Automatic Recognition Of Terrestrial Wildlife In Inner Mongolia Based On Deep Convolution Neural Network

Mastering wildlife resources through image information is the main way of wildlife monitoring at present.Automatic recognition of wildlife monitoring images based on deep convolution neural network can solve the problems of low efficiency and high cost of artificial classification,but it also faces the problems of "changing illumination conditions of images","complex background information of monitoring images","uncertain location and size of targets" Such difficulties affect the accuracy of automatic recognition.To realize automatic recognition and classification of wildlife monitoring images more accurately and efficiently,this thesis studies the automatic detection and identification of wildlife based on deep convolutional neural network.Major works of this thesis are presented as follows.1.Monitoring image database of major terrestrial wildlife in Inner Mongolia is established herein.Surveillance images from Saihanwula reserve in Inner Mongolia from 2012 to 2018 period are collected by using the infrared inductive camera and wireless image sensor network to collect.Six national protected species are studied including Red deer,Antelope,Boar,Roe deer,Raccoon dog and Lynx.Total number of acquired images is 10,720 and the calibration work is conducted at image-level and pixel-level.2.An adaptive illumination algorithm for wildlife monitoring image is proposed.A monitoring image adaptive enhancement algorithm is proposed based on the Retinex theory with color constancy.A horizontal/vertical gradient guide filter is proposed to estimate the illuminance component image regarding the halo artifact caused by uneven illumination.Then the contrast stretch function with Otus threshold value is proposed and the contrast component image is stretched to improve the adaptive ability of the algorithm.Lastly,the reflected component image is calculated by using the illuminance component gray image,which contributes to the color correlation and adaptive enhancement of wildlife monitoring images.3.An automatic identification algorithm for wildlife based on the deep residual network of self-attention mechanism is proposed.Classical convolutional neural network is limited in field and it is incapable of accurately extracting wildlife characteristics.A deep residual network based on self-attention mechanism(SA-ResNet)is proposed and network structure and training of self-attention module are introduced for parameter tested and analysis.The experimental results showed that the accuracy of the test set presented in this thesis reached 90.6%.4.A target detection and recognition algorithm for wildlife based on deep convolutional neural network is proposed.The recognition accuracy of low-level target is improved to address influence and species imbalance of target location uncertainty,complex background environment.A detection and identification method for wildlife targets with Faster RCNN is proposed.Firstly,SA-Resnetl52 is built as a backbone network to extract feature maps.Then the K-means algorithm is used to re-regress the Anchors of the regional suggestion network to improve the accuracy of the network detection of wildlife.An adaptive category weight loss function is proposed to solve the problem of low accuracy caused by data imbalance of different species.The experimental results showed that the target detection algorithm proposed in this thesis reached 92.2%.In view of the high cost and low efficiency of manual sorting of traditional wildlife monitoring images,this paper proposes a solution for intelligent identification of wildlife.The proposed research is beneficial to image enhancement,deep convolutional neural network feature extraction,target detection and identification,which provide theoretical support for intelligent wildlife protection.