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

针对群养猪攻击过程中身体形变和重叠影响猪身份识别精度的问题,开发一种基于注意力机制和自适应特征融合的深度学习方法以提高猪身份识别精度。录制猪栏中8只猪每天8小时共3天的视频,并筛选含攻击的16830帧作为数据集。首先,采用ResNet50提取猪的卷积神经网络(CNN)特征;然后,采用特征金字塔网络(FPN)在ResNet50中选择3层不同尺度的特征,以优化这些特征的定位和语义信息;接着,采用自注意力机制提高这些特征的区分度,并采用自适应空间特征融合(ASFF)以融合不同尺度的特征;最后,利用卷积和Sigmoid函数相结合的检测器对猪身份进行识别。使用该方法后,猪身份识别的均值平均精度(mAP)达到95.59%,速度达到37.6 f/s。结果表明,该方法能够在攻击场景下有效识别猪身份,有助于将攻击识别从群体级细化为个体级。

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.