Adaptive Loss Function And Label Smoothing Based On Prototype Network For Few-shot Learning

In recent years,deep learning has received extensive attention,but it has also revealed some shortcomings,that is,the model usually needs to rely on a large amount of data to obtain better performance.However,sometimes it is difficult to obtain this amount of data,so few-shot learning has gradually become a key research topic in the field of computer vision.Based on this background,this dissertation studies the prototype network algorithm based on metric learning in few-shot learning,and conducts experiments on international public datasets.The main contents are as following.(1)Aiming at the unbalanced contribution of the sample mean feature to the prototype representation in the measurement module,a method of redefining the prototype using weights and transduction is proposed.Combining the method of feature transformation to obtain more sample feature points,then according to the degree of these feature points deviating from the sample dense area,weights of different sizes are given,and the weighted prototype is improved by the query set instance transduction method,so that the prototype can be determined.The reliability of the prototype representation is improved by dynamically changing according to the contribution size of the sample features.(2)Aiming at the problem that the cross-entropy loss function can only measure the distribution of two samples and cannot optimize the parameters,a hybrid loss function algorithm that can use triple samples to optimize the hyperparameters is proposed.The algorithm integrates the Triplet Loss function,and obtains the adaptive boundary value hyperparameters through matrix transformation,logarithmic calculation,etc.At the same time,the loss is added to the cross-entropy loss to obtain a mixed loss function,so that the optimization of the model loss can be dynamic adjustment according to how the task changes within different images improves the generalization ability.(3)In view of the problem that the labels of the dataset samples are wrong and not accurate enough,a method of label smoothing and regularization is induced into the mixed loss function,and the regularization is used to add an error rate to the one-hot label for smoothing,and then added to the calculation of the boundary value of the hybrid loss function,the influence of the wrong label on the model loss update is minimized,and the accuracy of the loss calculation of the hybrid loss function is improved.Based on the existing prototype network,this dissertation strengthens the feature extraction network,and then redefines the prototype for the measurement module.From these,an adaptive loss function calculation method is proposed and label smoothing is added,and regularization is also used to weaken the influence of wrong labels,which can effectively improve the recognition performance of the model.