基于目标检测仔猪乳头计数及乳房形态评估方法

Counting and assessing piglet teats using object detection

母猪乳头数量是生猪选育中重要参考指标之一,也是生猪繁殖表型数据的组成部分。成年母猪其腹部视频较难获取,且容易受污渍干扰,乳头自动点数实现难度较大,人工计数母猪乳头数工作量大、强度高、效率低、容易产生人为误差。鉴于猪仔从出生到成年乳头数量保持一致性,该研究提出了一种基于仔猪腹部视频的深度学习乳头计数及乳房形态评估方法。通过架设在仔猪初生护理平台上的相机拍摄仔猪腹部视频,根据清晰度筛选出细节清晰的帧序列图像集,经过数据预处理再使用改进Pignip-YOLOv5s目标检测网络对仔猪乳头进行自动计数。为提高计数准确率,帧序列图像集的乳头计数使用滑动窗口取众数得到最终计数结果。试验结果表明,改进的Pignip-YOLOv5s平均精度值(mean average precision, mAP)高达0.97,较YOLOv5原模型具备更高的鲁棒性。该研究方案在113段仔猪腹部视频数据集上测试得到仔猪乳头计数方法准确率达90.26%。同时该研究提出仔猪乳房形态评估参数乳头成对数、乳头间距,从而量化仔猪乳头形态表型特征,构建了母猪乳房外在形态指标,可为母猪选育和繁殖工作提供重要的参考依据。

The teat count in a sow can serve as a key reproductive phenotype,thus offering valuable insights for selective breeding in a vital component of PSY(pigs weaned per sow per year).There is a positive correlation between the number of teats in pigs and their litter size.Particularly,the teat count in piglets was closely aligned with the average of their parents.Consequently,it is very necessary to select the piglets using teat count.The symmetry and shape of a sow's teats are two of the most important indicators of nursing ability.The better lactation performance was represented by the more orderly and regular arrangements.However,it is still challenging to capture clear videos of an adult sow's abdomen,due to the potential interference from stains.Automatic teat counting is also required for the labor-saving,higher efficiency,and higher accuracy,compared with the manual.In this study,a deep learning-based approach was proposed for the teat counting and evaluation using videos of piglets'abdomens.Among them,there was consistency in the number of teats from birth to adulthood of female piglets.Specifically,a camera was first installed on the piglet management platform,in order to capture the videos of the abdomen of piglets(2-7 days old).Then,these videos were screened using clarity.A sequence of images was preprocessed to facilitate the automatic teat counting via an enhanced Pignip-YOLOv5s object detection network.A sliding window majority voting mechanism was applied to the teat count sequence to acquire the final tally for high counting accuracy.Experimental results show that the improved Pignip-YOLOv5s achieved a mean average precision(mAP)of 0.97.Better performance was also obtained in the challenging conditions,such as tightly spaced teats at the piglet's abdominal end,complex body textures,and obscure vision from the umbilical cord and the shadow.The higher robustness was observed,compared with the original.There was an accuracy rate of 90.26%for teat counting in the dataset of 113 piglet abdomen videos.Some parameters were selected to quantify the piglet teat morphology,such as the number of paired teats and the distance between teats.A teat classification was also established for the left and right teats,according to the teat positions obtained from the Pignip-YOLOv5s object detection network.The image was divided into the quartile regions.The teat midpoints in each region were calculated to classify the left and right teats on the piglet belly.Additionally,a teat pairing algorithm was introduced to identify the teats in pairs,in order to calculate the pairing rates for the other data of teat morphology.A practical value was offered for the piglet teat counting and morphology assessment using images of the piglet abdomen.In summary,the target detection-based piglet teat counting and mammary evaluation can serve as a novel and effective way to extract breeding indicators in the livestock breeding industry.The finding can also provide high accuracy,speed,and efficiency in the realm of boar breeding.