Pansharpening via Frequency-Aware Fusion Network with Explicit Similarity Constraints

The process of fusing a high spatial resolution (HR) panchromatic (PAN) image and low (LR) multispectral (MS) to obtain an HRMS is known as pansharpening. With the development convolutional neural networks, performance pansharpening methods has been improved, however, blurry effects spectral distortion still exist in their fusion results due insufficiency details learning frequency mismatch between MS PAN. Therefore, improvement at premise reducing challenge. In this paper, we propose frequency-aware network (FAFNet) together with novel high-frequency feature similarity loss address above mentioned problems. FAFNet mainly composed two kinds blocks, where aware blocks aim extract features domain help discrete wavelet transform (DWT) layers, reconstruct from assistance inverse DWT (IDWT) layers. Finally, are obtained through block. order learn correspondence, also constrain HF derived PAN branches, so that can reasonably be used supplement MS. Experimental on three datasets both reduced- full-resolution demonstrate superiority proposed method compared several state-of-the-art models. codes available https://github.com/YinghuiXing/FAFNet.