Research On Wildlife Recognition Based On Deep Learning

Wildlife plays a role in regulating the structure of ecosystems and maintaining a healthy and balanced ecosystem.How to monitor and assess wildlife resources more efficiently and accurately has become a key issue in research.This study focuses on deep learning as the main research method,and conducts research on wildlife image acquisition and recognition in order to monitor wildlife more conveniently and efficiently,distinguishing wild animal species,with a focus on the construction of image acquisition platforms and the improvement of object detection algorithms.The main research content and conclusions are as follows:（1）A system platform is built based on the way of obtaining wildlife images.The system requirements and design scheme are determined,with a quadrotor unmanned aerial vehicle as the main equipment,and MAVLink as the communication protocol between the ground station and the unmanned aerial vehicle.The ground station software is used to complete the overall workflow of the system.In order to deal with the fisheye effect in the binocular camera,coordinate system transformation is used to calculate the internal and external parameters of the camera and complete Zhang’s camera calibration.（2）Due to the low quality and small quantity of wildlife datasets,this thesis selects 10 species of wildlife and expands the dataset to 4352 images through methods such as reverse cropping and image blending.Subsequently,the images are annotated using Labelme software,and the annotated dataset is converted into JSON format.Finally,feature visualization analysis is carried out for data classification.（3）In terms of algorithm optimization,this thesis adopts parameter transfer learning and uses the Ada Grad optimization algorithm to solve the loss function.The model performs well when Res Net＿101is used as the main network framework,with a learning rate of 1×10^-4 and a batch size of 32.Subsequently,the Grid R-CNN algorithm is improved from four aspects:adding grid feature expression areas,reducing non-maximum suppression thresholds,introducing attention mechanisms,and multi-scale feature fusion.Comparative experiments and analysis with other algorithms show that the detection speed,recall,precision,and mean average precision（m AP）are improved by 41.58%,3.52%,3.81%,and 3.75%,respectively,compared to the unimproved algorithm.Finally,ablation experiments are conducted,which show that adding grid feature expression areas improves detection speed by nearly30%,reducing non-maximum suppression thresholds improves precision and detection speed by 2.25%and 19%,and introducing attention mechanisms and multi-scale feature fusion improves recall,precision,and m AP by 4.92%,5.37%,and 6.38%,respectively.The algorithm’s performance is significantly improved.