| 摘要: |
| 针对生猪智能化管理中传统标识方法存在的易脱标，易引起生猪感染等问题，采用基于改进YOLOv3模型的非侵入式方法，对生猪多个体识别进行研究。针对原有的YOLOv3模型，在Darknet53特征提取器中引入密连块，结合下采样层组成新的骨干网络；在YOLOv3模型中添加改进的SPP单元，最终构建了YOLOv3_DB_SPP模型。试验采用的猪脸数据集共分为10类，数据增强后样本为8 512张，训练集和测试集比例约为9∶1。结果表明:1)YOLOv3_DB_SPP模型在各分类概率阈值下检测猪脸数据集时的平均精度均值均高于YOLOv3模型；2)当IOU阈值为0.5，分类概率阈值为0.1时，YOLOv3_DB_SPP模型的平均精度均值比YOLOv3模型高 9.87%；3)YOLOv3_DB_SPP模型检测远距离有遮挡的小目标样本时，平均精度值均高于YOLOv3模型。YOLOv3_DB_SPP模型用于猪脸识别时，能够提高基础特征提取器的特征提取能力以及检测器的准确率。 |
| 关键词: 猪脸识别 YOLOv3 DenseNet SPP |
| DOI：10.11841/j.issn.1007-4333.2021.03.06 |
| 投稿时间：2020-06-10 |
| 基金项目:国家自然科学基金项目(31671589)；安徽省科技重大专项(201903a06020009)；安徽省重点研究与开发计划项目(1804a0702130) |
|
| Pig face recognition based on improved YOLOv3 |
|
HE Yutong,LI Bin,ZHANG Feng,TAO Haobing,GU Lichuan,JIAO Jun
|
| (School of Information and Computer, Anhui Agricultural University, Hefei 230036, China) |
| Abstract: |
| In order to solve the problems existed in the traditional identification methods during the process of intelligent management of pigs, for example, ear tags are easy to fall off and pigs are easy to be infected, a non-invasive method based on the improved YOLOv3 model was adopted to study the multi-individual identification of pigs. Based on the original YOLOv3 model, this study introduced DenseBlock in Darknet53 feature extractor and combined it with down-sampling layer to form a new backbone, then added improved SPP unit in YOLOv3 detector, and finally constructed YOLOv3_DB_SPP network. The pig face data sets were divided into 10 categories. After data augmentation, there were 8 512 samples in total, and the ratio of training set to test set was about 9∶1. The results showed that: 1)In each classification probability threshold, the mean average precision of YOLOv3_DB_SPP detected on pig face data set was higher than that of YOLOv3 model; 2)When the IOU threshold was 0. 5 and classification probability threshold was 0. 1, the mean Average Precision of YOLOv3_DB_SPP was 9. 87% higher than that of YOLOv3; 3)When the YOLOv3_DB_SPP model detected remote and occluded small targets, the average precision was higher than YOLOv3. The experiment proves that YOLOv3_DB_SPP model can improve the feature extraction capability of the basic feature extractor and the accuracy of the detector. |
| Key words: pig face recognition YOLOv3 DenseNet SPP |
