Remote Sensing Image Registration Based Upon Extensive Convolutional Architecture With Transfer Learning and Network Pruning

Accurate registration of remote sensing images through automatic pipelines remains challenging. While bottlenecks have deferred the advancement of traditional approaches, more attention has been attracted on the incorporation of deep learning knowledge into the image registration process. This paper develops an efficient remote sensing image registration framework based upon a convolutional neural network (CNN) architecture, which is called the geometric correlation regression with dense feature network (GcrDfNet). To acquire deep features of remote sensing images, the DenseNet associated with partial transfer learning and partial parameter fine-tuning is exploited. The feature maps derived from the sensed and reference images are further analyzed using a geometric matching model followed by linear regression to compute their correlation and to estimate the transformation coefficients. Subsequently, a network pruning scheme is investigated to diminish the model structure while moderately escalating the registration accuracy. A wide variety of multitemporal and multispectral remote sensing images with distinctive scenarios were employed to evaluate the proposed image registration system. The ensemble parameter compression ratio was approximately 2.12 while slightly reducing the registration error. Experimental results indicated that our GcrDfNet outperformed the traditional and deep learning-based state-of-the-art methods both qualitatively and quantitatively. It is believed that this new image registration model is promising in many remote sensing image processing and analysis applications.