本文在python环境下，利用DETR模型对部分东北虎豹国家公园的动物进行了目标检测。DETR模型是Transformer架构在图像领域的应用，DETR模型包括三个部分其一是负责从图像中抽取特征信息的CNN（卷积神经网络）主干网络；其二是运用Transformer机制的编码器—解码器结构，用于对特征进行高层次的理解与建模；最后是负责进行最终目标检测结果输出的FNNs（简单前馈网络）。

作者选取了东北虎豹国家公园数据中的四个物种，总计五千多张图片。将DETR模型修改对应的参数之后在工作站上运行了121个epoch，经过长时间的训练，模型的mAP（平均精确度）达到了90%以上。选取过程中的最优权重文件之后，对验证集的图片进行了物种识别和定位，所有物种的召回率都在85%以上，除了样本数目较少的物种，其他物种的召回率都在95%以上，模型表现良好。

In this paper, the DETR model is used to detect and recognize some animals in the Siberian Tiger and Leopard National Park in Python environment. The DETR model is the application of Transformer architecture in the field of image. The DETR model includes three parts: the first is the CNN (convolutional neural network) backbone network responsible for extracting feature information from the image; the second is the encoder-decoder structure using the Transformer mechanism for high-level understanding and modeling of features; and the last is the FNNs (simple feedforward networks) responsible for the output of the final object detection results.

The author selected four species in the Siberian Tiger and Leopard National Park data, totaling more than five thousand pictures. After modifying the corresponding parameters of the DETR model, it was run on the computer for 121 epochs, and after a long period of training, the model's mAP (average accuracy) reached more than 90%. After selecting the optimal weight file in the process, the pictures of the validation set were identified. The recall rate of all species was above 85%, except for the species with a small number of samples, the recall rate of other species was above 95%, and the model performed well.