Learning Adversarial Augmentation Policies for Robust Garlic Seedling Detection

In precision agriculture, ground-based monitoring of seedlings is often hindered by severe and spatially heterogeneous illumination, such as harsh shadows and over-exposure. Existing detection methods either rely on computationally expensive preprocessing modules or are evaluated under stable conditions (e.g., greenhouses or UAV imagery). This paper addresses the challenge of robustly detecting garlic seedlings in real-world, dynamic outdoor environments without adding overhead to the final inference process.

The authors introduce a new garlic seedling dataset captured via a ground-based platform. To improve detection robustness, they propose an adversarial augmentation policy learning framework. This framework uses a policy agent that generates input-conditioned augmentations designed to challenge the detector. The detector and the policy agent are optimized in an alternating fashion: the policy agent learns to create difficult but structurally realistic training samples, while the detector learns to maintain high performance despite these perturbations. A structural penalty is included in the reward function to ensure that the augmented images remain representative of real-world garlic seedlings.

The proposed framework significantly improves detection performance compared to baseline models. Specifically, it achieved an AP50 of 91.6%, outperforming the previous best-performing method by 0.2 percentage points. Furthermore, the model demonstrated superior performance in downstream missing seedling localization, achieving a 75.0% precision and a 67.0% F1-score—improvements of 4.8 and 2.0 percentage points over the baseline, respectively. Because the adversarial augmentation is only applied during training, the final model maintains its original inference speed, making it highly suitable for practical, real-time agricultural monitoring.