软件导刊2024，Vol.23Issue(6)：25-31,7.DOI:10.11907/rjdk.241323
基于注意力机制与自适应特征融合的群养猪身份识别
Identification of Group-housed Pigs Based on Attention Mechanism and Adaptive Feature Fusion
摘要
Abstract
Aiming at the problem that body deformation and overlap during the attack reduce the accuracy of pig identification,a deep learn-ing method based on attention mechanism and adaptive feature fusion was developed to improve the accuracy of pig identification.Eight pigs were mixed into a new pen and 8 h of video/day was recorded for 3 days.From these videos,16,830 frames of aggressive behaviour were la-belled.Firstly,ResNet50 was chosen to extract the convolutional neural network(CNN)features of pigs.Secondly,feature pyramid network(FPN)was used to select three layers of features with different scales to optimise these features'locating and semantic information.Then a self-attention mechanism is used to improve the discrimination of these features,and adaptive spatial feature fusion(ASFF)was used to fuse features of different scales.Finally,the detector combined with convolution and sigmoid function was used to identify pigs.Using the proposed method,pigs were identified with mAP(mean average precision)of 95.59%and FPS(frames per second)of 37.6,respectively.These results indicate this method could identify pigs engaged in attack scenarios,helping to move the recognition of aggression from the group to the indi-vidual level.关键词
群养猪/身份识别/注意力机制/特征融合/深度学习Key words
group-housed pigs/identification/attention mechanism/feature fusion/deep learning分类
计算机与自动化引用本文复制引用
韩丁磊,陈晨,Steibel Juan,Siegford Janice,韩俊杰,王梦凡,徐雷钧,Norton Tomas..基于注意力机制与自适应特征融合的群养猪身份识别[J].软件导刊,2024,23(6):25-31,7.基金项目
国家自然科学基金项目(32102598) （32102598）
USDA National Institute of Food and Agriculture Program(2017-67007-26176,2021-67021-34150) （2017-67007-26176,2021-67021-34150）
江苏省高等学校大学生创新创业训练计划项目(202210299122Y) （202210299122Y）
