An efficient advanced YOLOv8 framework for sperm motility detection

Sperm motility detection is crucial for reproductive health research, which is closely related to human development. Due to the small size and fast movement of sperm, coupled with the complex background, traditional image processing methods can hardly obtain accurate detection results.

To address these issues, an advanced YOLOv8-based sperm detection model is proposed, named YOLOv8-STA. First, the SPD_Conv structure is integrated into the backbone network to improve the ability for extracting fine-grained information about sperm targets. Secondly, in order to construct effective relationships between input features while preserving more original image information, the partial C2F modules in the backbone and neck networks of the YOLOv8 model are replaced with C2F-Triplet modules, which can capture interactions across dimensions, build channel, and spatial attentions with minimal parameters. Finally, a new 104 $\times$ 104 pixel small object detection head is added specifically in the detection layer to construct advanced feature maps.

Experimental results on the public EVISAN dataset show that the improved YOLOv8-STA model can increase the mAP@0.5 by 4.2% and mAP@0.5:0.95 by 5.4%, compared with the baseline YOLOv8 model.

These experimental results verify the effectiveness and superiority of the proposed method in sperm motility detection.