Mahdi Chtourou

✓Creating , Training , Quantizing and optimizing deep learning algorithms to run efficiently on embedded systems: =>Training object detection models and optimizing them to run in real time on embedded systems

- ( Python / TensorFlow ) =>Implementing the pre-processing and the post processing of object detection models and classification models in optimized C code. =>Implementing the state of the art data augmentation to improve the performance of computer vision models.

- ( Python , TensorFlow , GPU optimizations ) =>Studying the latest research paper about object detection in order to improve and update our existing solutions performance. =>Processing different public and private datasets to train custom models and support ST customers with their computer vision use cases. => Created a Deep learning framework to automatically generate high quality object detection datasets and solve the issue of data scarcity. => Supervising interns and guiding them through their internship to successfully reach their project goals:

- Supervised an intern on an object detection project.

- Supervised an intern on a pose estimation project. ✓Representing ST in EEMBC Ultra low power Machine learning work group to benchmark AI solution in terms of accuracy and power consumption: => Implemented efficient and optimized embedded applications to run and benchmark Tiny MLPERF deep learning models on STM32 Boards. ✓Programming languages: { Python , C } ✓Deep Learning Frameworks : { TensorFlow , TFLite , OpenCv , Numpy, Pandas } ✓Embedded software development Frameworks: { STM32CubeMx , IAR EW , Keil , STM32CubeIDE }

Creating an object detection application based on deep learning and deploying it on STM32 board .The work consisted of :

- Creating deep learning models based on convolutional neural networks using Tensorflow.

- Processing Object detection data sets : COCO , OpenImage. ( Pandas , OpenCv )

- Creating the training script . ( Keras , TensorFlow )

- Creating a script to generate the different evaluation metrics in order to evaluate the model performance.

- Quantizing the model using TensorFlow Light and depploying it using STM32CubeMx on STM32 boards ( IN PROGRESS)

Developed and implemented tools to create and preprocess data in order to train a deep learning model "LipNet".Tools are:

- Used technologies: Python , Dlib , OpenCv.

- Used technologies: Python , Google Speech Recognition Api.