Multispectral Remote Sensing Image Matching via Image Transfer by Regularized Conditional Generative Adversarial Networks and Local Feature

Multispectral image matching is at the base for many remote sensing and computer vision applications. Due to the different imaging principles and spectra, there are significant nonlinear variations in intensity, texture, and style in multispectral images. This makes it difficult for many classic methods designed for the images of the same spectrum to achieve satisfactory matching performance. To cope with this problem, this letter proposes a new method based on image transfer and local feature for multispectral image matching. First, we propose a new regularized conditional generative adversarial network (GAN) for image transfer to preprocess the multispectral images. This step eliminates the differences in grayscale, texture, and style between the multispectral images. Then, we use a classic local feature to match the generated and original images. We evaluate our method on two commonly used data sets and compare with several state-of-the-art methods. The experiments show that our method performs well by significantly improving the matching accuracy and robustness, and slightly increasing the runtime.