Maritime aerial reconnaissance
As part of the DeepGreen project, Thales cortAIx has conducted research and development activities along two main axes: (1) a hardware-aware quantization method (Quantization-Aware Training, QAT) for Vision Transformer (ViT)-based algorithms, building on I-ViT, and (2) incremental learning for transformer-based models such as iCaRL and D3Former, through the implementation of data replay and knowledge distillation mechanisms. Both activities have resulted in functional implementations that are now integrated into Aidge and available as open-source software. To illustrate the second outcome, Thales has developed a demonstrator dedicated to a maritime recognition application based on aerial imagery. It leverages Aidge's graph-based representation capabilities, enabling the model to be updated online without requiring access to the original training data or the initial source code. This approach enhances the confidentiality of the data used to train AI models, which is particularly important in sensitive fields such as defence. In addition, the quantization method integrated into Aidge reduced the memory footprint by 66.7%, while increasing inference performance by approximately 30.6%. This use case highlights Aidge's ability to address key challenges related to data confidentiality, AI deployment, and the maintenance and operational updating of artificial intelligence solutions.
Contributors : Thales
Follow all the project news