Wang Jinwei, Du Zhiguo, Wen Bin, Wu Zhihui, Lin Xudong
College of Mathematics Informatics, South China Agricultural University, Guangzhou 510642, PR China.
Poult Sci. 2025 Aug;104(8):105221. doi: 10.1016/j.psj.2025.105221. Epub 2025 May 2.
The behavior of lion-head goose has a significant impact on their health status, activity levels, and productivity. It is therefore important to monitor the behavior of lion-head geese to enhance their health status, reproductive performance, and overall productivity. However, there is currently no specific behavioral recognition method for lion-head goose, which presents a significant challenge in quickly and effectively identifying various behaviors. To address this issue, this study proposes a model called EML-SlowFast, which is an improvement based on SlowFast. The model is capable of distinguishing five basic behaviors of lion-head goose: feeding, resting, preening, standing, and walking. The Efficient Channel Attention Bottleneck (ECAbneck) module and the Large Kernel Global-Local Feature Extraction (LGLE) module are designed and incorporated into the model. By combining and filtering channel information, the ECAbneck module enhances the model's ability to extract static characteristics from lion-head goose, increasing the accuracy of behavior recognition. The LGLE module captures temporal dependencies in lion-head goose behavior by integrating and extracting local and global features, thereby reinforcing the model's ability to model long-term temporal characteristics and further increasing accuracy. The experiment results showed that the average F1 score, average Precision, Accuracy, and average Recall of the EML-SlowFast model were 92.06 %, 91.60 %, 91.85 %, and 92.78 %, respectively, reflecting improvements of 4.03 %, 3.79 %, 4.14 %, and 4.45 % over the corresponding metrics of the SlowFast model. Furthermore, the FLOPs of the EML-SlowFast model was 10.807 G, which was a reduction of 7.358 G compared to the SlowFast model. Compared to commonly used behavior recognition models, the EML-SlowFast model has effective recognition of lion-head goose behaviors while maintaining low computational complexity, which is beneficial for deployment and use in scenarios with low computational resources. The EML-SlowFast model can rapidly and accurately recognize lion-head goose behaviors, providing a valuable reference for precision farming, reproduction, and health welfare monitoring of lion-head goose.
狮头鹅的行为对其健康状况、活动水平和生产性能有着重大影响。因此，监测狮头鹅的行为对于提高其健康状况、繁殖性能和整体生产性能至关重要。然而，目前尚无针对狮头鹅的特定行为识别方法，这在快速有效地识别各种行为方面构成了重大挑战。为解决这一问题，本研究提出了一种名为EML-SlowFast的模型，它是在SlowFast基础上的改进。该模型能够区分狮头鹅的五种基本行为：进食、休息、梳理羽毛、站立和行走。设计并将高效通道注意力瓶颈（ECAbneck）模块和大核全局-局部特征提取（LGLE）模块纳入该模型。通过组合和过滤通道信息，ECAbneck模块增强了模型从狮头鹅中提取静态特征的能力，提高了行为识别的准确性。LGLE模块通过整合和提取局部和全局特征来捕捉狮头鹅行为中的时间依赖性，从而增强了模型对长期时间特征进行建模的能力，并进一步提高了准确性。实验结果表明，EML-SlowFast模型的平均F1分数、平均精确率、准确率和平均召回率分别为92.06%、91.60%、91.85%和92.78%，与SlowFast模型的相应指标相比分别提高了4.03%、3.79%、4.14%和4.45%。此外，EML-SlowFast模型的浮点运算次数（FLOPs）为10.807 G，与SlowFast模型相比减少了7.358 G。与常用的行为识别模型相比，EML-SlowFast模型在保持低计算复杂度的同时，能够有效识别狮头鹅的行为，这有利于在计算资源较低的场景中进行部署和使用。EML-SlowFast模型能够快速准确地识别狮头鹅的行为，为狮头鹅的精准养殖、繁殖和健康福利监测提供了有价值的参考。
