Spatial Feature and Mish Loss Function Enhanced YOLOv8n for Efficient Volleyball Detection

To address the challenge of inaccurate volleyball height detection and localization due to complex environments and low image resolution during volleyball bobble assessments, a refined detection algorithm, YOLOv8n-SPD, is proposed. This algorithm enhances the YOLOv8n model by incorporating several key improvements. Firstly, a meticulously labeled dataset of volleyball bobble images is created. The algorithm integrates a Spatial Depth Convolution module (SPD-Conv), which preprocesses images prior to their entry into the neural network. By eliminating the max-pooling layer, redundant pixel information is reduced. Secondly, the activation function Mish is employed instead of SiLU to diminish the interdependence of parameters and enhance the adaptability of the network. Additionally, the DIoU replaces the original CIoU loss function, aiming to better align the predicted and actual bounding boxes. Experimental results show that YOLOv8n-SPD improves the [email protected] metric from 99.3% to 99.4% and the [email protected] metric from 86.5% to 87.6%, with an increase in GFLOPS from 8.1 to 11.5. This algorithm not only enhances accuracy but also improves real-time performance and stability in complex scenarios.