Research On Fine-grained Classification Based On Weakly Supervised Learning

Fine-grained visual classification is a challenging task due to the large intra-class variance and small inter-class variance,However,the current methods based on strong supervised information rely too much on annotated information,and the acquisition cost of annotated information is very expensive.As a result,It is difficult to apply in real life.Therefore,the research of fine-grained classification based on weak supervision information is of great practical significance.To this end,this thesis proposes two methods based on weak supervision information from the perspectives of global local area and feature interaction respectively.The main work is as follows:(1)Local regions play an important role in fine-grained classification,but how to better locate the discriminating local regions and better include more local features?Therefore,this thesis proposes a multi-module fine-grained image classification method based on the fusion of global and local features.The main idea is that in order to integrate more local regions with discriminating power,we adopt two branches:coarse classification and fine classification,in which the basic model is used as the coarse classification branch,and the mapping between randomly selected local features and original images is used as the input of the fine classification branch,so as to continuously optimize the classification effect of the whole model.In addition,in order to reduce the noise of random selection,a central loss is used to constantly shorten the distance between the model and the global feature,and the way of attention fusion makes the model pay more attention to the region with discrimination.At the same time,the validity of the proposed method is fully proved by experiments on three datasets in the fine-grained domain.(2)Model based on bilinear pooling has shown good results in fine-grained domain,But it ignores the semantic relationship between important layers in the process of propagation,and does not attempt higher order pooling method.Therefore,this thesis proposes a high-order statistical fine-grained image classification method based on bilinear.The main idea is to calculate the third-order interaction between the three feature layers by cross-level tri-linear pooling method,and then to merge the different combinations of third-order interaction features into the final representation,and by combining with the attention mechanism and measurement learning method,to obtain the superposition of the improved effect.The proposed method is a great improvement over the bilinear method and has achieved the best classification results in the field on the three publicly available datasets in the fine-grained domain.