Super-Resolution of Thermal Images Using GAN Network

Super-resolution (SR) reconstruction of thermal images has been one of the most active research areas specifically for industrial applications. However, most of the conventional RGB SR models available in the literature are not necessarily applicable to thermal images due to their difference in characteristics when compared to normal camera images. The recent advancement in the field of deep learning-based SR has helped achieve unbelievable results. Despite the advancement in models like deep convolution neural networks (CNN) and Generative adversarial networks, there remain multiple problems unsolved that will help improve the spatial resolution of thermal images. Not only the developed model should be computationally efficient but also easily implementable in industrial applications. Motivated to overcome the said limitations, in this work a generative adversarial network (GAN) based single images super-resolution architecture is proposed for thermal camera images. The developed model not only generates at par results with the other model but also is easy to implement and computationally efficient. The modified architecture has an identical layout inspired by SRGAN. In order to make the model faster to train while having less training parameters, the number of residual blocks was reduced to 5. The batch normalization layers were excluded from the residual blocks of both the Generator and Discriminator networks to remove the redundancy. Before each convolution layer, reflective padding is utilized at the edges to preserve the size of the feature maps. The comparative results revealed that the proposed network trained on thermal images produced high-quality images with enhanced details, while still maintaining image features and perspective throughout. The experimental results show that the proposed model has achieved a reduction in computation time compared to the State-of-the-Art method. The suggested strategy has outperformed the SOTA methods with the improvement of approximately 2dB in PSNR along with 0.9825 of SSIM.