A Comparative Study of Error-Correcting Codes for Multi-Cell Upsets in Memories: CLC and OPCoSA

Abstract

UAV-to-satellite cross-view geo-localization is an important technology for autonomous navigation in GNSS-denied environments. However, large differences in viewpoint, scale, and distortion between UAV and satellite images make feature matching difficult. This paper presents a dual-branch deep learning framework and compares two backbone architectures, ResNet-18 and MLP Mixer, under the same bidirectional InfoNCE training objective. Experiments on the UAV-Visloc dataset show that MLP-Mixer achieves better retrieval performance than ResNet-18, reaching Recall@1 of 74.09% versus 72.40%, with consistent improvements at Recall@5 and Recall@10. Ablation results further show that independent branches are important for handling the domain gap between UAV and satellite imagery. The results indicate that pure MLP architectures have strong potential for cross-view geo-localization when combined with contrastive learning.