UAV image recognition YOLOv8 detection network improvement

Abstract

An important skill in both civilian and military realms is the ability for UAVs to identify numerous targets. Even if deep learning approaches provide a better answer to the problem, there are still major obstacles in this area, such as variations in target size, form changes, occlusion, and illumination conditions as seen from the drone's viewpoint. The research presents a model for aerial picture identification that is very resilient and performs exceptionally well. These problems are the foundation of the model. The goal of introducing the Bi-PAN-FPN concept is to enhance the neck area. This is done because there is a common issue with aerial pictures where tiny targets might be missed or misdetected. A more sophisticated and comprehensive feature fusion technique may be developed with the goal of fully lowering parameter costs by taking into account and reapplying multiscale features in their entirety. One part of the C2f module is replaced with the GhostblockV2 structure in the benchmark model's backbone. In addition to drastically cutting down on the total number of model parameters, this also reduces the likelihood of data loss as features are sent over long distances. A globally famous dataset is used to analyze and examine the method's performance. To further confirm the efficacy and practicability of the suggested model, extensive ablation studies, comparative analyses, interpretability evaluations, and experiments with bespoke datasets are carried out.. This introduces a fresh approach to multitarget identification by unmanned aerial vehicles (UAVs) utilizing deep learning.