LiteYOLOv8: A Lightweight Model for Real-Time Object Detection in Remote Sensing

As the demand for object detection in remote sensing imagery increases, achieving model lightweighting without sacrificing detection accuracy has become a critical challenge in current research. The YOLO series of convolutional models, known for their high inference speed and accuracy, are widely employed in object detection tasks. Existing optimization techniques, such as Partial Convolution (PConv) and Depthwise Convolution (DWConv), are effective in reducing model parameters and Floating Point Operations (FLOPs), but they often lead to a decline in detection accuracy. To address this, we propose LiteYOLOv8, a lightweight object detection model based on YOLOv8. The model introduces a novel SPC2 module in the Backbone and Neck feature extraction stages, leveraging PConv for lightweight design and enhancing feature extraction through Channel Shuffle. In the Detection Head, DWConv is employed for lightweight classification and regression, while PConv is incorporated to resolve the limitations in cross-channel information interaction. Experimental results demonstrate that LiteYOLOv8 achieves detection accuracy comparable to the original YOLOv8 model on the SSDD, Sar-Aircraftv1.0, and FAIR1M datasets, while it reduces the parameter count by $\mathbf{4 1. 7 \%}$ and the FLOPs by $\mathbf{4 6. 9 \%}$.