology and convolutional neural networks through efficient edge filtering. Evenmore, we modify ConnNet and combine with deep edge filtering architecture to make use of pixel connectivity information and built an end-to-end system without requiring any post-processing techniques. In this paper, we focus on the comprehensive benchmark on various architectures on multiple datasets coupled with a novel vectorization step. Our experimental results on a new public dataset using COCO Panoptic metric exhibit very encouraging results confirmed by a qualitative analysis of the success and failure cases of our approach. Code, dataset, results and extra illustrations are freely available at https://github.com/soduco/ICDAR-2021-Vectorization.

Continue reading

Combining deep learning and mathematical morphology for historical map segmentation

Abstract

The digitization of historical maps enables the study of ancient, fragile, unique, and hardly accessible information sources. Main map features can be retrieved and tracked through the time for subsequent thematic analysis. The goal of this work is the vectorization step, i.e., the extraction of vector shapes of the objects of interest from raster images of maps. We are particularly interested in closed shape detection such as buildings, building blocks, gardens, rivers, etc. in order to monitor their temporal evolution. Historical map images present significant pattern recognition challenges. The extraction of closed shapes by using traditional Mathematical Morphology (MM) is highly challenging due to the overlapping of multiple map features and texts. Moreover, state-of-the-art Convolutional Neural Networks (CNN) are perfectly designed for content image filtering but provide no guarantee about closed shape detection. Also, the lack of textural and color information of historical maps makes it hard for CNN to detect shapes that are represented by only their boundaries. Our contribution is a pipeline that combines the strengths of CNN (efficient edge detection and filtering) and MM (guaranteed extraction of closed shapes) in order to achieve such a task. The evaluation of our approach on a public dataset shows its effectiveness for extracting the closed boundaries of objects in historical maps.

Continue reading

Two stages CNN-based segmentation of gliomas, uncertainty quantification and prediction of overall patient survival

Abstract

Continue reading

Using separated inputs for multimodal brain tumor segmentation with 3D U-Net-like architectures

Abstract

The work presented in this paper addresses the MICCAI BraTS 2019 challenge devoted to brain tumor segmentation using mag- netic resonance images. For each task of the challenge, we proposed and submitted for evaluation an original method. For the tumor segmentation task (Task 1), our convolutional neural network is based on a variant of the U-Net architecture of Ronneberger et al. with two modifications: first, we separate the four convolution parts to decorrelate the weights corresponding to each modality, and second, we provide volumes of size 240 * 240 * 3 as inputs in these convolution parts. This way, we profit of the 3D aspect of the input signal, and we do not use the same weights for separate inputs. For the overall survival task (Task 2), we compute explainable features and use a kernel PCA embedding followed by a Random Forest classifier to build a predictor with very few training samples. For the uncertainty estimation task (Task 3), we introduce and compare lightweight methods based on simple principles which can be applied to any segmentation approach. The overall performance of each of our contribution is honorable given the low computational requirements they have both for training and testing.

Continue reading

Segmentation of gliomas and prediction of patient overall survival: A simple and fast procedure

Abstract

In this paper, we propose a fast automatic method that seg- ments glioma without any manual assistance, using a fully convolutional network (FCN) and transfer learning. From this segmentation, we predict the patient overall survival using only the results of the segmentation and a home made atlas. The FCN is the base network of VGG-16, pretrained on ImageNet for natural image classification, and fine tuned with the training dataset of the MICCAI 2018 BraTS Challenge. It relies on the “pseudo-3D” method published at ICIP 2017, which allows for segmenting objects from 2D color images which contain 3D information of MRI volumes. For each n th slice of the volume to segment, we consider three images, corresponding to the (n-1)th, nth, and (n-1)th slices of the original volume. These three gray-level 2D images are assembled to form a 2D RGB color image (one image per channel). This image is the input of the FCN to obtain a 2D segmentation of the n th slice. We process all slices, then stack the results to form the 3D output segmentation. With such a technique, the segmentation of a 3D volume takes only a few seconds. The prediction is based on Random Forests, and has the advantage of not being dependant of the acquisition modality, making it robust to inter-base data.

Continue reading

Segmentation des hyperintensités de la matière blanche en quelques secondes à l’aide d’un réseau de neurones convolutif et de transfert d’apprentissage

Abstract

Dans cet article, nous proposons une méthode automatique et rapide pour segmenter les hyper-intensités de la matière blanche (WMH) dans des images IRM cérébrales 3D, en utilisant un réseau de neurones entièrement convolutif (FCN) et du transfert d’apprentissage. Ce FCN est le réseau neuronal du Visual Geometry Group (VGG) pré-entraîné sur la base ImageNet pour la classification des images naturelles, et affiné avec l’ensemble des données d’entraînement du concours MICCAI WMH. Nous considérons trois images pour chaque coupe du volume à segmenter, provenant des acquisitions en T1, en FLAIR, et le résultat d’un opérateur morphologique appliqué sur le FLAIR, le top-hat, qui met en évidence les petites structures de forte intensité. Ces trois images 2D sont assemblées pour former une image 2D-3 canaux interprétée comme une image en couleurs, ensuite passée au FCN pour obtenir la segmentation 2D de la coupe correspondante. Nous traitons ainsi toutes les coupes pour former la segmentation de sortie 3D. Avec une telle technique, la segmentation de WMH sur un volume cérébral 3D prend environ 10 secondes, pré-traitement compris. Notre technique a été classée 6e sur 20 participants au concours MICCAI WMH.

Continue reading

White matter hyperintensities segmentation in a few seconds using fully convolutional network and transfer learning

Abstract

Continue reading

Extraction of ancient map contents using trees of connected components

Abstract

Ancient maps are an historical and cultural heritage widely recognized as a very important source of information, but exploiting such maps is complicated. In this project, we consider the Linguistic Atlas of France (ALF), built between 1902 and 1910. This cartographical heritage produces firstrate data for dialectological researches. In this paper, we focus on the separation of the content in layers for facilitating the extraction, the analysis, the visualization and the diffusion of the data contained in these ancient linguistic atlases.

Continue reading