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

With the vigorous development of the Internet,images have gradually become the mainstream media of information dissemination.How to extract and integrate the corresponding feature representation and hidden information from images efficiently and stably to assist various tasks in the era of artificial intelligence is a key problem in the current industry and academia.As the basic task of image data processing,image classification is of great significance for the analysis and research of image data.Finegrained image classification is a further subdivision of the image classification task.Its main task is to distinguish different subcategories under the same category.Due to the very subtle inter-class differences between subcategories,fine-grained image classification is more precise than ordinary image classification tasks.The strongly supervised fine-grained classification algorithm needs to use additional supervision information(such as anchor points,annotation boxes,etc.)to accurately locate the target area and improve the recognition accuracy,but it is greatly limited in the real scene due to the expensive annotation cost.The weakly supervised fine-grained classification algorithm only needs the category label of objects,which has been widely used in industry and academia.Both the fine-grained image classification algorithms proposed in this paper are based on weakly supervised learning and optimized for the disadvantages of existing methods.The main research contents are as follows:(1)Aiming at the limitation that the attention-based method in the current weakly supervised fine-grained image classification task is not flexible enough,this paper proposes a spatial self-attention network(S-SANet)to obtain the highly available feature information of the relationship between compressed features along the spatial dimension and take it as fine-grained features,so as to alleviate the performance bottleneck of the convolutional neural network,It includes a self-attention distillation module,which is used to transfer the knowledge information output by the spatial self-attention module to the coarse-grained main features.Experiments show that SSANet has very competitive performance on three fine-grained image public datasets named CUB,dogs,and cars.(2)Aiming at the common problem of data noise and target conflict in weak supervised fine-grained classification methods,a cross integrated distillation model based on data enhancement(CEKD)is proposed in this paper.Although S-SANet uses data enhancement technology to alleviate the over-fitting problem,it does not consider the side effect caused by the introduction of enhancement operation,that is,data noise.Aiming at the problem of data noise,a cross distillation module is proposed to improve the robustness of the model against data noise;Aiming at the problem of target conflict,an optimized collaborative integration method is proposed to dynamically integrate the output Logits,so as to alleviate the performance gap between teacher-student networks.Through a large number of comparative experiments on three public fine-grained image datasets named CUB,cars,and aircraft,it shows that the recognition rate is better than other benchmark models,and the positioning of key parts is more accurate.