How can I improve the performance of an image classification model so that the loss on the test set will reduce

I’m working on an image classification task and the metric for evaluating the model performance is the log loss I have used pretrained model like efficient-net b1 and resnet18 to do fine tunning.
I have as well perform data augmentation such as gaussianblur, but the least loss I have gotten on the test set is 0.4 which is greater than the bench mark

Please how can I improve the model performance?

Thanks in advance.

there are few benchmark datasets for image classification .
I would suggest you to use them and evaluate your model and compare your model with the benchmark model for that dataset.

Hope it helps.