胡志伟, 杨华, 娄甜田, 等. 基于全卷积网络的生猪轮廓提取[J]. 华南农业大学学报, 2018, 39(6): 111-119. DOI: 10.7671/j.issn.1001-411X.2018.06.017
    引用本文: 胡志伟, 杨华, 娄甜田, 等. 基于全卷积网络的生猪轮廓提取[J]. 华南农业大学学报, 2018, 39(6): 111-119. DOI: 10.7671/j.issn.1001-411X.2018.06.017
    HU Zhiwei, YANG Hua, LOU Tiantian, HU Gang, XIE Qianqian, HUANG Jiajia. Extraction of pig contour based on fully convolutional networks[J]. Journal of South China Agricultural University, 2018, 39(6): 111-119. DOI: 10.7671/j.issn.1001-411X.2018.06.017
    Citation: HU Zhiwei, YANG Hua, LOU Tiantian, HU Gang, XIE Qianqian, HUANG Jiajia. Extraction of pig contour based on fully convolutional networks[J]. Journal of South China Agricultural University, 2018, 39(6): 111-119. DOI: 10.7671/j.issn.1001-411X.2018.06.017

    基于全卷积网络的生猪轮廓提取

    Extraction of pig contour based on fully convolutional networks

    • 摘要:
      目的  实现猪舍场景下非接触、低成本的生猪轮廓高效提取。
      方法  以真实养殖环境下的生猪个体为研究对象,提出一种基于VGG16 与UNET相结合的全卷积神经网络模型(VGG-UNET模型)。该模型采用批处理方法,迁移学习VGG16模型参数,通过在模型中构建复制通道深度融合图像深层抽象特征与浅层特征,实现对图像语义级别分割。在30头长白生猪的1 815张数据集上进行模型验证,通过设置不同批大小对比试验,并选取其中具有最佳效果的3组探讨批大小与评价指标值变化趋势间的关系。
      结果  测试集上的对比试验结果表明,VGG-UNET模型在像素精度与均交并比方面分别达到94.32% 和86.60%,比单独采用UNET模型分别高出0.89%和1.67%。不同指标值变化情况与批大小间的关系不尽相同。在本文试验环境下,批大小对模型收敛速度的影响不明显。不同批大小条件下PA及MIoU指标值变化综合分析得出,VGG-UNET模型具有较强稳定性和较高鲁棒性;批大小为8 的情况下VGG-UNET模型效果最佳。
      结论  本文提出的生猪轮廓提取方法(VGG-UNET模型)是有效的,能实现精确、稳定的生猪轮廓提取,且分割结果较为完整,同时模型具有较高鲁棒性,可为后续生猪个体识别研究提供参考。

       

      Abstract:
      Objective  To realize non-contacting and low-cost pig contour extraction under the piggery scene.
      Method  We took individual pig in the real culture environment as the research object, and proposed a full convolutional neural network model based on the combination of VGG16 and UNET (VGG-UNET model). We adopted the batch processing method in this model to transfer and learn the parameters of VGG16 model. We achieved semantic level segmentation of the image by combining the deep abstract feature and shallow feature in depth via building the duplicate channel. The model was verified on 1 815 datasets of 30 Large White× Landrace pigs. Comparison experiments of different batch sizes were performed, and three groups with the best results were selected to explore the relationship between batch size and the evaluation index.
      Result  Through comparison experiments on datasets, the pixel accuracy and mean intersection-over-union of VGG-UNET model were 94.32% and 86.60% respectively, which were 0.89% and 1.67% higher than those of the UNET model. The experiments showed different relationship between the change of different index values and batch size. Batch size had no obvious impact on the convergence rate of the model under this experimental environment. Through comprehensive analysis of PA and MIoU index values under different batch sizes, the VGG-UNET model showed the highest stability and robustness, and it was found to be the best when the batch size was 8.
      Conclusion  The VGG-UNET model is effective for accurate and stable extraction of pig contour. Such segmentation result is relatively complete and the model has higher robustness, which can provide a reference for follow-up identification of individual pigs.

       

    /

    返回文章
    返回