RESEARCH · RESEARCH · #2714
arXiv:2610.35984v1 — data-generation strategies improve EfficientNetV2B0 classification of olive tree diseases
This arXiv preprint (2610.35984v1) evaluates a comprehensive data-generation pipeline — traditional augmentation, CutMix, MixUp, and synthetic images from a DCGAN — used in three distribution scenarios (balancing classes, increasing per-class samples, and expanding while keeping classes balanced). The authors report substantial accuracy gains when training a pre-trained and fine-tuned EfficientNetV2B0 on the augmented olive tree disease dataset, arguing the approach improves dataset quality for plant disease detection.
KEY POINTS
- This arXiv preprint (2610.35984v1) evaluates a comprehensive data-generation pipeline — traditional augmentation, CutMix, MixUp, and synthetic images from a DCGAN — used in three distribution scenarios (balancing classes, increasing per-class samples, and expanding while keeping classes balanced).
- The authors report substantial accuracy gains when training a pre-trained and fine-tuned EfficientNetV2B0 on the augmented olive tree disease dataset, arguing the approach improves dataset quality for plant disease detection.
- The paper demonstrates that mixing traditional augmentation with GAN-based synthesis and controlled class-distribution strategies can mitigate data scarcity and imbalance, materially improving deep‑learning accuracy for plant disease diagnosis.
WHY IT MATTERS
The paper demonstrates that mixing traditional augmentation with GAN-based synthesis and controlled class-distribution strategies can mitigate data scarcity and imbalance, materially improving deep‑learning accuracy for plant disease diagnosis.