Super-Resolution Based on Residual Learning and Optimized Phase Stretch Transform

Abstract

High resolution infrared (IR) images are often required in military and industrial applications. Due to the limited properties of IR imaging sensors and camera lens, IR images exhibit poor spatial resolution with a blur phenomenon in the edge regions. In this correspondence, we develop a new super-resolution (SR)-IR image reconstruction method using the residual learning network in the wavelet domain (WRESNET) and optimized phase stretch transform (PST). Our algorithm first transforms the input low resolution (LR)-IR image into its low-frequency and high-frequency subbands using the discrete wavelet decomposition. Subsequently, we introduce the optimized PST to operate on the LR-IR image and extract the intrinsic edge structure. The PST behaves differently at low-frequency and high-frequency regions, thus capturing the intensity variations for edge detection. We incorporate the PST extracted edge map in the wavelet subbands to preserve the intrinsic structure of images. The resultant subbands are further refined based on the missing residuals obtained using the WRESNET. The proposed method is validated through quantitative and qualitative evaluations against the conventional and state-of-art SR methods. Results reveal that the proposed method outperforms the existing methods.