高级检索
    刘文定, 李安琪, 张军国, 谢将剑, 鲍伟东. 基于ROI-CNN的赛罕乌拉国家级自然保护区陆生野生动物自动识别[J]. 北京林业大学学报, 2018, 40(8): 123-131. DOI: 10.13332/j.1000-1522.20180141
    引用本文: 刘文定, 李安琪, 张军国, 谢将剑, 鲍伟东. 基于ROI-CNN的赛罕乌拉国家级自然保护区陆生野生动物自动识别[J]. 北京林业大学学报, 2018, 40(8): 123-131. DOI: 10.13332/j.1000-1522.20180141
    Liu Wending, Li Anqi, Zhang Junguo, Xie Jiangjian, Bao Weidong. Automatic identification method for terrestrial wildlife in Saihanwula National Nature Reserve in Inner Mongolia of northern China based on ROI-CNN[J]. Journal of Beijing Forestry University, 2018, 40(8): 123-131. DOI: 10.13332/j.1000-1522.20180141
    Citation: Liu Wending, Li Anqi, Zhang Junguo, Xie Jiangjian, Bao Weidong. Automatic identification method for terrestrial wildlife in Saihanwula National Nature Reserve in Inner Mongolia of northern China based on ROI-CNN[J]. Journal of Beijing Forestry University, 2018, 40(8): 123-131. DOI: 10.13332/j.1000-1522.20180141

    基于ROI-CNN的赛罕乌拉国家级自然保护区陆生野生动物自动识别

    Automatic identification method for terrestrial wildlife in Saihanwula National Nature Reserve in Inner Mongolia of northern China based on ROI-CNN

    • 摘要:
      目的利用红外自动感应相机对野生动物进行图像监测是对野生动物保护管理的有效手段,为了解决野外复杂背景环境导致的野生动物监测图像自动识别准确率低的问题,提出一种基于感兴趣区域(ROI)与卷积神经网络(CNN)的野生动物物种自动识别方法。
      方法以红外自动感应相机在内蒙古赛罕乌拉国家自然保护区内拍摄的马鹿、斑羚、猞猁、狍和野猪这5种国家级陆生保护动物的图像为实验样本,采用基于回归算法的目标检测方法,对监测图像中野生动物区域进行检测并分割,生成ROI图像,减少复杂背景信息对物种识别的干扰;利用裁剪、仿射变换等方式对样本数据进行扩充;构建基于全局-局部的VGG16双通道网络模型对样本图像进行训练,最后接入分类器输出物种识别结果。同时,构建了基于VGG19的双通道网络模型对样本图像进行训练,并与本研究训练结果进行比较;另外,将样本图像分别输入本研究算法与VGG16、R-CNN、Fast R-CNN算法进行训练,对比不同算法下的识别效果。
      结果利用本研究模型对样本图像进行训练时,测试集的平均识别精度均值MAP达到0.912,相对于VGG19结构下的训练模型和VGG16、R-CNN、Fast R-CNN,得到了更高的MAP值。
      结论相比于其他算法,本研究提出的物种识别模型更适合于复杂背景下的野生动物监测图像的物种识别,可以得到更高的MAP值与更优的识别效果。

       

      Abstract:
      ObjectiveThe use of infrared automatic sensing cameras for image surveillance of wildlife is an effective method for the protection and management of wildlife. In order to solve the problem of low accuracy of the automatic identification of wildlife monitoring images caused by the wild complex background environment, a method based on the region of interest (ROI) and convolutional neural network (CNN) was proposed.
      MethodImages of five kinds of national protected terrestrial animals such as red deer, goral, pheasant, badger, and wild boar taken in Saihanwula National Nature Reserve, Inner Mongolia of northern China by an infrared automatic sensing camera were used as experimental samples. Using target detection based on regression algorithm, the wild animal images in regional monitoring detection segmentation generated ROI images to overcome the interference of background information on wild animal identification; then perform affine transformations and cropping to achieve random sample data augmentation; proposing a global local VGG16 based on dual channel network to train sample images and use a classifier to output the recognition result finally. At the same time, a dual-channel network model based on VGG19 was constructed to train sample images and compared with the proposed model. The sample images were trained by the algorithm of this paper and VGG16, R-CNN and Fast R-CNN, respectively. The recognition results under different algorithms were compared.
      ResultResults showed that the mean average precision (MAP) of the test set was 0.912 compared with the other models and algorithms when sample images were trained by the model in the paper.
      ConclusionCompared with other algorithms, the species identification model presented in this paper is more suitable for species identification of wildlife monitoring images under complex backgrounds, and can obtain higher MAP and better recognition effects.

       

    /

    返回文章
    返回