基于幅值迭代剪枝的多目标奶牛进食行为识别方法

Detection Method of Multi-objective Cows Feeding Behavior Based on Iterative Magnitude Pruning

针对奶牛进食行为监测通常要为每头奶牛配备监测设备,但受限于设备成本,很多应用于奶牛养殖场的奶牛行为监测方法难以普及的问题,提出了一种多目标奶牛进食行为识别方法,基于YOLO v3算法,根据目标差异,将牛舍中的奶牛分为3类目标来实现奶牛进食行为监测,以通过单台设备监测多头奶牛的进食行为。YOLO v3算法具有计算成本高、能源消耗大、设备依赖性强等不足,针对该问题,参考彩票假设,提出了一种基于幅值迭代剪枝算法的更优稀疏子网络筛选方法,使参数数量下降了87.04%,平均精度均值(mAP)达到了79.9%,较原始网络提高了4.2个百分点。说明了通过幅值迭代剪枝技术降低奶牛行为监测任务成本的可行性,验证了基于彩票假设从奶牛进食行为识别模型中筛选出更优稀疏子网络的有效性,为降低动物行为监测任务的成本提供了参考。

The existing methods for monitoring the cow's dietetic behavior do not allow monitoring of multiple cows simultaneously through a single device.A multi-objective cow dietetic behavior identification method was proposed based on the YOLO v3 algorithm.According to the difference in the goals,the cows to be monitored were classified to three groups to achieve dietetic behavior monitoring of multiple cows with a single device.However,the YOLO v3 algorithm has some disadvantages,such as high computational cost,large energy consumption,and strong equipment dependence.So the lottery ticket hypothesis was referred to apply this approach.And an iterative magnitude pruning algorithm for the identification of cow dietetic behavior based on the YOLO v3 network was proposed.Using this approach,the number of parameters was decreased by 87.04%,the mean average precision(mAP)value reached 79.9%,which was increased by 4.2 percentage points.Nevertheless,results proved that through the iterative magnitude pruning technique,the cow behavior monitoring task was feasible at a reduced cost.The effectiveness of screening out the optimal sparse subnetwork from the cow dietetic behavior identification model based on the lottery ticket hypothesis was verified.