PIRN: Phase Invariant Reconstruction Network for Infrared Image Super-Resolution
Jun Dan,Tao Jin,Hao Chi,Mushui Liu,Jiawang Yu,Keying Cao,Xinjing Yang,Luo Zhao,Haoran Xie
DOI: https://doi.org/10.1016/j.neucom.2024.128221
IF: 6
2024-01-01
Neurocomputing
Abstract:Single image super-resolution (SR) reconstruction plays a crucial role in various fields, including surveillance and remote sensing. However, the majority of available SR reconstruction methods are designed primarily for visible images, making it challenging to obtain satisfactory results when applied to infrared images. Given the low contrast, low resolution, and limited dataset size of infrared images, we propose a new method for infrared image SR reconstruction called Phase Invariant Reconstruction Network (PIRN), which employs a phase invariant data augmentation (PIDA) strategy, deformable convolutional layers, and a phase invariant content loss (PICL) function. PIDA, in particular, effectively expands the diversity of infrared images by randomly perturbing their amplitude spectrum, improving generalization and robustness of the model. Learnable offsets are used by deformable convolutional layers to assist the model in extracting fine-grained spatial features from infrared images. Moreover, PICL is proposed to capture texture information differences between high-resolution (HR) and SR images in order to enhance the texture details of the reconstructed images. A series of experiments and analyses on two public benchmarks demonstrate the effectiveness of our method. On the fusionA-22 dataset, the proposed PIRN outperforms the SRGAN baseline model by +1.142 dB in the Peak Signal-to-Noise Ratio (PSNR) metric and +0.041 improvement in the Structural Similarity Index (SSIM) metric. Additionally, our PIRN performs the best reconstruction on the fusionC-22 dataset, with a PSNR of 34.412 dB and an SSIM of 0.891.