Abstract: Presently pig aggression recognition is still in the research stage of group level/pairwise level，and the identification of individual pig has become a necessary condition for further realizing individual level aggression recognition. In order to solve the problem of difficult identification caused by body deformation，occlusion，overlapping and other factors under pig aggression situation，an improved deep learning algorithm based on ResNeXt50（Residual networks with next‑50）was proposed to recognize pig identity in aggression situation. 18 000 frames were generated from the labelled 600 1‑second aggressive video episodes as the dataset.Firstly，the GECA（Ghost‑based efficient channel‑coordinate attention）module was embedded in the backbone network ResNeXt50 to enhance feature discriminative ability.Secondly，bidirectional feature pyramid network（BiFPN）was introduced to enhance the fusion capability of multi‑scale features.Then，position attention mechanism（PAM）was cascaded after BiFPN to improve the discrimination of global spatial features generated by pig body deformation，and channel attention mechanism（CAM） was used to optimize feature utilization through channel adaption.Finally，Fovea Head was used to recognize the identity of pigs under aggression situation.The identity of pigs could be recognized by using the proposed algorithm GBPC‑ResNeXt50（GECA‑BiFPN‑PAM‑CAM‑ResNeXt50）with a mean average precision（mAP）of 95.6%，which was 3.8 percentage points higher than that of the benchmark network ResNeXt50.The result indicates that this method can be used to recognize the identity of individual pig under aggression situation.This provides a foundation for promoting the conversion of pig aggression recognition from group/pairwise level to individual level，and also provides a reference for identification of other livestock under aggression situation.

Key words: Group?housed pigs, Identification, ResNeXt50, Deep learning, Bidirectional feature pyramid network, Double attention mechanisms

摘要： 目前猪攻击识别还处于群体级/成对级研究阶段，识别猪个体身份已成为进一步实现个体级攻击识别的必要条件。为解决猪攻击状态下因身体形变、遮挡、重叠等因素引起的身份难以识别的问题，提出一种改进ResNeXt50（Residual networks with next‑50）的深度学习算法以识别攻击状态下猪身份。从标记的600 段1 s 攻击视频片段中产生18 000 帧作为数据集。首先，在主干网络ResNeXt50 中嵌入GECA（Ghost‑based efficient channel‑coordinate attention）模块提升特征判别力。然后，引入双向特征金字塔网络（Bidirectional feature pyramid network，BiFPN）以提高多尺度特征的融合能力。接着，在BiFPN后级联位置注意力机制（Position attention module，PAM）以增强由猪体形变产生的全局空间特征的区分度，采用通道注意力机制（Channel attention mechanism，CAM）通过通道自适应优化特征利用率。最后，采用Fovea Head 识别攻击状态下猪身份。结果表明，提出的GBPC-ResNeXt50（GECA-BiFPN-PAMCAM-ResNeXt50）算法能够以95.6%的平均精度均值（mAP）识别猪身份，较基准网络ResNeXt50提升3.8百分点。表明该方法能够识别攻击状态下个体猪身份，为推动猪攻击识别从群体级/成对级到个体级的转变提供基础，同时也为其他家畜的攻击状态下身份识别提供参考。

关键词: 群养猪, 身份识别, ResNeXt50, 深度学习, 双向特征金字塔网络, 双重注意力机制