Research On Fine-grained Image Recognition Based On Convolutional Neural Network Optimization

With the continuous progress of the Internet,all kinds of images have become the main data media and information carriers in the Internet era.How to quickly and efficiently extract the key information of the images and accurately identify the images is the main problem at present.Fine-grained image recognition is different from traditional image recognition.The former distinguishes different subclasses of a certain type of object.Since fine-grained images have the characteristics of small differences between subclasses and large differences within subclasses,fine-grained image recognition has become a challenging research task in the field of computer vision.Bilinear convolutional neural networks can be trained end-to-end with only image labels,thus becoming the mainstream model for fine-grained image recognition.This paper will focus on the fine-grained image recognition model,and propose a fine-grained image recognition model based on CNN optimization.This model solves two problems,one is how to effectively acquire the shallow features of the network,and fuse and identify the shallow surface features and the deep abstract features.The second is how to train a model with good performance with the help of less existing labeled sample information,or with data sample information only labeled with image labels.In order to ensure that useful shallow features can be obtained through fine-grained image recognition models and fused with deep features,a fine-grained image recognition model based on feature fusion is proposed in this paper.Considering that the deep features of the network and the shallow features have different effects on image recognition,this paper improves the bilinear convolutional neural network.A high-resolution network is used to construct a bilinear network to perform multiple feature fusions on the network,thereby improving the recognition accuracy of the network.In this paper,the constructed network is tested on CUB-200-2011 dataset and Stanford Cars dataset,and compared with the bilinear convolutional neural network proposed by Lin and three common fine-grained image recognition models,which verifies the effectiveness of the method.In order to ensure that the fine-grained image recognition model can have strong learning ability even with fewer training samples,this paper proposes a fine-grained image recognition model based on small sample training.Considering that there are few annotated data samples at present,the model proposes to use the prototype network to train fine-grained images with small samples.In this paper,the bilinear network constructed by the Mobile Net network and the Res Net-18 network is used as the embedding function of the prototype network,and the labeled samples of the test set are mapped to the embedding space and averaged to obtain the category prototype.Finally compute the feature vector of the validation set,the Euclidean distance from the class prototype,and classify it according to these.This method can quickly learn features from a small number of samples,and then classify image samples accurately.In this paper,the constructed network is tested on the CUB-200-2011 dataset and Stanford Cars dataset,and compared with three common small-sample learning models,the recognition accuracy is higher than other networks.It is verified that the model has the characteristics of high recognition accuracy and strong generalization ability.