Research On Complex Small Sample Image Classification Based On Capsule Network

Small-sample image classification is an important task in the field of computer vision,and most real-life application scenarios have the problem of small sample data size,which makes this field receive extensive attention from scholars at home and abroad.In this thesis,based on the capsule network model,the classification model of fused capsule network and Darknet is proposed for relatively complex small sample datasets without noise,and the model of fused capsule network and deep residual shrinkage network is proposed for the classification of complex small sample datasets with noise,and the effectiveness of the proposed fusion model is verified by experiments.The specific work in this thesis is as follows:(1)A classification model that incorporates bilinear Darknet into a capsule network is proposed for the classification of relatively complex small-sample datasets.Firstly,Darknet is improved into a bilinear structure,and the deep feature extractor uses a 3*3 convolutional kernel to extract the deep detail features of the image,and the shallow feature extractor uses a5*5 convolutional kernel to capture the long-range edge features of the image,and the deep and shallow features of the image are extracted simultaneously by the deep feature extractor and the shallow feature extractor for fusion to increase the key features of the image.Secondly,the features are fused by Secondly,the features are characterized as vectors through capsule units to increase the spatial information of key features,and the lower layer capsules are directed to activate the higher layer capsules for feature transfer to avoid the loss of effective information during feature transfer.Finally,the L₂regularization term of the weights of each layer of the network is added to the loss function of the capsule network to make the network weight values smoother and prevent the problem of overfitting caused by too much weight of a certain feature.The experiments show that the classification accuracy of the fusion model is significantly higher than that of convolutional neural networks and capsule networks such as Res Net and Xception in the classification task of small sample datasets with relatively complex data without noise,and the model classification accuracy is significantly improved.(2)For the classification of complex small-sample datasets with noise,we propose a classification model that incorporates the deep residual systolic network into the capsule network.First,the deep residual systolic network is improved into a bilinear structure,and the residual connection of the shallow feature extractor is eliminated to avoid the introduction of image noise in the shallow features,and the noise in each channel of the deep and shallow features is effectively eliminated by the residual systolic module to reduce the redundant information of the features.Then the deep features are fused with the shallow features to increase the key features of the image;then the capsule unit is used to increases the spatial information of key features,while avoiding information loss during feature transfer.Finally,focal loss is used as the loss function of the model to solve the problem that some categories are difficult to distinguish and effectively improve the classification accuracy of the model.The experimental results show that the classification accuracy of the fused model is improved to a certain extent in the classification task of small-sample household waste images with noise.