基于E-DCN-Cascade RCNN的牛群目标检测算法

Cattle object detection based on E-DCN-Cascade RCNN

提出一种基于可变形膨胀卷积级联网络的牛群目标检测算法。首先,采用可变形膨胀卷积网络(Expand-Deformable Convolutional Networks,E-DCN)提取牛身纹理特征,并使卷积区域始终覆盖在牛身周围,提高了牛身与牛舍背景的区分度;然后,结合E-DCN和特征金字塔网络(Feature Pyramid Network,FPN),对各层特征信息进行深度融合,增强了遮挡牛身的纹路特征,得到各层牛身特征图,解决了因牛身相互遮挡导致的漏检问题。实验结果表明,交并比(Intersection Over Union,IOU)阈值为0.75时,提出算法的平均精确度达到91.6%,比Cascade RCNN算法提高了4.2%。

A cattle object detection method based on E-DCN-Cascade RCNN is proposed.Firstly,a deformable-expandable convolutional network is used,so that the convolution area is always covered around the shape of the object,more suitable for the body shape of the herd,which can also improve the distinction between the cattle body and the background of the cowshed.Secondly,combining E-DCN with FPN,the feature information of each layer is deeply integrated,and the features of the cattle body pattern that shield the herd are enhanced,and the cattle body feature map of each layer is obtained,which solves the problem of missed detection caused by mutual occlusion.Experimental results show that when the IOU threshold is 0.75,the average precision of the proposed algorithm reaches 91.6%,which is 4.2%higher than the Cascade RCNN.