Analysis And Implementation Of Fine-grained Image Retrieval Algorithm Based On Deep Convolution Features

With the advancement of Internet,the image library has gradually expanded and perfected.Image processing has become a central issue in research field,and image retrieval has gained more and more attention as a key part of image processing.Fine-grained image retrieval is a popular trend in recent years.It draws on the research experience of traditional image retrieval,and in the beginning,most of the studies ignored the era of hand-crafted features that were complex and professional and could not obtain excellent retrieval results,most of the studies used deep features obtained by convolution.Each class of fine-grained images belongs to a sub-class of a meta-class,which make retrieval more difficult.Classical convolutional neural network loss functions can achieve good retrieval results on traditional image retrieval tasks,but their performance in fine-grained image retrieval tasks is unsatisfactory.In the training phase,on the one hand,the inner classes of the fine-grained images have great differences and over-aggregating the images within the classes with such great differences will not obtain a good embedding space.Maintaining the intra-class differences has become one of the problems that need to be studied.On the other hand,the traditional pair-based loss function selects sample pairs within a batch,but the samples within a batch are limited and cannot represent the global information of each category,and the random combination of sample pairs will exponentially increase the computational cost of loss,which also becomes a difficulty that needs to be improved.In the testing phase of the neural network,the process of aggregating the threedimensional convolutional features into one-dimensional image feature descriptors cannot avoid the loss of information because of the dimensionality reduction operation.The way on moving forward the retrieval accuracy is to obtain a more representative image feature descriptor in one-dimensional space.This thesis proposes a unified fine-grained retrieval framework,which improves and innovates the loss function of neural networks and weakly supervised aggregation of convolutional features to solve the above problems.1.This thesis designs a unified image retrieval framework,innovating in the network training and feature aggregation respectively,to solve the fine-grained image retrieval task that is more challenging than traditional images.2.In the neural network training stage,this thesis designs a loss function based on adaptive multi-class centers and applies this loss to the Res Net50 deep convolutional neural network end-to-end.The adaptive multi-class center loss function maintains the intra-class difference while optimizing the similarity between samples,an embedding space more suitable for fine-grained image retrieval tasks is obtained.3.In the testing phase,this thesis proposes a weakly supervised channel-weighted feature aggregation algorithm.First,the three-dimensional convolution features of the image are obtained through the Res Net50 neural network by the adaptive multi-class center loss function,and then the one-dimensional image feature descriptor is obtained through the channel weighted feature aggregation algorithm,and the weight coefficient of important channels is increased to obtain a more representative image feature descriptor.Use it for subsequent similarity measures to obtain a more accurate retrieval ranking.4.This thesis conducts multiple experiments on several representative fine-grained retrieval datasets.The effectiveness of adaptive multi-class center loss function and weakly supervised channel feature aggregation algorithm are analyzed respectively.The comparative experimental outcomes indicate that both the adaptive multi-class center loss function and the weakly supervised channel feature aggregation algorithm designed in this thesis can improve the accuracy of fine-grained image retrieval,and they can be unified into the fine-grained retrieval framework and promote each other to achieve better results of retrieval.Compared with the previous representative finegrained image retrieval methods,the accuracy has been essentially enhanced.On five classic fine-grained image datasets,Stanford Dog,FGVC aircraft,Oxford Pets,CUB-200-2011 and CARS196,better retrieval results than previous fine-grained image retrieval methods have been obtained.