| In recent years,with the vigorous development of large-scale data,many deep learning algorithms have made key breakthroughs,but at the same time,some shortcomings have been exposed,such as the training process of deep learning models that require a large number of labeled samples to obtain a good classification result in a given task.However,in the actual research process,the labeled training sample data is often difficult to obtain.Few shot learning is a research topic that has attracted much attention in the field of computer vision,the purpose of which is to use only a small number of labeled samples,so that the classifier can quickly classify the input samples.In addition,the network depth of the few shot learning model is relatively small,which can effectively reduce the consumption of computing resources and better cope with tasks in actual scenarios.Based on this background,this dissertation studies the algorithm of metric learning in few-shot image classification,and the main contributions are as follows:Starting from the perspective of channel and position information in images,this dissertation solves the problem of few shot classification by constructing attention mechanisms from different angles to enable the network to better pay attention to these two types of sample information.The model adopts a mutual learning mechanism to mine the detailed information of the sample between the respective attention branches,thereby enhancing the model’s understanding of the fine-grained features in the sample.Experimental results show that integrating the multi-angle student sub-attention mechanism into the measurement network can greatly improve the learning ability of the model,so as to obtain good performance in few shot classification.In multi-branch mutual supervised learning,the network is prone to fall into the local optimal solution,which affects the performance of the model.In this dissertation,an online ensemble distillation method is proposed,which first integrates the probability distributions predicted by different branches,and then performs peak elimination processing on the low probability distributions to improve the quality of global soft targets.Next,we use the worst performing branch of the average probability distillation branch after peak elimination to improve its classification accuracy and further improve the generalization ability of the global model.In the process of continuous learning and updating of branch sub-networks,the entire network model will also be constantly updated.In the process of continuous learning and updating of branch sub-networks,the entire network model will also be constantly updated.There is a difference in the balance of the loss function between the mutual supervision task and the few-shot classification task between branches.In order to solve this problem,this dissertation conducts in-depth research on two aspects:optimizing task parameters and mining features.We design a network structure that can balance the differences between different tasks and improve the accuracy of the few-shot classification model,effectively avoiding the overfitting phenomenon in the classification task and greatly improving the accuracy of the model.Experimental results show that the task model proposed in this dissertation has higher classification accuracy and can mine sample information more fully. |
