Research On Bird Species Image Recognition Based On Convolution Neural Network

China has vast territory and abundant animal resources.The proportion of bird resources occupies a large proportion of animal resources,and the protection of birds is one of the keys to promoting sustainable development.Bird recognition belongs to fine-grained image recognition.For bird image recognition,there are problems such as small differences between subclasses and large differences within classes.To solve problem,this thesis studies it from the following aspects:Two methods based on weakly supervised learning are proposed.They are the bird recognition method based on repeated CAM(Class Activation Mapping)network and bilinear network and the method based on adversarial complementary attention and cross-layer high-order bilinear network.First,a bird recognition method based on iterative CAM network and bilinear network is proposed.The method consists of three parts,iterative CAM object localization network,improved bilinear network,and embedding space(improved loss function).Using the CAM network to extract the discriminative area only,it is easy to lead to the lack of other supplementary discriminative area feature,so after using the CAM network to extract the salient area,set a threshold,suppress this part of the area,and send the processed image to the CAM network again,Extract other supplementary discriminative regional features,and use probability weighting to fuse two different discriminative regional feature maps.The bilinear network based on deformable convolution is selected to extract high-order image features.Finally,it is sent to the embedding space,and the discriminativeness of the features is further improved through the improved loss function.Experimental results show that this scheme has excellent classification performance in CUB-200-2011,and the classification accuracy reaches 87.5%.A method based on adversarial complementary attention and cross-layer highorder bilinear network is proposed.The method consists of three parts,the object positioning network against the complementary attention mechanism,the global branch,and the cross-layer high-order bilinear network.The anti-complementary attention object positioning network can effectively find the complementary object area and locate the main object area through the dynamic erasure mechanism.The cross-layer high-order bilinear network extracts the features of different convolutional layers,thereby enhancing the feature representation ability.Introduce the global branch to extract the global features of the image.Experimental results show that the classification accuracy of this method on CUB-200-2011 reached 88.4%.Finally,this thesis builds a small bird image recognition system based on the improved method.It mainly realizes the functions of bird image recognition.To sum up,this thesis has proposed two different bird image recognition solutions for the problems of bird image recognition,and neither of them requires strong supervision information,which effectively solves the problems of bird recognition,and has great practical application value.