Projects

DocPRESERV: Preserving & Processing Historical Document Images with Artificial Intelligence

The main idea of the project is to provide AI-based solutions to make digital historical handwritten documents more accessible. The challenge lies in the fact that digitized documents are of very complex structure and exhibit variable writing styles due to different authors or ages amongst other issues. The project will push innovation in historical handwritten document analysis and recognition and will develop innovative tools to enhance the capabilities of historical document processing and retrieval. 

Partner : Blekinge Institute of Technology in Karlskrona (Sweden) 

SmartSURV: Vidéo Protection Urbaine dans les villes intelligentes 

Intelligent video surveillance of urban areas by means of artificial vision is a real challenge in smart cities. It consists in exploiting the images from one or more cameras to extract an interpretation that allows to diagnose a situation and to report it. The SmartSURV project aims to develop an intelligent video protection platform to improve security in urban areas. The platform automatically analyzes a huge amount of multimedia data to combat risky road behaviors such as speeding or traffic light violations, provide parking surveillance and access control to combat threats such as vehicle theft, assaults and violence. 

PEJC project, financed by Tunisian ministry of higher education and scientific research 

READ: Reconnaissance et Extraction d’information d’imAges de Documents 

The objective of this project is to develop a system for extracting relevant information (names, dates, places...) from images of digitized civil status records. The corpera is a mix of printed and handwritten documents with different formats (Birth, Marriage and Death) and different qualities (noise, quality of the writing, degradation...). We combine the techniques of OCR/ICR and NLP to extract the named entities form document images.

Partner Studia company 

The DECRYPT Project: Automatic Decryption of Historical Manuscripts

Within the DECRYPT project, we release resources and tools with open access to facilitate research in historical cryptology, allowing collection, analysis and decryption of historical ciphertexts. Resources are collections of encrypted sources, and historical texts and language models. The tools facilitate the processing of the encrypted sources from transcription to decryption incl. cryptanalysis.

Partner : CVC Barcelone, Uppsala University, Sweden

More Infos...

SOS-OLIVIER: Suivi multi-apprOche de la Santé de l’OLIVIER en Tunisie par télédétection optique 

The analysis of satellite and aerial images is one of the fundamental topics in the field of remote sensing. In recent years, technological advances have increased the availability of large-scale images. Earth observation satellites provide multispectral and panchromatic data with different spatial, spectral, temporal, and radiometric resolutions. We propose different deep learning based techniques for image fusion which consist in combining several images, taken by one or several satellites during the same or different time. The goal is to propose solutions to merge the best characteristics of each image into a single product. We propose innovative solutions to the problems of pansharpening and multi-source multi-temporal fusion.

PRF project financed by the Tunisian ministry of higher education and scientific research.

IntelliDOC: Intelligent Document processing with Transformer 

The objective of this project is to develop models based on the Transformer architecture for information extraction from images of handwritten documents. We are first interested in optimizing the Transformer architecture to generalize it to the processing of long sequences (whole documents or paragraphs).  This first objective will be experimentally validated on public databases for the recognition and extraction of named entities from images of handwritten documents.  The second objective of the project concerns the learning of Transformer models. These models usually require large amounts of labeled data to converge. We intend to propose a self-supervised learning technique to learn these models on unlabeled document corpora. Then this first model could be refined by transfer learning (in supervised mode) using reduced quantities of labeled data. This two-step learning process will then be validated on several applications in document image processing such as handwriting recognition, named entity recognition, classification...

Partner: Instadeep company 

REIVE: Reconnaissance de l'identité de Véhicule Tunisiens

REIVE project provides robust actionable vehicle knowledge and critical insights for safe city, law enforcement, access control, speed detection, light violation... The need to recognize the identity of vehicles has increased due to the increase of vehicle number. To identify vehicles, numerous parameters are recognised using Artificial Intelligence techniques such as license plate number, vehicle logo, vehicle mark, model and color. All parameters are combined to improve the vehicle identity recognition process and help check correlation with data stored on police and homeland security databases. 

VRR project financed by the Tunisian ministry of higher education and scientific research.  

ADIP: Automatic Detection of in-process pieces using robot control

ADIP project aims at developing a vision-based robotics system for the automatic assembly of pipe fittings. Different parts can be randomly located across the conveyor. The developped system allow robot control by properly detecting in-process pieces using computer vision and image processing techniques . 

Partner: Sopal company 

SMART-CASTLE : SMART Contrôle Aérien, Satellitaire et Terrestre de L'activité terroristE 

The project consists in proposing a re-identification system of suspicious persons from the analysis of images captured by a fixed camera or carried by a drone.  We propose a new approach based on advanced deep learning techniques. It is composed of 3 complementary modules: a person re-identification module based on the person's appearance, a person attribute prediction module (clothing, accessories, color of clothing...) and an aggregation module that allows to reclassify the hypotheses returned by the appearance model from the set of predicted attributes. 

PRF project financed by the Tunisian ministry of higher education and scientific research.