The paper introduces YOLO-SK, an improved version of the lightweight YOLOv5s model, designed to enhance multiscale object detection accuracy while maintaining low computational complexity. The key contributions of YOLO-SK include: 1. **Weighted Dense Feature Fusion Network (WD_FPN-PAN)**: This network dynamically fuses features at different scales using learnable parameters and cross-layer fusion capabilities, enhancing the richness of effective information in the fused feature maps. 2. **SK Attention Prediction Head (SK_PH)**: This head broadens the model's receptive field and sharpens its focus on target characteristics, improving the extraction of target information from the fused feature maps. 3. **SiOu Loss Function**: This loss function reduces positioning errors between predicted and ground truth boxes, ensuring more accurate detection of targets of different scales. 4. **Ghost Convolution**: Replacing conventional convolutions with Ghost Convolution reduces computational complexity and parameter count, maintaining model accuracy while keeping the model lightweight. Experiments on the PASCAL VOC 2007 and 2012 datasets show that YOLO-SK achieves a 2.6% increase in mAP@.5 and a 4.8% increase in mAP@.5:95 compared to YOLOv5s, while keeping the same level of model complexity. The proposed improvements effectively enhance the model's detection accuracy, particularly for multiscale targets, and demonstrate better convergence performance, making it suitable for low-performance devices.The paper introduces YOLO-SK, an improved version of the lightweight YOLOv5s model, designed to enhance multiscale object detection accuracy while maintaining low computational complexity. The key contributions of YOLO-SK include: 1. **Weighted Dense Feature Fusion Network (WD_FPN-PAN)**: This network dynamically fuses features at different scales using learnable parameters and cross-layer fusion capabilities, enhancing the richness of effective information in the fused feature maps. 2. **SK Attention Prediction Head (SK_PH)**: This head broadens the model's receptive field and sharpens its focus on target characteristics, improving the extraction of target information from the fused feature maps. 3. **SiOu Loss Function**: This loss function reduces positioning errors between predicted and ground truth boxes, ensuring more accurate detection of targets of different scales. 4. **Ghost Convolution**: Replacing conventional convolutions with Ghost Convolution reduces computational complexity and parameter count, maintaining model accuracy while keeping the model lightweight. Experiments on the PASCAL VOC 2007 and 2012 datasets show that YOLO-SK achieves a 2.6% increase in mAP@.5 and a 4.8% increase in mAP@.5:95 compared to YOLOv5s, while keeping the same level of model complexity. The proposed improvements effectively enhance the model's detection accuracy, particularly for multiscale targets, and demonstrate better convergence performance, making it suitable for low-performance devices.