Research On Wild Animal Image Recognition Method Based On Deep Learning

The identification of wild animals is a necessary part of protecting wild animals,and fast and accurate identification methods will make the identification effect more effective.At present,the introduction of deep learning technology has solved the problems of low manual recognition efficiency and high recognition cost,but there still exist "target features are not obvious","target itself and its background are complex and changeable","target scale deformation",etc.Difficulties,which in turn affected the effectiveness of wild animal identification.This article aims to study the identification methods of wild animals,which is mainly divided into the identification of wild animals under a single angle and different angles.The identification of wild animals under a single angle is mainly used in the field detection environment,which can effectively collect characteristic photos of wild animals.The recognition application of wild animals from different angles is more extensive,mainly in scenes with complex background information.The main research contents are as follows:Firstly,a wild animal recognition algorithm based on local features is proposed.Aiming at the problem that wild animals have only local feature information under a single angle,the backpropagation method is used to calculate the contribution of wild animal feature map pixels to the target classification,so that the local features have a certain degree of discrimination.The cost-sensitive algorithm is introduced in the stage of training the target classifier.Finally,the local features of the target object are fused with the overall features,and the classifier is trained.Through experiments,it is concluded that this algorithm has higher recognition accuracy than other related algorithms.Secondly,a SSD target detection model based on Dense Net-169 is proposed.In view of the complex and changeable background of wild animals from different angles and their different shapes,this model uses Dense Net-169 network to replace the feature extraction network in the original model,which improves the detection effect of complex targets.Using the target detection model to locate and detect wild animals in images can effectively reduce the interference of external factors on wild animal detection.The experimental results show that this model has better detection effect than other target detection models,and the accuracy rate is increased by 1.67% on average.Thirdly,a bilinear convolutional neural network recognition model based on multiangle feature fusion is proposed.In view of the unstable feature extraction and poor performance of wild animals under different angles,the model starts from two aspects:establishing a network model with a good feature extraction effect and improving the loss function.By introducing the feature extraction network Dark Net53,and adopting a multiscale fusion method,the features of different scales are fully displayed.Introduce the central loss function and make it learn together with the Softmax loss function.At the same time,the wild animal images are restricted in the feature space,and the distance between the same species is reduced and the distance between different species is increased,so that the features obtained by the network have stronger discriminativeness.Finally,through experiments,the method has a significant improvement in average accuracy and average recall,reaching 93.63% and 90.85%,respectively.