Image Analysis and Pattern Recognition

Incremental & interactive Methods - Graph based Methods

Incremental & interactive Methods for Image Analysis and Pattern Recognition

Incremental & Interactive Machine Learning
With Deep learning, Machine Learning is actually a very hot topic with.wide range of applications. Nevertheless, such methods still need a huge amount of training data to define an optimized but static classification model. 
In order to overcome these limitations, my research about system architecture for incremental and interactive learning try to propose more flexible and adaptable classification systems.


• During, the PhD of Gaetan Galisot and until now, we develop (in collaboration between INRAE, iBrain-INSERM, ILIAD3) a new technique for the segmentation of 3D MRI Images based on the combination of local classifiers (HMM or U-Net) dedicated to specific object associated with a strategy for learning spatial relationships between objects (MLWA22, MIA2019, VISAPP2019). It has resulted in an open  plateform SILA3D based on SLICER-3D .
• In 2019, I propose several orientations for Deep Learning research showing the interest of more convergence between works achieved in Document Images Analysis and general Computer vision (ICDAR2019, IDAKS slides).        My Github about that...Before, I have already proposed several original Machine Learning architectures based on the combination of Incremental One class claissifiers (IJDAR2017, DAS2016, PhD Ngoho).
• Biometric applications of Pattern Recognition: AutoID05, ICPR06 
  

Incremental & Interactive Image Segmentation (3D medical Images, LIDAR images, Document Images)

• PhD of Gaetan Galisot: Interactive segmentation of 3D MRI Images based on the combination of local classifiers (MLWA22,,MIA2019, VISAPP2019 - SILA 3D / NeuroGeo Project)
• I also worked on using Machine Learning techniques for Detection of Archeological elements in LIDAR images (CAA2018)
• PhD of Ludovic Paulhac:  Interactive segmentation of 3D ultrasound images: see a video and main publications: [MVA2013] [SIVP2015]    

Document Image Analysis
Most of the works achieved in this field concern low level processing or symbol recognition but very few studies deal with interpretation strategies or knowledge management. So today, the extraction of element of contents in document image is most often achieved in static, predefinite and hard encoded way. We propose to explore new solutions for document image analysis by integrating more interaction between analysis systems and users. To integrate this cooperation, it is necessary to decompose the global problem into two different subproblems:

• The first one concerns the definition of the primal sketch features and primitives in order to have a precise and concise represesentation of the image content whatever its type (drawing, printed document, musical score …). See next points.
• The second one deals with the strategies of analysis of this representation to extract the desired and interesting knowledge from the images Our opinion is that instead of encoding a unique extraction strategy dedicated to a precise set of documents, it is preferable to let the end users define by themselves different strategies depending of the type of document and of the desired results.
• A paper about system architecture [IJDAR2007]
  

Graph based methods for Image Analysis and Pattern Recognition 

In this topic, my past and current works deal with :
• Graph Neural Networks for 3D images segmentation and structural data comparaison (Brain analysis) +  CodeGNN  ANR in progress
• Anytime Graph Matching (PRL2016, PRL2019, PRL2020, PhD of Zeina Abu-Aisheh) - Our method of GED has been included in NetworkX Python library. See Github1 GitHub2
• Optimization and evaluation of Graph Mlatching Algorithms (GED optimization and associated benchmarks (PhD of Zeina Abu-Aisheh)
• Graph based representation of degraded and complex document images
• Graph Embedding techniques for image analysis and Pattern Recognition (PhD of Nicolas Sidère and Muzzamil Luqman) [ICDAR2009] - [LNCS2010] - [PR2013]
• Symbol spotting using graph representation (PhD works of RJ Qureshi) [LNCS2008]
• Symbol recognition using inexact graph-matching methods (PhD works of RJ Qureshi). The proposed approach are evaluated in the context of "Technovision EPEIRES" project.

An evaluation version of VectoGraph.exe can be downloaded here. VectoGraph allow to transform a line drawing into a structural representation using a original vectorisation algorithm. This video can help you to learn how to use Vectograph. Vectograph uses mainly the algorithms described in my PhD manuscript (see below) and in nd a journal paper explaining its origanilty [IJDAR2000]

Chemestry formula recognition
vectograph has also been used by a small company to automatically recognize handwritten chemestry formula The first prototype version of this software is available here.

Analysis and indexation of Historical and Degraded Documents

In this field, research is directed towards pre-processing, layout analysis and character recognition. Major regional (BVH CESR) and national (ANR Navidomass-Madonne - 2006/2010) projects focus on the creation of a complete approach (from scanning to OCR to semantic-based storage) to historical document conversion into digital archives (video describing the context of this work).
Another national project called DIGIDOC supported by the ANR (French National Research Agency) starts in february 2011 dealing with "intelligent digitization of documents".

A demo version of AGORAv2005 prototype for layout analysis and specific content extraction is available here: Demo AGORAv2004 with a [Pdf version of the technical report (in french)]  and a journal paper explaining its originality [IJDAR2007]

A second version of AGORA (v2008) is available (starting from an original idea of Nicholas Journet) : Download it
-Two videos of Agora : A Agora 2005 - A Agora 2008 .

We also develop an interactive Text Transcrption software called RETRO that used the meta-data provided by AGORA to recognize (OCR) the text parts using pattern clustering techniques .A Video demonstration.

Google supports our research work on this topic. We have obtained a Digital Humanities Award in december 2010 to support our activities dealing with old books transcription. See Google Research Blog -

The name of this project is PaRADIIT (Pattern Redundancy Analysis for Document Image Indexation and Transcription.
NEW VERSION of AGORA and RETRO (2014) have been developped during this project. Pleasse visit the associated Web site

On-line handwritting recognition:

I have been interested in on-line handritting recognition during industrial projects:

• The first one was concerning User authentication using on-line handwitten signature (M. Wirotius PhDdealing with biometrics )
• The last one was concerning on-line handwitten gesture recognition using electronic teamboard (for an e-learning application)

A little prototype of gesture recognition is available here with a video demonstration

Remind Performence Evaluation

• Benchmarls for evaluation of Graph Mlatching Algorithms (GED optimization and associated benchmarks (PhD of Zeina Abu-Aisheh)
• Benchmarks for evaluation of 3D images segmentation
• Benchmark for evaluation RGB-D calibration (for Augmented Reality)
• Some images for Document image analysis :

             

A summary of my reseach works until 2008 . . .

You can download my HDR manuscript (in french) entitled "Propositions pour la representation et l'analyse de documents numériques" defended in november,the 8th 2006.

Some results of my works -open source and open data) are available on the RFAI Resources web page ==> wiki RFAI

Others very old works . . .

Last modif. : 2/01/2013.