Research On Weakly Supervised Algorithm For Fine-grained Image Classification Based On Deep Learning

With the rapid advent of the era of intelligence,the task of classifying images in the computer vision field is not only limited to identifying large categories of objects,but more detailed sub-categories of images of the same category are required.Therefore,this paper uses the deep learning method to identify the mainstream fine-grained image classification data set.The main research contents of this paper are as follows:1.Analyze the data set used in mainstream fine-grained image classification tasks and make reasonable divisions to facilitate subsequent research work.2.Based on the method of applying Weakly Supervised Data Augmentation Network(WS-DAN),introduce the mainstream feature extraction networks VGG,Inception and Res Net,and adopt the method of transfer learning to determine Res Net 152 as the feature extraction network.3.The global features are extracted through the Res Net 152 pre-training model,and then the deep separable convolution is used to extract part of the attention region on the feature region.4.A fine-grained classification algorithm based on Attention-Attention Bilinear Pooling(AABP)is proposed,which can effectively distinguish small differences between classifications and reduce the interference of background factors.5.The work of integrating 2-4 will linearly fuse the extracted attention regions in the AABP algorithm to improve the accuracy of the final classification with single-mode and multi-state data augmentation.Finally,this paper conducts an experimental comparative analysis on the mainstream algorithms.The experimental results show that the method has a top1 accuracy rate of89.9% on the CUB-200-2011,a top1 accuracy rate of 90.0% on the Stanford Dogs,a top1 accuracy rate of 95.2% on the Stanford Cars,and a top1 accuracy rate of 94.1% on the FGVC-Aircraft.Compared with the experimental results of mainstream network models,it has improved.