Ali Assafiri

Recently, with the help of artificial intelligence and especially through computer vision, new solutions are proposed to identify the roofs of buildings and objects of smaller dimensions and this on satellite images of high resolutions. On the other hand, identifying these same objects on aerial images remains an open research problem. Moreover, whether with satellite images or aerial images, the current solutions based on deep neural networks do not allow to detect several objects of different dimensions at the same time. Thus, the present research project proposes to use convolutional deep neural networks with a multi-scale architecture and a Mask R-CNN in order to identify several objects (building, sheds, pools and cars) simultaneously to various resolutions on aerial images.

To do this, a first study will be done comparing the impact of the use of a pre-trained network compared to an non train network on aerial images. Subsequently, the work will focus on a comparative analysis of the accuracy of the results in the case of using a Fast R-CNN versus a Mask R-CNN. Finally, validations will be made to evaluate the performance of the network developed on Quebec images. In terms of expected results, the four types of objects mentioned above will be extracted from Quebec images in two formats (polygon and bounding box). In addition, the research carried out will lead to obtaining a trained and calibrated deep neural network that will be available for application on other datasets. As for the advancement of knowledge, this research will propose a new multi-detection-resolution approach adapted to the specificities of aerial images.

Direction: Sylvie Daniel

Codirection :