Enhancing the Explainability of the YOLOv8 Model in Cattle Face Detection: A Numerical Integrated Gradients-Based Approach

In recent years, deep learning-based object detection models, such as the YOLO series, have been applied to cattle face detection. However, the complexity of actual pasture environments (such as occlusion, lighting variations, and long distances between the camera and target) reduces detection precision, leading to false and missed detections. However the “black-box” nature of deep learning models makes these errors inexplicable, thereby undermining model's reliability in practical applications. To enhance the explainability of the YOLOv8 model in cattle face detection under complex environments, this paper proposes a numerical integrated gradients explainability algorithm based on integrated gradients. The proposed method designs a gradient computation strategy tailored for object detection and incorporates numerical analysis to achieve both qualitative and quantitative feature attribution for specific detection boxes. By leveraging visualization and statistical analysis, the method explains the degree of dependence of YOLOv8's decision outputs on key cattle face features. Experimental results demonstrate that NumIG-YOLO exhibits strong explanation performance in cattle face detection using YOLOv8, and it generates more precise heatmaps than Grad-CAM and D-RISE explainability algorithms. Additionally, statistical analysis on the test set reveals that YOLOv8 relies more heavily on the eyes and mouth regions when detecting cattle faces.