Domain Generalization for Fruit Classification Using EfficientNet-B0

Hello everyone,

I am working on a fruit classification problem using EfficientNet-B0, and I would appreciate some advice from anyone experienced with domain generalization / cross-dataset evaluation.

I have a custom fruit dataset collected under real-world conditions, with different backgrounds, lighting conditions, object sizes, and image acquisition environments. I am training EfficientNet-B0 on my custom dataset and then evaluating the trained model on the Fruit-360 dataset, which has a substantially different visual domain.

My main goal is to improve the model's generalization to the unseen Fruit-360 domain, rather than simply maximizing accuracy on the training-domain test set.

I would like to ask:

1. What preprocessing or normalization techniques would you recommend before training and cross-dataset evaluation?
2. Should the same preprocessing pipeline be applied to both the custom dataset and Fruit-360?
3. Would techniques such as stronger data augmentation, color normalization, background augmentation, MixStyle, domain randomization, or feature normalization be useful for this problem?
4. Would you recommend any modifications to EfficientNet-B0, such as attention mechanisms, domain-generalization modules, feature alignment, or other architectural changes?
5. Is there a recommended training strategy for improving cross-domain generalization?
6. Are there any MATLAB examples, papers, or implementations that you would recommend for this type of experiment?

The experimental setup is:

Source domain: My custom real-world fruit dataset

Target domain: Fruit-360

Model: EfficientNet-B0

Training: Source/custom dataset only

Evaluation: Target/Fruit-360 dataset without using its images for training

I would especially appreciate suggestions from anyone who has experience with domain generalization, domain shift, or cross-dataset fruit classification.

Thank you in advance for any advice or references.