融合视觉上下文与跨通道信息的生猪脸部轻量化检测模型

Lightweight model of pig face detection by fusing vision contextual and cross-channel information

生猪脸部的检测,是实现生猪个体身份、情绪状态和行为模式智能化自动识别与分析的关键环节。针对生猪在真实养殖环境中脸部区域通常存在遮挡、视角多变以及模型部署端资源有限等问题,提出了一种融合视觉上下文与跨通道信息的生猪脸部轻量化检测模型。本文采用ShuffleNet v2作为检测模型的特征提取模块;设计Nonlocal-PAN特征融合模块,融合不同尺度特征图中的视觉上下文信息和跨通道信息;利用NanoDet检测头模块来完成对生猪脸部目标的检测。实验结果表明,该模型在IoU为0.50、0.95、0.50:0.95水平下,分别获得了96.98%、81.16%和67.44%的平均精度,且模型大小保持在8.2 MB的轻量化水平,具有良好的边缘端移植性和较高精度,为生猪养殖智能化监测与分析提供了新技术。

Pig face detection is one of the most important prerequisite step for pig individual identification,emotion status and behaviors recognition in the field of smart pig breeding.To solve the problem of face occlusion and variety about visual angle in real senses of breeding,a lightweight model of pig face detection by fusing vision contextual and cross-channel information is proposed in this paper.Firstly,a lightweight feature extraction network,i.e.,ShuffleNet v2,is used as backbone network of the proposed model.Secondly,a nonlocal-PAN module is designed for fusing vision contextual information and cross-channel information of extracted feature maps with different scales.Finally,a lightweight NanoDet head is adopted as the object detection module based on those aggregated features.Experimental results show that the proposed model has achieved an accuracy of 96.98%,81.16%,and 67.44%at IoU level of 0.50,0.95 and 0.50:0.95 respectively with model size of 8.2 MB.The proposed model has an excellent ability of deployment in edge computing hardware and superior detection performance,which can be applied effectively in the field of smart pig breeding.