AN IMPROVED DEEP NETWORK METHOD TO INCREASE THE EFFICIENCY OF UAV REMOTE SENSING IMAGE REGISTRATION

Abstract

Accurate registration of remote sensing images acquired by UAVs is one of the essential steps in applications such as aerial mapping, photogrammetry, 3D reconstruction, environmental change monitoring, and smart agriculture. Despite the advancement of deep learning-based methods, many existing algorithms still face reduced accuracy and stability when faced with changes in viewing angle, scale differences, lighting changes, and diversity of imaging scenes. Therefore, the aim of this research is to present an improved method based on deep networks to increase the efficiency of registering UAV remote sensing images. In this research, a ResNet-50 network-based architecture was designed that improves the process of extracting and matching corresponding features by utilizing the channel and spatial attention module (CBAM), multi-scale feature extraction, and the RANSAC algorithm. To evaluate the model, the UAVPairs dataset, which consists of 21,622 high-resolution images from 30 diverse scenes with significant differences in viewing angle, scale, and imaging conditions, was used. Before training the model, the images were preprocessed, including resizing, normalization, data augmentation, and image pair preparation, and then the model was trained using the appropriate cost function and Adam optimizer. The model performance was evaluated based on the criteria of Registration Accuracy, Matching Accuracy, RMSE, SSIM, and processing time, and was compared with ResNet-50 and SIFT methods. The results showed that the proposed method performed better than the comparative methods by achieving 98.1% image registration accuracy, 97.4% matching accuracy, 1.18 pixels for RMSE, and 0.986 for SSIM index. Also, the model training process showed that the proposed architecture converged faster and was more stable in the face of changes in the scene and imaging conditions. Overall, the results of this study show that the combination of the improved ResNet-50 network, the CBAM attention module, and the RANSAC algorithm can significantly increase the accuracy, stability, and generalizability of the drone image registration process and can be used as an effective solution in remote sensing, photogrammetry, 3D reconstruction, and intelligent environmental monitoring systems.