Attention-Driven Dual Feature Guidance for Hyperspectral Super-Resolution

Benefiting from the high spectral resolution, hyperspectral image (HSI) owns the property of discriminating material. However, the spatial resolution of HSIs is limited by the hardware and, thus, makes HSI super-resolution (SR) a necessary and hot topic. Existed HSI SR methods rarely consider the high-frequency edge information, resulting in unsatisfactory quality of spatial reconstruction. To address the problem mentioned above, we propose a novel method for HSI SR named attention-driven dual feature guidance net (AD-DFGNet), which makes full use of the spatial-spectral information. Specifically, AD-DFGNet mainly consists of three modules, which are shallow feature extraction, DFG blocks, and upsampling fusion module. First, the three adjacent bands of an HSI cube are selected sequentially as one of the inputs to the proposed AD-DFGNet. Their feature dimension is upgraded through shallow feature extraction. Second, middle band of three adjacent bands is used as the other input to DFG block, and it makes the network focus on spatial feature extraction by DFG, which includes feature aggregation guidance (FAG) and gradient-texture attention (GTA) guidance. Then, the output features of DFG blocks are upsampled and fused, in turn, to obtain the single-band SR result. Finally, the SR results of the whole HSI are obtained from the sequential concatenation of each single band. In addition, the proposed method reduces the size of the generated model and makes it possible to apply on different datasets. The efficiency of the AD-DFGNet is validated on three publicly available hyperspectral datasets and yields the state-of-the-art performance.