Research And Application Of Fine-grained Image Classification Model Based On Cross-domain Fusion

Fine-grained image classification is a more refined classification of the category to which the image belongs,in other words,the subcategory is divided under the large categories to determine that the image belongs to a subcategory.In contrast to traditional image classification,which is mainly based on the classification of large categories,fine-grained image classification is characterized by large intra-class differences and small inter-class differences,which are very challenging,and because it has a wide range of applications in people's production and life,it has attracted a large number of scholars to conduct in-depth research on it.In this paper,firstly,some fine-grained image classification algorithms are studied.At present,most of the algorithms rely on depth learning techniques,by changing the convolutional neural network structure,adjusting the model parameters,optimizing the loss function,expanding the types of data sets and other ways to improve the accuracy of fine-grained image classification.Inspired by the process of human cognition,this paper will improve the accuracy of image classification by expanding the types of data sets,and research on fine-grained image classification model based on cross-domain fusion and its application.Cross-domain fusion combines image and text information to simulate human cognitive process in order to improve the accuracy of image classification.The images needed for fine-grained image classification are images of different subcategory under a certain category.In the process of data collection,the identification of sub-categories often requires certain professional knowledge,and the acquisition of images is difficult,few-shot learning method can be used to reduce the cost of data acquisition,so this paper will focus on few-shot learning based on cross-domain fusion fine-grained image classification algorithms and applications.The main research work and achievements of this thesis are as follows:(1)A cross-domain fusion mechanism combining image information with text information is proposed.The text information is transformed into word vector by GloVe algorithm,and the text vector is represented by its average value,then the text feature vector and the image feature vector are combined to generate a new class prototype and the ratio of the two is adjusted by the learning parameter ?c,so that the model can adaptively adjust the dependence on the image and the text information,finally,the distance between the query set sample vector and the new class prototype is calculated,and the final image classification result is obtained by generating the probability distribution of the distance using the Softmax function.(2)Based on the above cross-domain fusion mechanism,a fine-grained image classification model is constructed.In order to solve the problem of small sample size,ITCD-ProtoNets++(image and text cross-domain prototypical network)and ITCD-TADAM++(image and text cross-domain Tadam network)are constructed based on prototype network and task dependent adaptive metric network in few-shot learning.(3)In order to verify the validity of fine-grained image classification model for cross-domain Fusion,experiments were conducted on CUB-200-2011 and Flowers data sets.The results show that the cross-domain fusion model can effectively improve the accuracy of fine-grained image classification based on few-shot,the influence of the number of sentences on the result of image classification is analyzed.(4)The ITCD-TADAM++model with higher classification accuracy is applied in this paper.The method is applied to the field of bird recognition,and the bird recognition system is realized,and the application prospect of the method is analyzed through application scenarios.