Abstract: ‍ ‍When using image data sets to train neural network classification model， a large number of accurately labeled image data sets are needed， but the actual image data sets often contains a large number of mislabeled images， which is not conducive to the training of accurate neural network classification model. However， the production of annotated accurate data sets requires a lot of time and labor costs. Therefore， this paper proposes a classification framework based on inaccurate image data cleaning. Experimental results on natural cat and dog images show that the classification accuracy of the model with cleaning is improved and the loss value of the loss function is decreased. In the study of the relationship between the proportion of mislabeled images in the data set and the classification accuracy， it is found that the deeper neural network has certain robustness to the error images in the data set， but when the proportion of tag noise images in the image data set is high， The introduction of cleaning makes the shallow neural network classification model achieve the same classification effect as the deeper neural network classification model， and the shallow neural network classification model has faster operation speed. This paper provides a new sightseeing into constructing fast and accurate classification model.