Research On Fine-grained Image Classification Based On Deep Convolutional Neural Network

Fine-grained image classification is a challenging task in the field of image classification,which aims to classify sub-categories.Because the discriminative information exists in small local areas,most of the fine-grained image classification algorithms are based on the discriminative region,and then train convolutional neural network to perform fine-grained classification based on the discriminative region.By analyzing the algorithm of obtaining the discriminative region,it is found that these algorithms either rely on manual annotation information or the discriminative region contain large redundant information,which seriously affects the performance of model classification.In this thesis,it focues on the research of fine-grained image classification algorithm based on the discriminative region localization and training method,the main research content and innovation work are as follows is:1.Aiming at the discriminative region localization,this thesis proposes a finegrained image classification algorithm based on multi-layer feature fusion of the attention perception.Firstly,this algorithm uses a bilinear operation layer to improve the attention network,and uses the improved attention network and localization network to automatically locate the discriminative region.Then,the discriminative region is cropped and enlarged to extract features,and the multi-lager feature fusion is used to fuse the multi-stage convolution features of the network.Finally,the global feature and the fused feature of each layer are fused to improve the expression ability of the feature,and the classification is based on the fused feature.The experimental results show that the algorithm achieves good results on CUB-200-2011 and Stanford Cars datasets.2.In order to improve the classification accuracy and training efficiency of the model,this thesis proposes an ensemble transfer learning algorithm.Firstly,the optimal finetuning depth of the model is determined by initial iterative training of the network.Then,in the process of training,we use the learning rate setting and the stochastic weight averaging algorithm to get multiple local model weights.Finally,the weights of each local model are ensembled,and average value is taken as the weight of the final prediction model.At the same time,a deep attention feature fusion network is proposed,which does not have the unique advantages of complex models and complex training processes.Firstly,the improved attention network is embedded in the multiple residual block of Res Net,and then the multi-layer convolutional features of the network are used for classification.The model is trained in the way of ensemble transfer learning algorithm,and the experiment shows that the ensemble transfer learning algorithm can steadily improve the classification accuracy of model without additional training,and the classification performance is almost the same as the model with higher complexity.