Abstract
The field of image denoising has undergone significant advancements over the years. Recently, Convolutional Neural Networks (CNN) based denoising methods have shown remarkable performance in image denoising. Most of these adopt single-scale features, which may have limitations in denoising real-world images. Real-world noise is complex and non-Gaussian in nature. The multi-scale strategy of the Gaussian pyramid (GP) facilitates the attenuation of noise while preserving image details. Additionally, this multiscale architecture inherently reduces the data's dimensionality, resulting in decreased computational complexity. Over the past few decades, this method has been employed for image denoising; however, its application to real-world images remains computationally challenging. In this study, we implemented the GP method for denoising X-ray, MRI, non-medical images, and SIDD datasets. Furthermore, its denoising performance is compared with the wavelet transforms (Coiflet4, Haar, Daubechies, and Symlets). Quantitatively, GP achieves a significant improvement in PSNR, SSIM, and computational complexity compared to the wavelet method. PSNR of 36.8024 dB, SSIM of 0.9428, and computational complexity of 0.0046 s have been achieved, thereby offering an effective and practical solution for real-world image applications.