Fine-Grained Image Recognition Based On Feature Encoding

With the rapid development of artificial intelligence technology,image recognition has been widely used in daily life.As an important branch in the research of image recognition,fine-grained image recognition aims to identify and classify multiple subcategories in the same large category.Because different sub-categories have large intra-class differences and small inter-class differences,distinguishing multiple sub-categories usually requires expertlevel domain knowledge.The current methods mainly use feature coding techniques to learn the discriminative attributes of fine-grained objects.However,the current feature coding methods have the problems of high computational complexity,lack of interpretability,and difficulty in effectively using complementary information between different convolutional features in neural networks.In view of the above problems,this paper proposes new feature encoding algorithms,which reduce the computational complexity of the network model,and use the complementary information between different convolutional features to effectively improve the performance of fine-grained image recognition.Aiming at the problem that the existing feature coding method is difficult to effectively use the complementary information between different convolutional features in the neural network,this paper proposes a hierarchical bilinear pooling model(HBP).Using the crosslayer bilinear coding algorithm,the neural network is guided to learn the interactive information between different convolutional features,which effectively improves the neural network's ability to express features.Experimental results show that,compared with a variety of classical fine-grained recognition methods,the recognition accuracy of the model in the three standard datasets has been significantly improved.In view of the problem of high computational complexity of feature encoding methods,this paper proposes a deep hypersphere embedding model(DHE),which uses feature phase information to encode the attributes of fine-grained objects,effectively reducing the computational complexity of the network model.And the feature phase can better characterize the subtle semantic difference information between fine-grained objects,which improves the learning ability of the neural network.Aiming at the lack of interpretability of feature coding methods,this paper proposes phase activation maps and category contribution maps to intuitively explain the coding learning process of subtle differences between fine-grained objects by deep hypersphere embedding networks.The phase activation map can highlight the object area in the input image and has a significant object detection ability;the category contribution map quantitatively analyzes the classification decision of the neural network and intuitively shows the discriminative properties of finegrained objects.Experimental results show that the deep hypersphere embedding model proposed in this paper has higher computational efficiency than many classical fine-grained identification methods,and can accurately locate the discriminative attribute area of finegrained objects.