Abstract: In order to construct a longer term constraint in the process of visual positioning, and facilitate the establishment of globally consistent trajectory estimation for aircraft equipped with visual positioning equipment, a lightweight weakly supervised simultaneous localization and mapping (SLAM) loop closure detection algorithm is proposed based on contrastive learning. By establishing consistent global descriptors for images and using similarity measures to judge loops in trajectories, long term timing constraints are established. Considering the application on resource constrained platforms, an EfficientNet (efficient neural network) is used to achieve more efficient feature extraction, and the NSE (need squeeze and excitation) attention module is combined to improve the screening of effective data in the process of data dimensionality reduction. Global descriptor is integrated through a complete VLAD (vector of locally aggregated descriptors) layer. The network model can still have efficient recognition ability in environmental changes such as lighting conditions, viewing angles, and seasons. The experimental results show that, while maintaining a 2% difference in the TOP 5 recall index compared to the baseline model, the proposed method can effectively reduce model volume by 57%, training time by 35%, and improve execution efficiency by 48%, which is beneficial for deployment on resource constrained embedded platforms such as small drones

Key words: loop closure detection, lightweight, weakly supervised, comparative learning