Ahmed M.

Senior Machine Learning Researcher at the crossroad of physical systems like computational fluid dynamics and deep neural networks. I work on AI models for numerical simulation. It includes technology watch, design and implementation of efficient deep learning models. 6 years experience.

Geometric Deep Learning, Mesh / Graph Neural Networks, large scale graph mesh datasets,Equivariant Neural Networks, Statistical Signal Processing, Fluid Mechanics, Turbulence Modelling, Navier-Stokes equations, Partial Differential Equations (PDEs), Multi-Scale Representations, Neural Operators, Hybrid Simulation, Uncertainty Quantification, Geometry Morphing

Deep Video Representations, Multiple Aggregation Learning, Kernel Methods, Signal Processing, Graph Neural Networks, Computer Vision

The resurgence of deep neural networks have lead to major breakthroughs and successes mainly in computer vision and natural language processing. In the context of optical character recognition (OCR), there is a considerable growth of document images (e.g scanned invoices) which makes their manual annotation and analysis out of reach. Thus, there is a need to come up with a reliable automatic solution to be able to annotate and extract detailed information in a reasonable time. Scanned documents are usually noisy and their resolution is variable. When the quality of document is poor some characters are not well extracted. The presence of noise and blur make the extraction more challenging. In this internship, we evaluated OCR engines and benchmarked OCR algorithms : Markov models, recurrent and deep neural networks. The existing system that Dhatim uses is based on bi-gram model. It shows poor performance : up to about 45% accuracy. We then proposed an end-to-end trainable neural network which is convolutional recurrent neural network (CRNN). It predicts sequence labels without any pre-segmented inputs or post-process outputs. This deep neural network consists of : convolutional neural network (CNN) which takes the input images, 2 bidirectional long short term memory (Bi LSTM) and a connectionist temporal classification (CTC) layer. We obtained 95.69% accuracy at a sequence level prediction and 98.41 % at a character level prediction.