Optimizing image quality through UNET-based architectures in the denoising of facial and CT-scan images

Image denoising holds significant importance across various fields. However, this study specifically focuses on facial recognition systems and medical imaging. The impact of noise on facial recognition, stemming from factors like lighting changes, camera quality, and reflections, poses challenges for accurate identification. As the noise level increases, the difficulty for facial recognition software intensifies, leading to less accurate predictions. While CT scans play a crucial role in assessing COVID-19 severity, the necessity to minimize radiation exposure, while essential, results in a trade-off: a significant reduction in radiation intensity introduces noise into the CT images. In this research, novel UNET-based architectures for denoising single and mixed-noise CT-scan images and mixed-noise facial image are proposed. The improvements in existing U-Nets are as follows: (1) Residual blocks are employed to enhance the depth of the network, preventing degradation issues. (2) An inception mechanism is introduced in the residual block, allowing multi-scale feature extraction and increasing adaptability to distinct noise patterns. (3) An attention mechanism is incorporated within skip connections to refine image feature extraction and selectively preserve crucial details. (4) Multiple loss functions and their combinations are used to train the model to prevent over-smoothing problems.Among all of the proposed models, it is found that the Inception Residual Attention U-Net (IRAUNET) achieved the highest metrics value for facial image and COVID CT-scan denoising. The experimental results reveal that the proposed IRAUNET achieves significant performance in facial image denoising, outperforming seven state-of-the-art methods in both quantitative metrics and visual assessments. Similarly, in COVID CT-scan denoising, the IRAUNET architecture demonstrates substantial advancements, yielding superior outcomes across diverse scenarios, including single and mixed noise. Validated across multiple datasets, this enhances the model's robustness and generalizability in CT-scan image denoising.