Deep learning-based object detection technology has been increasingly applied in various scenarios. However, existing object detection networks face challenges in balancing computational complexity and detection accuracy. Neural networks for object detection often struggle to achieve an optimal balance between these two factors, with high computational cost hindering efficiency while maintaining accuracy. In this paper, we propose GS-YOLOv8, an enhanced version of YOLOv8, incorporating two key improvements: a Hybrid Ghost Convolution Network, which integrates lightweight GhostConv modules and C3Ghost bottlenecks to reduce computational redundancy while preserving feature representation capability, and a Dynamic Spatial-Kernel Attention (DSKA) mechanism, implemented through DSK blocks, that adaptively fuses multi-scale features with channel-wise attention. Evaluated on the Tsinghua-Tencent 100K (TT100K) benchmark, GS-YOLOv8 achieves an mAP@0.5 of 74.8%, surpassing the baseline YOLOv8 by 3.7% in accuracy while reducing the number of parameters by 1.3 M. The model demonstrates significant improvements in handling scale variations and complex backgrounds due to its multi-branch adaptive design. The experiments demonstrate that the proposed improved network achieves higher detection accuracy while reducing the model parameters.