Structure-Aware Complex-Scene Image Denoising via Deep Convolutional Autoencoders

Abstract

To address challenges in complex scene images—such as overlapping noise types, non-stationary distributions, and susceptibility of structural details to degradation—this paper develops a structure-aware image denoising model based on Deep Convolutional Autoencoders (DCAE). It characterizes the trade-off between noise attenuation and detail fidelity at the feature-representation level. A symmetric encoding–decoding architecture is adopted to strengthen the representation of structural principal components through layer-wise downsampling and channel expansion in the encoder, while hierarchical reconstruction in the decoder enables high-resolution detail recovery. Meanwhile, a composite loss that combines pixel-domain reconstruction with gradient-consistency regularization is formulated to stabilize the preservation of edges and textures. Quantitative comparisons with traditional denoising algorithms and representative deep models on datasets under diverse complex degradation settings indicate consistent advantages in Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM), together with effective suppression of oversmoothing under strong and compound noise. Further analysis shows that structure-dominant modeling in the latent feature space improves discrimination of complex noise distributions and reconstruction consistency. Although large-scale global correlation modeling remains a limitation, the proposed approach offers an engineering-feasible pathway for structure-preserving restoration in complex scenes and can support noise-robust visual analysis in cross-disciplinary imaging tasks, including remote sensing interpretation, medical image enhancement, and experimental/industrial vision inspection.