An effective image despeckling and reconstruction approach using U-Net based model and comparative analysis

U-Net-based deep learning models have garnered significant attention in recent years due to their strong denoising capabilities in image restoration tasks. This study critically evaluates both the strengths and limitations of these models, with a particular focus on their architectural design and constituent components, in an effort to further advance denoising performance. Based on the insights derived from these analyses, a novel architecture–termed U-Tunnel-Net–is proposed. The model is trained on the UNS and Waterloo datasets, each augmented with Rayleigh-distributed speckle noise at four distinct intensity levels (= 0.10, 0.25, 0.50, and 0.75), and evaluated on the UNS, BSD68, and Set12 datasets. Performance assessment is conducted using peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), and runtime as evaluation metrics. Leveraging a novel network design strategy and a newly introduced convolutional block, U-Tunnel-Net consistently achieves superior performance compared to other U-Net-based denoising models. A key architectural innovation lies in the repositioning of the pooling operation within the Tunnel Blocks, which differentiates the proposed model from conventional U-Net variants. Experimental results on both benchmark and real-world datasets confirm that U-Tunnel-Net outperforms several state-of-the-art denoising methods.