Improving the lightweight anti drone target detection algorithm for YOLOv8

To address the challenges posed by the small imaging targets, complex and diverse backgrounds, and strong motion capabilities of unmanned aerial vehicles (UAVs), as well as the issues of large model parameters and slow inference speeds in object detection models, an improved lightweight anti-drone object detection algorithm based on YOLOv8 is proposed. Firstly, the integration of the SASnet backbone network increases the number of feature extractions without introducing redundant information, effectively decreasing the model’s parameter count. Secondly, the incorporation of variable kernel convolutions overcomes the limitations of fixed local windows and sampling shapes in information acquisition. Additionally, the adoption of an efficient C2f_ODConv module dynamically adjusts the convolution kernel’s shape and size adapt to different types of input data, thereby enhancing detection accuracy and precision. The high-precision C2f_SaENet module, combined with the squeeze-and-excitation network and an aggregated multilayer perceptron model, further strengthens the neural network’s ability to capture both channel and global information, thus improving detection accuracy and accelerating model inference speed. In tests conducted on the VisDrone2019 dataset, our model reduced the parameter count by 9.6M, lowered computational complexity by 6.7G, and achieved an FPS of 187, outperforming the original model. On a self-built dataset, our model demonstrated a 3.6% increase in accuracy, a 1.1% increase in recall, a 2.1% improvement in detection precision, along with a 9.6M reduction in parameter count, a 6.7G reduction in computational complexity, and an FPS of 285 compared to the original model. These results indicate that the improved model offers strong detection performance.