Research On Automatic Recognition Method Of Wildlife Monitoring Images Based On Convolutional Neural Network

Using deep learning to automatically identify and classify wildlife can greatly improve the efficiency of wildlife monitoring and provide reliable data support for the formulation of wildlife protection strategies.However,the current automatic identification of wildlife also faces the problems of low accuracy due to the complex background information and low quality of the monitoring images,which has affected the application of deep learning technology in the field of wildlife protection.In order to achieve high-accuracy automatic identification of wildlife,based on the self-built main terrestrial wildlife dataset in Inner Mongolia and the Snapshot Serengeti dataset,this paper carried out research on automatic identification methods of wildlife monitoring images based on convolutional neural networks and analyzed the effect of image quality on the recognition effect.The method of improving the recognition accuracy of low-quality images was explored.The main contents are as follows:1.An automatic recognition algorithm for wildlife monitoring images based on regions of interest and convolutional neural networks is proposed.In order to reduce the interference of complex background information on species identification,the target detection method based on regression algorithm is used to detect and segment the wildlife area in the monitored image to obtain the ROI image.A global-local VGG16 dual-channel network model is constructed to extract features from sample images.Finally,a classifier is used to identify wild animals.The proposed recognition model is superior to the two-channel network model under the VGG19 structure and other models such as VGG16,R-CNN and Fast R-CNN in the recognition of five wildlife such as red deer,antelope,roe deer,lynx and boar.The mean average precision of the model reached 0.912,and the maximum improvement of MAP reached 0.256.2.An automatic recognition algorithm for wildlife monitoring images based on SE-Res Ne Xt is proposed.In order to avoid the problem of high labeling cost caused by the target monitoring algorithm and at the same time to ensure the accuracy of automatic identification,a wildlife automatic identification model based on SE-Res Ne Xt is proposed.Five wildlife images of red deer,antelope,roe deer,lynx and raccoon dog are selected as training samples,and the recognition effects of SE-Res Ne Xt and Res Net at 50,101 and 152 layers were compared respectively.Experiments show that when the learning rate is 0.01,SE-Res Ne Xt101 has the best training effect and the test accuracy reaches 93.5%,which is higher than of Res Net.Experiment of performance verification is based on the monitoring images of 26 species in Snapshot Serengeti dataset.It shows that accuracy of SE-Res Ne Xt101 on 13 categories is better than Res Net-101 and the maximum difference is 31.2%.3.The effect of different image qualities on the automatic identification algorithm of wildlife is analyzed,and the preprocessing method is used to improve the recognition accuracy of low-quality images.Aiming at the low-quality monitoring images collected by wireless image sensor network,the effects of different qualities on six different recognition models are analyzed through experiments.Experimental results show that the low quality of the image will degrade the recognition performance of all models.When the bit error rate is less than 5%,the model of the SE-Res Ne Xt performs optimally,and the test accuracy of SE-Res Ne Xt152 decreases only by 3.16%.In order to improve the accuracy of lowquality images,preprocessing methods of image pixel expansion and image sharpening and denoising are introduced.A comparative experiment is conducted on the recognition effect of low-quality images before and after processing.The experimental results show that when the bit error rate of the image is less than 5%,the accuracy of the model SE-Res Ne Xt 50 is increased by 1.21% after the image is preprocessed,proving that preprocessing can reduce the impact of low quality on model performance.The automatic recognition algorithm of wildlife monitoring images based on convolutional neural network proposed in this paper can provide a solution for the application of deep learning in the intelligent monitoring of wildlife and improve the intelligent level of wildlife protection.