Few-shot Image Classification Based On CNN

Few-shot image classification can classify image only based on the minor number of samples.Compared with image classification methods based on numerous image data,few-shot image classification has the characteristics of the smaller amount of data,more kinds of classification categories,and much difficulty of modeling and learning the deep network.However,the transfer learning method and meta learning method,based on convolutional neural network(CNN),have made great process in few-shot image learning classification in recent years.In the optimization view of the network modeling and algorithm,they share the same features of different task areas to guide,and optimize network learning based on previous network experience to improving the learning performance,where they avoid repeating the learning process to accelerate the training speed.Our paper study few-shot learning based on transfer learning and meta-learning of CNN,the specific contents are as follows:(1)A Model Transfer-Fusion Network(MT-Fusion Net)is proposed in this paper based on transfer learning,which transfers the parameters of the deep network model trained by the multi-sample dataset,it can also extract feature form the minor number of target domain dataset to address the over-fitting problem effectively.In order to increase the feature description of small sample dataset,this paper introduces Inception V3 and Res Net18 models to transfer trained parameter and aggregate models,which extracts multi-scale features to describe the deep and shallow semantic features of the target data image.Then adaptive adjustment layers are utilized to calculate the distance between the source domain and the target domain,which can fulfil the requirements of the target data classification.In comparison to other networks without transfer and fusion,MT-Fusion Net can improve the generalization of the model effectively,and promote the classification performance to some extent.(2)A Node embedding and Edge labeling Graph Neural Network(NEGNN)is formulated based on the meta learning,where the feature coding network describes the node features and edge features,the Edge-labeling Graph Neural Network EGNN updates the topological relationship between nodes and edge features,and the weight of meta-tasks adjust according to the loss of each meta-task adaptively.However,the node embedding network ignores the correlation between features through some shallow CNNs stacking,this paper combines SE-Res Net and Inception module into a novel node embedding network,which can combine multiple scale features,and relabel the features by adjusting the weight coefficient of each feature channel adaptively.Besides,the Dense-Block is applied in the node update network to transfer the feature map to all subsequent layers,where the edge characteristics of dominant markers are combined in the neighborhood to enhance intra-cluster similarity and the inter-cluster dissimilarity.Compared with the existing meta learning method,the proposed method can obtain a well-generalizable model and improve the classification effect effectively.