A method for improving the robustness of neural network for aerial image matching

Abstract

The subject of study in this article is neural network–based methods for aerial image matching, which are widely used in navigation, localization, and mapping tasks. A key challenge lies in the sensitivity of such methods to visual disturbances and scene novelty caused by shadows, illumination changes, and terrain variability, which limits their robustness in real-world conditions—particularly under constrained computational resources. This paper investigates an approach to enhancing the robustness and cross-domain generalization of computationally efficient aerial image matching models by combining adversarial procedural noise with a modified activation function. The goal is to develop a training methodology that simultaneously increases the resilience of models to perturbations and improves their transferability across different observation domains. The research objectives are as follows: (1) to analyze existing methods for improving the robustness of neural networks and assess their applicability to aerial image matching tasks; (2) to develop a training approach incorporating the synthesis of adversarial procedural noises (Perlin, Gabor, Worley) and the replacement of the standard ReLU with a hybrid activation function, LeakyReLU6, which constrains activation amplitudes and reduces sensitivity to local disturbances; (3) to conduct a comprehensive experimental evaluation of detector-based architectures (SuperPoint + LightGlue) and detector-free models (EfficientLoFTR) using the Aerial Image Matching Benchmark dataset; (4) to verify cross-domain generalization on the HPatches dataset; and (5) to perform an ablation study to isolate the contribution of each component. Results. The proposed methodology achieved over a 4.2% absolute improvement in AUC@1px matching accuracy on noisy test data for both classes of models. The ablation study revealed a synergistic effect from combining procedural noise with LeakyReLU6 — in particular, for the SuperPoint + LightGlue combination, improvements reached +3.0% AUC@1px and +2.7% AUC@3px, while for EfficientLoFTR, gains of +2.2% and +2.6% were observed, respectively. Additionally, testing on HPatches showed a 0.83% smaller performance drop compared to baseline training, confirming a higher level of cross-domain generalization. Conclusions. The proposed approach enhances the noise robustness and cross-domain generalization of feature-matching models and can be easily extended to various neural network architectures. Future work will focus on investigating the influence of procedural noise hyperparameters, applying meta-learning on corrupted data, and introducing architectural improvements to further strengthen resilience and robustness. Scientific novelty. The novelty of this work lies in the first integration of adversarial learning with procedural noise and a bounded activation function (LeakyReLU6, using the Straight-Through Estimator (STE) in the backward pass), which produced a synergistic effect that improved the robustness and generalization of aerial image matching models without a significant increase in computational cost.