YOLO-CG: an enhanced YOLOv8 model with multi-scale feature fusion and adaptive loss for remote sensing images

Object detection is essential for many remote sensing applications. Traditional object detection methods struggle with large scale variations and complex backgrounds in high-resolution remote sensing images. To address these challenges, we introduce YOLO-CG, an enhanced YOLOv8-derived framework for remote sensing object detection. Our design fuses the Spatial Pyramid Pooling and Cross Stage Partial Channel (SPPCSPC) module with a Global Attention Mechanism (GAM), bolstering both multi-scale feature aggregation and contextual representation. Specifically, SPPCSPC replaces the original Spatial Pyramid Pooling Fast (SPPF) to retain richer spatial information and improve feature fusion, while GAM is applied after each Convolution-to-Feature (C2F) module in the neck to emphasize salient spatial and channel-wise features. Furthermore, the standard Intersection over Union (IoU) loss is replaced with Wise-IoU (WIoU), a dynamic non-monotonic loss that adapts to object-specific characteristics and improves detection across different scales and orientations. Experimental results on benchmark datasets demonstrate that YOLO-CG achieveds higher detection accuracy than baseline methods. Experiment results show that the YOLO-CG model attained an AP50 of 94%on the Pose Bowl: Detection Track dataset outperforming YOLOv8 by 12% and YOLO-SE by 2%. IThe model’s robustness and efficiency support sustainable development applications by providing reliable data for informed decision-making and plicy implementation.

Actions (login required)

View Item