Your question does not follow StackOverFlow Community Guidelines on how to post a question. Specifically, your question is non coding question, and is asking for recommendations which is generally frowned upon. But anyway, let me help you :)

First thing, I could think of from your question is sheer quantity of images that you have in your dataset. I would recommend you to collect more good quality images to improve your model accuracy and performance.

If, you can't get hold of more images, you can apply Data Argumentation Techniques like

1. Geometric Transformations: Try to rotate/scale/flipping your images.
2. Color and Brightness Adjustments: Play with your dataset's color and brightness levels.

This will significantly, increase the number of images. The next important thing is to make sure that you have proper Data distribution. That is, if your classes are skewed towards one class, model performance will degrade. Make sure you use techniques like Oversampling to minority class. Ensure that the train and validation sets have a similar class distribution to avoid skewed evaluation.

Last piece of advise, play with different models and constantly hypertune your model params.