What if the accuracy of some categories in the results obtained by the classification model is obviously on the low side?

such as the title.
what I"m doing is language classification of picture text

the following figure shows the results of my experiment

I would like to ask what measures can be taken to further improve accuracy?


some of my methods:

  1. check whether the sample number of the corresponding category in the training set is relatively small, and try to expand the sample set.
  2. if you can't enlarge the sample set, train multiple models, balance the number of samples in each model as much as possible, and then merge the results with the idea of decision tree
  3. .
  4. is it true that the content of the sample dataset is recognized by human eyes with low accuracy?

  • coarse adjustment: vgg hyperparameter adjustment. Grid search for a moment
  • fine tuning: consider typing out the confusion matrix CM to see which results are misclassified to which categories with low accuracy, and whether these results have any commonalities. Make adjustments during the preprocessing phase.
Menu