Research On Fine-grained Birds Classification Model Based On Transfer Learning

Efficient monitoring of forests and their associated wildlife resources is an important task for forest managers.In order to solve the practical problems in bird observation in the wild,a fine-grained image classification system based on transfer learning is studied and implemented.With deep learning as a tool and CUB-200-2011 bird data set as the basic data set,centering on transfer learning and fine-grained image classification and recognition and utilizing forestry,computer science,statistics and other methods and technologies,the author carried out the research by improving the feature extraction algorithm,optimizing the model structure,and fusing the feature model.The five main aspects covered in the paper are as the following::(1)Propose an improved fine-grained feature extraction algorithm based on transfer learningIn order to realize fine-grained feature extraction with the basis of transfer learning,an optimized KLT-SNE algorithm was proposed.The KLt-SNE algorithm is based on Kullback-Leibler divergence and t-SNE.Firstly,an unknown distribution p(x)is given,and then a q(x|?) with the same dimension as the unknown distribution is established.The parameter ? to be configured is estimated by taking N samples from p(x).The results show that this algorithm can effectively achieve dimensionality reduction of data.(2)Construct a feature extraction model based on an improved fine-grained feature extraction algorithmA feature extraction model which is composed of "convolution basis" layer deriving from VGG16 network and "dense connection layer" is constructed.According to the characteristics of the data set,six connection layers such as linear connection layer,Re LU layer and Dropout layer are added to the dense connection layer.In addition,a linear connection layer containing 200 classifiers is included too.By using this model and the improved KLT-SNE algorithm for feature extraction,the data dimensions can be reduced to 338 in the case of 80% variance contribution rate,and the data dimension reduction effect is obvious.(3)Design a fine-grained image classification model based on transfer learningModel fine-tuning strategy,model fine-tuning algorithm and data enhancement technology are applied in the design of fine-grained image classification model for transfer learning.By comparing the accuracy rate,recall rate,precision and other indicators of the alternative base model,Res Ne Xt was finally selected as the base model.Experimental results show that the proposed model scheme can achieve 84.43%accuracy in CUB-200-2011 data set,which is better than most public models.(4)Propose and implement a weighted fusion fine-grained classification modelIn this paper,a weighted weight allocation fusion model is proposed,which can record the optimal training results after dynamically allocating weights in the training process,and then obtain the optimal model weight allocation model.It can optimize training for different models or set the step size of parameters to control the precision of weights.Compared with the specified hyperparameter,the weighted model is more flexible and has practical application and popularization value.In the experiment,the Res Ne Xt152 model and VGG16 model were fused,and it was calculated that when the VGG16 model allocated 0.3 weight and the Res Ne Xt152 model allocated 0.7 weight,the model classification effect was the best.This paper focuses on fine-grained image classification and recognition,based on cub-200-2011 bird data set,through improving the fine-grained feature extraction algorithm based on transfer learning,designs and implements a new bird classification model based on deep transfer learning.Using this model,the function of the existing monitoring application system is expanded,which makes the application system more intelligent and can collect bird information more accurately.