Recognition of Marine Products Based on SA-YOLOv10

Current object detection technology still faces several challenges, including low image clarity, complex backgrounds, and blurred targets when identifying marine products. To address these challenges, this paper proposes an enhanced object detection model based on YOLOv10, named SA-YOLOv10. Firstly, the downsampling method is optimized by replacing SCDown with ADown, which not only reduces model size but also improves object detection accuracy. Secondly, a sliding window mechanism is introduced to restrict attention computation within a window while allowing the model to learn across-window information, significantly reducing computational costs. Finally, through a hierarchical design of downsampling layers, the model is capable of focusing on both global and local information. Experimental results demonstrate that the proposed model achieves 83.6% precision, 78.4% recall, and 85.6% [email protected] in real marine life recognition. Compared to YOLOv10n, these results show improvements of 1.5%, 1.8%, and 1.7%, respectively, with only a 0.3 ms increase in inference time. This indicates that, with a minimal increase in inference time compared to YOLOv10n, the model achieves a substantial improvement in accuracy.