The selection of a suitable navigation area is pivotal in aircraft scene matching guidance technology. This study addresses the challenge of identifying suitable reference image ranges for precise scene matching, which is crucial for enhancing aircraft positioning accuracy. Traditional methods for image matchability analysis are often limited by their reliance on manual feature parameter design and threshold-based filtering, resulting in suboptimal accuracy and efficiency. This paper proposes a novel network architecture for selecting suitable navigation areas using image Matching Level Segmentation (MLSNet). The approach involves two key innovations: a method for generating segmentation labels that quantify matchability levels and an end-to-end network architecture for rapid and precise prediction of reference image matchability segmentation maps. The network includes two core modules: the saliency analysis module uses multi-layer convolutional networks to accurately detect image saliency features across various levels and scales; the multidimensional attention module utilizes attention mechanisms to focus on feature channels and spatial neighborhood scenes to assess the image’s matchability. Our method was rigorously tested on an extensive collection of remote sensing images, where it was benchmarked against a range of both traditional and cutting-edge deep learning methods. The findings indicate that MLSNet is significantly superior to traditional methods in accuracy and efficiency of matchability analysis, and is also relatively ahead of state-of-the-art deep learning models.

The navigation system plays a pivotal role in guiding aircraft along designated routes, ensuring precise and punctual arrival at destinations. The integration of scene matching with an inertial navigation system enhances the capability of providing a dependable guarantee for successful accomplishment of flight missions. Nonetheless, assuring reliability in scene matching encounters significant challenges in areas characterized by repetitive or weak textures. To tackle these challenges, we propose a novel method to assess the reliability of scene matching based on the distinctive characteristics of correlation peaks. The proposed method leverages the fact that the similarity of the optimal matching result is significantly higher than that of the surrounding area, and three novel indicators (e.g., relative height, slope of a correlation peak, and ratio of a sub peak to the main peak) are determined to conjointly evaluate the reliability of scene matching. The proposed method entails matching a real-time image with a reference image to generate a correlation surface. A correlation peak is then obtained by extracting the portion of the correlation surface exhibiting a significant gradient. Additionally, the matching reliability is determined by considering the relative height, slope, and ratio of the peak collectively. Exhaustive experimental results with two sets of data demonstrate that the proposed method significantly outperforms traditional approaches in terms of precision, recall, and F1-score. These experiments also establish the efficacy of the proposed method in achieving reliable matching in challenging environments characterized by repetitive and weak textures. This enhancement holds the potential to significantly elevate scene-matching-based navigation.