| Deep learning method has made great progress in the field of image classification,but the labeled data required by the algorithm is massive.When encountering tasks where the available data is scarce,the only data is difficult to support the existing methods to complete sufficient networks training.Therefore,how to realize image recognition with few labeled samples has become a research hotspot in the field of computer vision in recent years.More and more studies have shown that the metric-based few-shot learning method has the advantages of simple model and high efficiency.In this paper,we study the problems of few-shot learning,such as insufficient labeled samples,easy model over-fitting and poor generalization ability,and proposes effective solutions and establishes the corresponding network model.Metric-based few-shot learning method can help deep learning method make better use of prior knowledge through feature extraction and metric learning in task space,but this kind of method is restricted by the quality of feature extraction and the choice of measurement methods.To solve this problem,this paper designs a few-shot learning model which is based on the architecture idea of Siamese Network,integrates the main parts of the Capsule Network in the embedding module.By using the dynamic routing between capsules,the internal relationship between the part and the whole is encoded,which helps learning to obtain more advanced feature;In the metric module,the obtained feature vector is calculated by L1 distance and the result is mapped to the similarity between [0,1] through two fully connected layers combined with the Sigmoid activation function.The model does not uses a single convolutional neural network in the feature extraction module,which can overcome the shortcomings of CNN translation invariance,preserve the spatial relationship between features,and help to better solve the common intra-class variation in few samples.The experiment was carried out on the low-complexity few-shot data set Omniglot,and compared with the existing methods.The experimental results show that the method has the advantages of few parameters and high accuracy.The existing metric-based few-shot learning methods usually optimize the model directly under the meta-learning framework,but the feature representation obtained directly by meta learning strategy has weak transitivity,which limits the classification ability of the metric method.To solve this problem,this paper proposes a model structure,which includes two parts: feature mapping module and piecewise metric module.Firstly,the classifier based on bilinear CNN is trained in advance on the base class data and transfer it to the feature extraction module of meta-learning;In the meta-learning stage,the extracted features are divided into multiple segments to calculate the cosine similarity to predict the category.Effective pre-training parameters,more powerful feature extraction structure combined with segmentation metric methods help the model to achieve better classification performance when facing complex scene data sets Mini-Image Net. |
