文献类型： 外文期刊

第一作者： Zhao, Shida

作者： Zhao, Shida;Bai, Zongchun;Huo, Lianfei;Han, GuoFeng;Duan, Enze;Meng, Lili;Zhao, Shida;Bai, Zongchun;Huo, Lianfei;Han, GuoFeng;Duan, Enze;Meng, Lili;Gu, Yue;Meng, Lili

作者机构：

关键词： Cage-reared ducks; Deep learning; Cage-net completion; Behavior recognition; Pose estimation

期刊名称：COMPUTERS AND ELECTRONICS IN AGRICULTURE （ 影响因子：8.9； 五年影响因子：9.3 ）

ISSN： 0168-1699

年卷期： 2025 年 236 卷

页码：

收录情况： SCI

摘要： Intensive farming of cage-reared ducks is an innovative approach being actively explored by the Chinese poultry breeding industry. In this process, obtaining behavioral and pose information is vital for assessing the physiological health of ducks, ensuring their welfare, and monitoring environmental conditions. However, the cage-rearing method leads to the loss of duck body features due to obstruction by the cage-net. This is not conducive to accurate object detection. Therefore, this study explores cage-net elimination using deep learning techniques and proposes a method for recognizing the behavior and estimating the poses of cage-reared ducks. Image datasets for cage-reared ducks aged 10, 20, and 30 days were constructed by applying a time gradient of 10 days under two conditions: with and without cage-nets. Cage-net segmentation was implemented using Mask R-CNN, and CycleGAN was utilized to restore obscured duck body features. The structural similarity index (SSIM) and peak signal-to-noise ratio (PSNR) values for cage-reared duck images at these three different ages, both before and after image completion, were 88.36 %, 91.78 %, and 91.27 %, as well as 37.33 dB, 40.45 dB, and 39.50 dB, respectively. An improved YOLOv7 model, YOLOv7-CRD (Cage-reared Ducks, CRD), was developed for recognizing typical behaviors (drinking, preening, and flapping) in cage-reared ducks. This model enhanced real-time performance by replacing the original YOLOv7 backbone network with the lightweight MobileOne neural network. Additionally, the network neck incorporated the SimAM attention mechanism to better focus on the key features of cage-reared ducks. The EIoU loss function, used instead of CIoU, further improved the accuracy of the bounding box regression and accelerated model convergence. This method achieved accurate recognition of behaviors in three types of day-old cage-reared ducks, with improvements of 2.60 %, 1.86 %, and 1.95 % in mean average precision (mAP); 0.033, 0.023, and 0.025 in recall; and 0.067, 0.018, and 0.027 in F1 score. Additionally, FPS increased by 21.05 %, and memory usage decreased by 24.66 % compared with the original YOLOv7. To further refine duck behavior, a cage-reared ducks' pose estimation method was proposed based on HRNet-48 by employing eight keypoints, including the duck's beak and head. It accurately estimated multiple postures for cage-reared ducks of various ages, improving the object keypoint similarity (OKS) by 13.22 %, 18.07 %, and 16.98 % compared to the results obtained before cage-net elimination. The proposed method effectively eliminated the interference caused by cage-net, enabling accurate recognition of behaviors and postures in cage-reared ducks. This advancement can significantly contribute to the scientific management of duck cage-rearing models and provide technological insights into the intelligent farming of other cage-reared poultry species.

分类号：