Protecting and Mitigating Against Cyber Threats

17Object Detection in Aerial Imagery Using Object Centric Masked Image Modeling (OCMIM)

Aarthi Pulivarthi, Jitta Poojitha Reddy^*, Vanka Eshwar Prabhas, T. Satyanarayana Murthy, Ramesh Babu and Ramu Kuchipudi

Department of Information Technology Chaitanya Bharathi Institute of Technology (A), Hyderabad, Telangana, India

Abstract

To overcome the difficulties in identifying objects in aerial images, the authors of this work introduced Object-Centric Masked Image Modeling (OCMIM), an innovative pretraining technique. The main obstacle is the backdrops’ intricacy and the objects’ different scales, particularly for small objects. The authors developed OCMIM, which consists of two components: Attention-Guided Masking Generation (AGMG) and Object-Centric Data Generation (OCDG), to address this. OCDG combines and crops objects in a creative way to enhance learning about representations by generating object-level image input. Specifically, AGMG concentrates on highly observed object regions when creating guided masks for the pretraining assignment. OCMIM performed exceptionally well when tested on a variety of benchmark datasets, with the NW-PUVHR10 dataset yielding the greatest mean Average Precision (mAP) scores—93.07. Findings of this study demonstrate how OCMIM can greatly improve object detection in the field of remote sensing, which has significance for applications including intelligent security systems, traffic management, and urban development. The study also investigates more ...