Smarter Animal Detection: Upgrading YOLOv8 with Scalable Attention and Loss Optimization

Accurate wildlife observation is important not only in research on ecosystems, but also in maintaining endangered species and preserving diversity. Conventional methods of observation are generally time consuming, labor intensive, and unproductive. In this paper, this study suggests an improved animal detection approach using YOLOv8 to overcome difficulties with moving targets of different sizes and intricate background in wildlife conditions. The new model includes three important updates: Bidirectional Feature Pyramid Network (BiFPN) multiscale feature fusion, dynamic head mechanism with adaptable prediction, and SIoU loss function to enhance the accuracy of bounding box regression. In particular, multi-scale feature integration is enabled efficiently with the help of BiFPN, the dynamic head enhances feature representations, and SIoU improves the accuracy of localization. All these updates significantly enhance detection performance. Experimental results demonstrate that the improved model outperforms the baseline YOLOv8 in both precision and recall, achieving an F1-score of 85.8 %. Moreover, <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\text{mAP} {@} 0.5$</tex> increases by 7.8 percentage points, and <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\text{mAP} {@} 0.5: 0.95$</tex> improves by 7.4 points, validating the effectiveness of the proposed approach.