Intelligent Recognition Of Wetland Birds Based On Deep Learning

Wetland is not only an important ecological science gene bank,but also the main habitat of birds.The species and quantity of birds can be used as an important evaluation index of wetland natural environment and ecosystem.The construction of wetland bird intelligent identification system is not only conducive to the management and research of birds in the natural ecological reserve,but also can monitor and track endangered birds in the reserve,so as to provide more targeted measures for wetland bird protection.With the development of artificial intelligence,the object detection method based on deep learning shows its excellent feature extraction ability in bird recognition,but it also needs a large number of evenly distributed data support.However,the actual wetland bird data is very complex compared with the public data set.It usually has the problems of small number of samples and uneven distribution of samples.Taking the bird data of Dahuangpu wetland in Tianjin as an example,this paper formulates the following research ideas:First,expand the data set based on conditional generative adversarial network.Conduct statistical screening on the collected data sets and eliminate the categories with too few samples.The data set is preliminarily expanded by using affine transformations such as rotation and scaling and adding Gaussian noise.Build a conditional generative adversarial network and use the generator to intelligently generate new bird images to increase the diversity of data sets.Second,improve the object detection algorithm and build an intelligent recognition system for wetland birds.Three one-stage object detection algorithms,yolo-v3-spp-net,SSD and Retinanet,are selected to carry out the intelligent recognition of wetland birds.According to the characteristics of wetland bird images,the model is locally optimized and improved,including:introducing the spatial pyramid pooling module,adjusting the network parameters,modifying the NMS to soft-NMS,selecting the loss function and anchor initial value with the best effect,so as to improve the detection effect of the intelligent recognition system.The generalization ability of bird intelligent recognition system is verified on the test set.Based on the above ideas,this paper has carried out a large number of experiments and Analysis on the bird data set of Dahuangpu wetland in Tianjin.The model effect is good and the recognition accuracy is significantly improved.The mAP on yolo-v3-spp-net reaches 0.6114 and AR reaches 0.6002;The mAP on SSD reaches 0.6766 and AR reaches 0.6876;The mAP on Retinanet reaches 0.7929 and AR reaches 0.7466.