基于Transformer与自适应空间特征融合的群猪目标检测算法研究

Study on pigs target detection algorithm based on Transformer and adaptively spatial feature fusion

针对深度学习群猪目标检测算法精确度低和模型占用内存大等问题,提出基于Transformer与自适应空间特征融合的群猪目标检测算法。搭建群猪图像采集设备,以视频帧作为数据源,提取关键帧并剔除模糊图像,采用Labelme标注图像中猪只,建立群猪图像数据集;将Swin Transformer网络作为主干网络,在FPN后引入自适应空间特征融合方法作为特征融合网络;提出RIoU作为预测框回归损失计算方法。结果表明,该算法在精确率、召回率、F1值和平均精确率指标方面分别达到93.6%、97.2%、0.953、96.5%,检测速度为34.9 Hz且模型大小仅为20.6 MB,与YOLOv4相比上述指标分别提高1.5%、1.7%、1.6%、2.4%,模型占用内存量缩小12.5倍,检测速度提高13 Hz。研究有助于智能化猪场建设,为养殖场动物计数和行为识别等方面提供技术支持。

The deep learning algorithm for pigs target detection has low accurate and a large model size,a pig target detection algorithm was proposed based on transformer and adaptively spatial feature fusion.Set up pig image acquisition equipment,use video frames as the data source,extract key frames and eliminat blurred images,pigs image dataset was made after using Labelme to label the pigs in the image.Introduct Swin Transformer network as the backbone network,a method of adaptive spatial feature fusion was used as a feature fusion network after FPN.The RIoU as the regression loss calculation method of prediction box was proposed.The results showed that the precision,recall,F1score and average precision of the algorithm were 93.6%,97.2%,0.953 and 96.5%,respectively,the detection speed was 34.9 Hz and the model size was only 20.6 MB.Compared with YOLOv4,the above indicators were increased by 1.5%,1.7%,1.6%and 2.4%,respectively,the memory footprint of the model was reduced by 12.5 times,detection speed was increased by 13 Hz.This research is helpful to the construction of intelligent pig farm,animal counting and behavior recognition.