基于卷积神经网络的猪脸特征点检测方法

Feature Point Detection Method of Pig FaceBased on Convolutional Neural Network

针对牲畜面部识别在养殖行业广泛需求的问题,提出一种基于卷积神经网络的猪脸特征点检测方法,解决了猪脸特征点难检测的问题.首先,采集猪面部数据并进行特征点标注,使用新的采集方法以解决猪口部通常不可见的问题;其次,对猪脸数据和人脸数据进行结构计算,匹配相似度较高的猪脸和人脸,构建猪脸人脸匹配数据集;再次,利用匹配数据集训练TPS(thin plate spline)形变卷积神经网络,得到形变后的猪脸数据集以适配人脸特征点检测模型;最后,使用形变猪脸数据集对人脸特征点检测神经网络模型进行微调,得到猪脸特征点检测模型.实验结果表明,用该方法进行猪脸特征点检测,错误率仅为5.60%.

Aiming at the problem of wide demand of livestock facial recognition in the breeding industry,we proposed a feature point detection method of pig face based on convolutional neural network,which solved the problem that it was difficult to detect feature points of pig face.Firstly,the pig face data was collected and the feature points were marked,and a new collection method was used to solve the problem that the pig mouth was usually invisible.Secondly,we calculated the structures of the pig face data and the human face data,matched the pig face and human face with high similarity,and constructed the pig face and human face matching data set.Thirdly,TPS(thin plate spline)deformed convolutional neural network was trained by matching data set,and the deformed pig face data set was obtained to fit the feature point detection model of human face.Finally,the feature point detection neural network model of human face was fine-tuned by using the deformed pig face data set,and feature point detection model of the pig face was obtained.The experimental results show that the error rate is only 5.60%by using the proposed method to defect feature points of pig face.