In order to improve the performance of image classification models in practical applications, the main issue is to address the possibility that the performance of the model may vary under different lighting conditions. Should we enhance the generalization ability of the training set during training, or adjust the brightness of the input image before making predictions when using the model?