Sina Ghassemi




I received my M.Sc. degree in Telecommunication engineering with a specialization in information theory and signal processing from Politecnico di Torino, Turin, Italy, in 2016. Afterward, I started my Ph.D. program in the area of deep learning with application in image processing and computer vision at Politecnico di Torino. Proceeding with my research as postdoc for another year, I have four years of experience in studying and developing machine learning algorithms addressing problems in different areas such as image processing, remote sensing and time series forecasting. I use my expertise for the current project of automatic personality assessment at Vrije University, organizational psychology department, to develop methodologies for extracting, processing, and analyzing verbal, paraverbal, and nonverbal personality cues from interview videos based on AI techniques.

PhD project

During my Ph.D., I developed an image classification methodology based on deep learning and attention windows to be used in intelligence traffic systems. We showed that an attention mechanism combined with convolutional neural networks can significantly improve vehicle detection and classification accuracy. The proposed approach resulted in state-of-the-art performance in multiple publicly available vehicle classification datasets.
In another project, I proposed a learning and adapting framework relying on neural networks and active learning, to facilitate post-disaster damage assessment over satellite images in collaboration with ITHACA (Information Technology for Humanitarian Assistance, Cooperation and Action) association. In this project, the goal was to develop an AI-based method that can measure and locate the damage caused by natural disaster over satellite images, and also to generalized over any location on the globe. Thus, we proposed domain adaptation techniques that have shown outstanding performance in two publicly available satellite image datasets.
The third project of my Ph.D. involved developing a neural network for cloud screening to be able to be implemented on low-power satellite platforms. As most neural networks contain millions of parameters to perform well, we study methods to optimize the network’s resource consumption such as its memory footprint and inference time while preserving its performance.

Research Interests

  • Data Analysis
  • Machine Learning
  • Signal Processing
  • Computer Vision


Google Scholar

Recent publications

S Ghassemi, A Fiandrotti, G Francini & E Magli (2019) Learning and adapting robust features for satellite image segmentation on heterogeneous data sets. IEEE Transactions on Geoscience and Remote Sensing 57 (9), 6517-6529
S Ghassemi, A Fiandrotti, E Caimotti, G Francini & E Magli (2019) Vehicle joint make and model recognition with multiscale attention windows. Signal Processing: Image Communication 72, 69-79
S Ghassemi & E Magli (2019) Convolutional neural networks for on-board cloud screening. Remote Sensing 11 (12), 1417
S Ghassemi, C Sandu, A Fiandrotti, FG Tonolo, P Boccardo, G Francini & ... (2018) Satellite image segmentation with deep residual architectures for time-critical applications. 2018 26th European Signal Processing Conference (EUSIPCO), 2235-2239
S Ghassemi, A Fiandrotti, E Magli & G Francini (2017) Fine-grained vehicle classification using deep residual networks with multiscale attention windows. Multimedia Signal Processing (MMSP), 2017 IEEE 19th International Workshop …

View full list of publications on Google Scholar