基于关键点检测的南美白对虾体尺估测方法

Whiteleg Shrimp Body Length Estimation Approach Based on Keypoint Detection

摘要:

南美白对虾体尺的准确测量在工厂化养殖中具有重要意义,然而在实际养殖环境中,观测台上常存在残余饲料干扰、虾只密集聚集、相互遮挡与粘连等问题,导致现有计算机视觉算法面临关键点定位不准、饲料误识别、重复检测以及遮挡下虾只无法完整识别等挑战。 为解决上述问题,提出了一种对虾关键点检测模型 YOLO 11-Shrimp。 首先采用 Diverse branch block(DBB)模块优化主干网络中的 Conv 模块,以显著增强特征表达能力;其次,融入坐标注意力机制,进一步提升关键点定位精度;最后,在检测头部分引入 C3k2_OREPA 模块,实现高效特征融合与表达。 实验结果表明,与基准模型 YOLO 11-Pose 相比,YOLO 11 Shrimp 在PCK0. 1 指标上提升了 10. 35 个百分点,在 PCK0. 2 指标上提升了 7. 82 个百分点;同时,饲料误检现象显著减少,密集分布场景下的重复检测问题得到有效缓解,虾体关键点偏移幅度明显降低,边界框置信度得分更加可靠。 此外,设计了一种对虾体尺自动化估测系统。 在虾只因遮挡无法完整检测时,能够自动根据可见的局部部位估测体尺。 研究结果可为南美白对虾智能化养殖提供方法依据和参考。

Abstract:

Precise measurement of body length and morphological dimensions in whiteleg shrimp is critical for intelligent industrial aquaculture. However, practical farming environments present significant challenges, including interference from residual feed pellets, high stocking densities, mutual occlusion, and individual adhesion. These factors often lead to inaccurate keypoint localization, false positives from feed pellets, duplicate detection, and incomplete recognition of occluded individuals. To overcome these limitations, YOLO 11-Shrimp, an enhanced model for shrimp keypoint detection was proposed. Specifically, YOLO 11-Shrimp incorporated three key architectural enhancements. Firstly, standard convolutional modules within the backbone were replaced by diverse branch block (DBB)modules to enhance multi-scale feature representation. Secondly, the coordinate attention (CA)mechanism was integrated to capture long-range spatial dependencies, thereby refining keypoint localization precision. Finally, the C3k2_OREPA module was introduced into the detection head to substitute the original C3k2 structure, facilitating more robust feature fusion and representational efficiency. Experimental results demonstrated that, compared with the baseline YOLO 11-Pose model, YOLO 11-Shrimp improved PCK0. 1 by 10. 35 percentage points and PCK0. 2 by 7. 82 percentage points. Meanwhile, false detection caused by residual feed pellets were significantly reduced, and duplicate detection under dense scenes were suppressed. Also, the shrimp keypoint shift was markedly decreased, and bounding box confidence scores became more reliable. Furthermore, an automated shrimp body length estimation system was designed. When a shrimp cannot be fully detected due to occlusion, the system can automatically estimate the body length based on visible parts. The research results can provide a methodological framework for precision monitoring in smart aquaculture for whiteleg shrimp.