Research On The Application Of Deep Metric Learning In Fine-Grained Image Classification

At present,deep learning has become the preferred method to solve the task of image classification,and more and more deep learning models have come out.However,with the progress of research,researchers found that there is a kind of image classification task that can not be effectively solved by deep learning.This kind of task is fine-grained image classification task,which also belongs to one of image classification tasks,but the discriminant difference in the image is more subtle.The images of general image classification tasks belong to different types,but the discriminant samples in fine-grained image classification can belong to the same kind.To solve this problem,this paper studies the common methods of ordinary image classification tasks,then distinguishes the similar regions between images according to the characteristics of fine-grained image classification tasks and combined with depth measurement learning,calculates the similarity of image features through measurement learning,and proposes a fine-grained image classification model based on depth measurement learning,The main work of this paper is following:Firstly,for the task of fine-grained image classification,this paper uses the common model of deep learning as the feature extraction network.Considering the characteristics of fine-grained image distinguishing regional subtle differences,firstly,the bilinear attention pooling method is used to generate the matrix of some feature regions.In order to further locate the same part with subtle differences in different images,It is proposed to use depth measurement to learn a specific loss function to calculate the measurement between feature matrices,so that the measurement distance between the same feature is closer and the measurement distance between different features is farther.Depth metric learning is generally used in open set classification tasks such as face recognition and fingerprint recognition.The number of categories is often large and the number of samples in the category is relatively small.The experimental results show that depth metric learning can also be applied to the field of fine-grained image classification.Compared with the fine-grained image classification model using only weak supervised learning,it has more accurate recognition ability for discriminant regions,so as to improve the accuracy of final fine-grained image classification.Secondly,for image classification task,convolutional neural network in deep learning can automatically learn features from large-scale data,but different network depth,network width and image resolution will affect the classification accuracy of the network.Efficientnet obtains the best basic model through composite parameter expansion and grid search.Combined with bilinear attention pooling,this paper improves the network structure model of efficientnet and improves the ability of model feature extraction.At the same time,inspired by the data enhancement method,the data enhancement based on attention guidance provides more diverse sample data for the model.The model can more easily obtain the sample characteristics and improve the generalization ability of the model.Experiments show that the fine-grained image classification model using Efficientnet has higher classification accuracy,but with the improvement of network parameters,the amount of computation also increases relatively.Finally,the fine-grained image classification model combined with improved Efficientnet and metric learning is used to experiment on the commonly used fine-grained image classification data set.The experiments show that the model can further improve the accuracy of fine-grained image classification,and can be successfully applied to insect recognition projects to obtain good results.