Enhancing Vehicle Detection in Aerial Images Through Improved YOLOv8

Introducing a drone image vehicle detection method, it is specifically devised to counter challenges posed by intricate backgrounds, multispectral variances, and numerous minute targets. This approach incorporates adaptive convolutions within the backbone's C2f module, dynamically adjusting kernel sampling points, enhancing feature portrayal for multi-scale and crowded objects. Incorporation of the Biformer attention mechanism within the bi-directional feature pyramid network utilizes sparse computations to dampen complex background effects and protect fragile small target features. We've also developed WloU-S, a unique target boxing loss function integrating angle penalties and a dynamic, non-linear focus system, boosting model flexibility and accelerating learning. Trials confirm significant improvement in detection accuracy over existing methods. Remarkably, compared to the original YOLOv8 framework, our formulation boosts the mean average precision (at 0.5 IoU) by 5.3%, coupled with a parallel increase in recall rate of 5.6%. Our method stands out as an efficient solution for drone image vehicles detection.