DRCDNet:Adaptive Deep Residual CNN for Real-Time Blind Denoising and Quality Enhancement

Abstract

The proliferation of digital imaging systems in medical diagnostics, autonomous driving, surveillance and consumer photography has intensified the demand for robust real-time image denoising and quality enhancement frameworks. Conventional denoising approaches including Gaussian filtering, non-local mean and BM3D are constrained by fixed noise assumptions and prohibitive computational costs in real-time scenarios. To overcome these limitations, this paper proposes a Deep Residual Convolutional Denoising Network that leverages hierarchical feature extraction residual dense blocks and channel attention mechanisms to achieve superior noise suppression while preserving fine-grained structural details. The proposed DRCDNet incorporates a multi-scale encoder-decoder backbone augmented with squeeze andexcitation attention modules and adaptive noise level estimation to handle spatially variant noise distributions. Extensive evaluations on benchmark datasets including BSD68and Urban100 demonstrate that DRCDNet achieves state oftheart Peak Signal-to-Noise Ratio and Structural Similarity Index values across Gaussian, Poisson, and real-world noise conditions, while maintaining real-time inference throughput on standard GPU hardware.