Image classification requires the collection of a large number of images for model training and optimization， but the image collection process will bring noisy labels inevitably. To cope with this challenge， robust classification methods have emerged. The setting of hyperparameters in the current robust classification methods needs to be adjusted manually， which brings a lot of loss in human and material resources. Therefore， Meta Hyperparameter Adjuster （MHA） was proposed， which adopted a two-layer nested loop optimization method to learn noise-aware hyperparameter combinations adaptively， and a Meta-FPL （Feature Pseudo-Label adaptive learning algorithm based on Meta learning） algorithm was proposed too. In addition， in order to solve the problem that the backpropagation process in meta training phase consumes a large amount of GPU computing power， the Select Activation Metamodel Layer （SAML） strategy was proposed， which restricts the update of some metamodel layers by comparing sizes of the average gradient of the backpropagation and the meta-gradient in virtual training phase， which improves training efficiency of the model effectively. Experimental results on four benchmark datasets and one real dataset show that compared with MLC （Meta Label Correction for noisy label learning）， CTRR （ConTrastive RegulaRization） and Feature Pseudo Label （FPL） algorithms， Meta-FPL algorithm has a higher classification accuracy. In addition， after introducing SAML strategy， the training duration of the backpropagation process in the meta training phase was reduced by 79.52%. It can be seen that Meta-FPL algorithm can effectively improve the accuracy of classification testing in a shorter training time.