Remote Sensing Image Registration Based on Improved Geometric-Matching CNN

Environmental change detection is one of the uses of satellite images. This process is performed by subtracting image pairs obtained by different time series or sensors. Therefore, image registration is an important pre-processing step in detection of environmental changes. Currently, image registration methods based on deep learning are gaining attention. In general, higher satellite image resolution results in more accurate registration. However, the increase in image size leads to higher computational costs during training and estimation of deep learning models. Then, we propose a method that reduces the number of parameters of the model to lower the computational cost while maintaining the accuracy. This method makes it easier to handle high-resolution satellite images. The proposed method modified the GMCNN (Geometric-matching CNN) architecture by adding CSA (Cosine Similarity Attention) and SE (Squeeze-and-Excitation) layers to enhance the feature map, and point-wise convolution to reduce the number of parameters. The improved GMCNN decreases the grid MSE by 0.0037 compared to the conventional GMCNN. It also reduces the number of parameters by 49.6%.