OBJECT DETECTION MODELS SENSITIVITY & ROBUSTNESS TO SATELLITE-BASED ADVERSARIAL ATTACKS
Résumé
The use of object detection algorithms for the analysis of satellite imagery is increasing in various fields, including en- vironment and defense, as they enable the automatic detec- tion, recognition and localization of targets. Satellite images often exhibit significant variations, including differences in resolution and noise levels between different satellites. Ad- ditional distortions can be caused by factors such as the po- sition of the satellite and the specific area being scanned, re- sulting in changes in tangential distortion, brightness and sat- uration. Depending on the severity, these variations can af- fect the visual clarity of objects in the images and thus impair the effectiveness of object detection algorithms. This study therefore investigates the effects of such fluctuations on the performance of 3 categories of object recognition algorithms - YOLO, FASTER-RCNN and RT-DETR - by applying the principle of adversarial attacks to the inference phase of the algorithms. This experiment makes it possible to uncover the weaknesses of the algorithms and then provides information on how these models could be improved to be more robust to variations in satellite imagery. The case study presented is based on the automatic detection of 3 types of oil and gas in- frastructure: compressor, tank and well in the Permian Basin (USA).
Origine : Fichiers produits par l'(les) auteur(s)