Super Resolution Image Denoiser Network: Toward Real-world Image Noise Removal

Image denoising techniques have been greatly improved in recent years by deep learning, with approaches like transformers and GANs pushing the limits of performance. These cutting-edge models are usually resource-intensive, need a lot of processing power, big datasets, and a long training period, which slow down real-time inference performance. In order to overcome these obstacles, we present the Super-Resolution Image Denoiser Network (SRIDNet), a small and incredibly powerful model that uses a fraction of the data and processing power to provide high quality results. SRIDNet maintains competitive performance in image denoising while drastically reducing the model size. Across a range of datasets, our model demonstrated the fastest inference times while maintaining high quality output when compared to other models, achieving avg. PSNR of 39.50 dB and 34.23 dB on SIDD and Urban100 datasets respectively.