| The deep learning method which relies on the Internet and high-performance computers has obvious advantages in image recognition in recent years,and its effect has been proved in animal recognition.But there are few researches on primates living in the complex ecological environment.This paper takes the snub nosed monkey in the Shennongjia National Park in Hubei Province as an example,to discuss the recognition effect of snub nosed monkey in this area with deep learning-based object detection and instance segmentation,aiming to solve the problem of accurate identification and positioning of the snub nosed monkey with any pose and size in the pictures containing complex environmental background information,and solve the problems of existing research on the snub nosed monkey modeling data set with single source and recognition part limitation.This paper takes the pictures crawled online and shot on-the-spot as the training set,builds the Shennongjia National Park snub nosed monkey image recognition models of YOLO v3,Faster R-CNN and Mask R-CNN respectively,and determines the optimization models with optimal parameters.It analyzes the effect of these models based on different training sets,and then explores the optimal supplementary scheme of modeling data with limited samples.And this paper also makes comprehensive comparison of the above models from quantitative,qualitative and training time dimensions,in a bid to make clear the effect of these models on the rapid and accurate recognition of the snub nosed monkey in the pictures of real scene in the wild.The results show that:1.The YOLO v3 model with optimized parameters has better recognition effect on the faces of the monkeys than on their bigger-target bodies,and the training time is shorter.At the same time,compared with machine learning model and convolution neural network,the recognition accuracy of the model is greatly improved.2.The Faster R-CNN model with optimized parameters is not only better than the YOLO v3 model in overall recognition effect,but also more suitable for recognizing the monkeys’ bigger-target bodies.3.The Mask R-CNN algorithm based on the deep transfer learning,compared with the deep learning-based object detection,can not only improve the recognition accuracy,but also segment the monkey along the edge in the image to achieve precise recognition of the monkey image.4.Through comprehensive analysis of the above three models,We find that the Mask R-CNN model compared with YOLO v3 and Faster R-CNN models,has higher recognition accuracy,though with the training speed slightly slower than the YOLO v3 model but faster than the Faster R-CNN model.Therefore,the Mask R-CNN algorithm is more suitable for the monkey image recognition from the results of comprehensive comparison of operation efficiency and recognition accuracy.And the Mask R-CNN compared with YOLO v3 and Faster R-CNN,is more suitable for extending the samples by using the network crawled pictures as the alternative data source of the field photos,to solve the problem of limited samples of field photos.On this basis,this paper with the help of the YOLO v3,Faster R-CNN and Mask R-CNN models with optimized parameters combined with field shot photos and network crawled pictures,not only solves the problem of limited samples obtained from the wild,but also improves the automatic recognition level in complex ecological environment background,and then enhances the ability to monitor their survival state.This study not only has important practical significance for the field discovery and tracking of the snub nosed monkeys in Shennongjia National Park,but also provides technical support for the wildlife protection in this area,and thus helps to promote the further construction of the Shennongjia National Park. |
