Research On Fine-Grained Image Classification Based On Deep Learning

Generally speaking,in the process of studying computer vision,the traditional image classification took objects of different basic categories as a foothold.Up to now,image classification was no longer used only for large category classification in practical application scenarios,but sought for more meticulous and accurate classification on the basis of traditional classification.Therefore,fine-grained image classification has become the key research direction in the field of computer vision.In most case,the assignment of fine-grained image classification studied different subcategories which have belonged to the same underlying category.The challenge of this classification task was to address the characteristics of large inter class differences and small intra class differences.In addition,the complexity of the background,the different light rays and the difference in the shooting angles also brought the difficulty of the classification task to different degrees.According to the research prospect of fine-grained image classification and the above interference factors and task difficulty,this thesis focuses on fine-grained image classification based on deep learning,and the main research and work content is as follows:(1)By analyzing and comparing the principles and advantages of existing FGIC network algorithms based on Deep Learning,NTS-Net,which does not rely on additional manual annotation of information,performs localization and classification tasks simultaneously,and does not require strict experimental equipment,was selected as the base network.CUB-200-2011,Stanford Cars and FGVC-Aircraft are selected as the reference datasets.Then,through analysis and research,it is found that the attention mechanism used for fine-grained classification should contain the following characteristics: Firstly,the detected feature of the object location should be relatively evenly distributed on the object,so as to extract the unrelated feature;Secondly,the feature of each part of the object should be distinguished separately to separate different classes of objects.Thirdly,the part extractor should be lightweight in order to expand its scale for practical application.Therefore,based on the infrastructure of the network,an improved attention mechanism module MA-FPN is presented,and from this module,the first improved network in this thesis,MA-FPN-Net is presented.The experiments showed that MA-FPN-Net improved the classification accuracy by 0.2% and 0.1% on bird and aircraft datasets,respectively,compared with the basic network.(2)This thesis researches on fine-grained image classification tasks,so feature extraction at different scales of input images has become very important.Therefore,this thesis constructs a novel multi-scale feature fusion module SgPS by improving and combining several classic multi-scale feature fusion modules,and proposes the second improved network SgPS-Net.Through experiments,it can be concluded that the classification accuracy of SgPS-Net has been improved by 0.4%,0.2%,and 0.4%compared to the basic network on three datasets.(3)This thesis proposes the third improved network,AMFF-Net,based on the attention mechanism module and feature fusion module.Then,compared with the basic network,the classification accuracy of the AMFF-Net proposed in this article was improved by 0.8%,0.9%,and 1.1% on three datasets.Meanwhile,the AMFF-Net network algorithm has good robustness and can be applied to most fine-grained image classification tasks.(4)Today,significant breakthroughs have been made in the field of Deep Learning,which shows that binding energy.Thus,it can be concluded that data is the most effective fine-grained classification method.Therefore,based on the AMFF-Net,this thesis expands the original training set by adding sub datasets Web bird,Web car,and Web air craft from the mixed dataset WebFG-496 of the network.Through experiments,it has been shown that the classification accuracy of the AMFF-Net on three datasets has been improved by 0.8%,0.4%,and 1.2%,compared to the AMFF-Net without increasing the training set.