Wang Richeng, Zheng Xiongsheng, Chen Yan
School of Marine Engineering Equipment, Zhejiang Ocean University, Zhoushan, 316002, People's Republic of China.
School of Food and Pharmacy, Zhejiang Ocean University, Zhoushan, 316022, People's Republic of China.
Sci Rep. 2025 Jan 25;15(1):3216. doi: 10.1038/s41598-025-87339-3.
Accurate and rapid segmentation of key parts of frozen tuna, along with precise pose estimation, is crucial for automated processing. However, challenges such as size differences and indistinct features of tuna parts, as well as the complexity of determining fish poses in multi-fish scenarios, hinder this process. To address these issues, this paper introduces TunaVision, a vision model based on YOLOv8 designed for automated tuna processing. TunaVision incorporates enhancements in instance segmentation through YOLOv8m-FusionSeg, improving the segmentation of small and complex targets by increasing channel depth and optimizing feature fusion. Additionally, the YOLOv8s RSF model improves feature extraction speed and accuracy, ensuring each fish is correctly identified and localized before segmentation and pose estimation. Furthermore, TunaVision employs a vector-based approach for pose estimation, utilizing detection and segmentation results to determine fish posture and orientation. Experiments show that YOLOv8m-FusionSeg achieves an mAP@0.5 of 93.3%, while YOLOv8s RSF achieves an mAP@0.5 of 96.1%, with a mean absolute error (MAE) of 1.81 degrees in angle estimation, significantly outperforming other methods. These findings highlight TunaVision's effectiveness in segmenting, detecting, and estimating poses of frozen tuna, offering valuable insights for the development of automated processing systems.
对冷冻金枪鱼的关键部位进行准确、快速的分割，以及精确的姿态估计，对于自动化加工至关重要。然而，金枪鱼各部位的尺寸差异和特征不明显等挑战，以及在多鱼场景中确定鱼的姿态的复杂性，阻碍了这一过程。为了解决这些问题，本文介绍了TunaVision，一种基于YOLOv8设计的用于金枪鱼自动化加工的视觉模型。TunaVision通过YOLOv8m - FusionSeg在实例分割方面进行了改进，通过增加通道深度和优化特征融合来改善小而复杂目标的分割。此外，YOLOv8s RSF模型提高了特征提取速度和准确性，确保在分割和姿态估计之前每条鱼都能被正确识别和定位。此外，TunaVision采用基于向量的方法进行姿态估计，利用检测和分割结果来确定鱼的姿态和方向。实验表明，YOLOv8m - FusionSeg的mAP@0.5达到93.3%，而YOLOv8s RSF的mAP@0.5达到96.1%，角度估计的平均绝对误差（MAE）为1.81度，显著优于其他方法。这些发现突出了TunaVision在分割、检测和估计冷冻金枪鱼姿态方面的有效性，为自动化加工系统的开发提供了有价值的见解。
