基于目标检测的东北虎个体自动识别

东北虎个体的自动识别是种群数量评估和制定有效保护策略的重要基础。以东北虎林园和怪坡虎园38 只虎为研究对象，将目标检测方法首次应用到东北虎个体识别研究中，采用多种深度卷积神经网络模型，以实现虎个体的自动识别。首先通过相机在不同角度对 38 只东北虎进行拍摄取样，建立包含13579张图像的虎样本数据集。由于虎的体侧条纹信息不具有对称性，所以运用单次多盒目标检测（Single Shot MultiBox Detector， SSD）方法，对虎的躯干左侧条纹、右侧条纹以及脸部等不同部位图像，进行自动检测并分割提取，极大节省手工截取时间。在检测分割出的左右侧及脸部不同部位图片基础上，运用上、下、左、右平移变换进行数据增强，使图片数目扩大为原来的5 倍。采用LeNet、AlexNet、ZFNet、VGG16、ResNet34共5 种卷积神经网络模型进行个体自动识别。为了提高识别准确率，运用平均值和最大值不同组合方式来优化池化操作，并在全连接层引入概率分别为0.1、0.2、0.3、0.4的丢弃（Dropout）操作防止过拟合。实验表明，目标检测模型耗时较少，截取分割老虎不同部位条纹能达到0.6 s/张，远快于人工截取速度，并且在测试集上准确率能达到97.4%。不同姿态下的目标部位都能正确识别并分割。ResNet34模型的准确率优于其他网络模型，左右侧条纹以及脸部图像识别准确率分别为93.75%、97.01%和 86.28%，右侧条纹识别准确率优于左侧条纹和脸部图像。研究为野生虎自动相机影像的识别提供技术参考。在未来研究中，对东北虎个体影响数据进行扩充，选取更多影像数据进行训练，使网络具有更强的适应性，从而实现更准确的个体识别。

Amur tiger individual automatic identification based on object detection

The automaticly individual identification of Amur tigers (Panthera tigris altaica) is important for population monitoring and making effective conservation strategies. In this paper, we applied the object detection method and deep convolutional neural network models to individual identification with the images of 38 Amur tigers in the Northeast Tiger Forest Park (West 126°36'', North 45°49'') and Guaipo Tiger Park (West 123°37'', North 42°4''). Firstly, the Canon EOS 200D camera was used to establish the data set containing 13579 pictures of Amur tiger from different angles. Since the tiger''s body stripes of two sides are not symmetrical, single shot multibox detector (SSD) was used to automatically intercept and distinguish the left and right body stripes as well as the face of tiger, which greatly saved the time of manual interception. On the basis of the interception image results, the data were enhanced to 5 times by the up, down, left and right transformation. Then, LeNet, AlexNet, ZFNet, VGG16 and ResNet34 were used for individual identification. Furthermore, the pooling model was optimized by using different combinations of average pooling and maximum pooling, and the dropout operations with probabilities of 0.1, 0.2, 0.3 and 0.4 were introduced to prevent overfitting. The experiment shows that the target detection model in this study takes less time than manual interception. It can intercept and segment the stripes in different parts of the tiger with 0.6 seconds for one image, which is much faster than manual interception, and the accuracy rate on the test set can reach up to 97.4%. The target parts can be correctly identified and segmented for Amur tiger images with different kinds of positions. Finally, based on the experiment results, we found that the accuracy of ResNet34 is better than that of other network models. The recognition accuracies of left, right stripe and face images are 93.75%, 97.01%, and 86.28%, respectively. The recognition accuracy of right stripes is better than left stripes and face parts. This study can provide technical support for automatic camera image recognition of wild tigers. The method in this paper can be applied to the identification of individual in the species which have distinct streaks or spots on the body. In the future work, the image dadaset of individual Amur tigers will be expanded and much more image data will be selected for training set so as to make the network more adaptable and realize more accurate individual identification.