Research on Remote Sensing Image Matching Based on Improved SuperPoint

Image matching is a fundamental technology for data processing and analysis in remote sensing robots. To address the limitations of deep learning models in image matching, particularly the poor matching performance and limited number of matching points in remote sensing images, this paper presents an enhanced SuperPoint model based on deep learning. It incorporates a sub-model called DC-SuperPoint for feature point extraction and description. Within the feature extraction network, the CBAM (Convolutional Block Attention Module) attention mechanism is introduced to enhance the salience of point features in the graph. Additionally, the network extraction branch for feature points and descriptors is replaced with deformable convolution DCNV2 (Deformable Convolutional Networks v2), which is specifically designed to learn different types of features in various scenarios. Comparative experiments are conducted on two different datasets to evaluate the algorithm. The experimental results demonstrate that the proposed improved algorithm achieves a matching accuracy of 49.2% for the images, surpassing the original algorithm by 30% in terms of the number of matching points. Moreover, it exhibits better adaptability to different orientations compared to the original model. The accuracy is significantly improved to a certain extent, and the overall performance surpasses that of the original algorithm. These findings indicate that the proposed method effectively enhances matching efficiency and accuracy in scenarios with varying levels of complexity.