Hyperspectral Image Residual Denoising Network Based on Mixed-Domain Attention Mechanism

Abstract

Hyperspectral images (HSIs) contain not only spatial information, but also detail information on spectrum that reflects the internal features of objects, which can be used to monitor crop growth, for example. It is noteworthy that noises are inevitably introduced in the obtained HSIs due to the imperfection of imaging equipment and data transmission process, which will probably lead to misjudging the species of objects. Currently, HSIs-denoising methods based on deep learning have received considerable amount of attention and achieved promising results. However, these methods did not consider the interdependence among the three domains of HSIs. Based on this, we present a mixed-domain attention-based residual denoising network (for short named MA-RDN), so as to better the noises suppression by taking all the three domains into consideration. Different from existing methods, we introduce a mixed-domain attention module, which consists of three branches, respectively modeling the correlation between two of the domains. In this way, the model is guided to simultaneously focus on all the cross-domain features that are influential in denoising tasks. We take the average value of the three branches as the module's output. Then, a sparse feature extraction subnetwork is designed to preserve spatial-spectral features of HSIs as many as possible, which contains several multiscale structures and channel attentions. In order to avoid the gradient disappearance and model degradation caused by the deepening of the network, we utilize two weighted skip connections in the output. Simulation experiments show that, in different noise conditions, the peak signal-to-noise ratio PSNR of our method is increased of about 1.6 that compared with the Cao et al's GRN [11] method, and the structural similarity SSIM is slightly better than it.