DURN: Data uncertainty-driven robust network for mural sketch detection

Abstract

Mural sketches reveal both the content and structure of the murals and are crucial for the preservation of murals. However, existing methods lack robustness, making it difficult to suppress noise while preserving sketches on damaged murals and fully capturing details on clear murals. To address this, we propose a Data Uncertainty-Driven Robust Network (DURN) for mural sketch detection. DURN uses uncertainty to quantify noise in the murals, converting prediction into a learnable normal distribution, where the mean represents the sketch and the variance denotes the uncertainty. This enables the model to learn both the sketch and the noise simultaneously, achieving noise suppression while preserving the sketches. To enhance sketches, we design an Adaptive Fusion Feature Enhancement Module (AFFE) to dynamically adjust the fusion strategy according to the contribution of features at different scales and reduce the information loss caused by feature dimensionality reduction to maximize the utility of each feature. We develop a novel Deep-Shallow Supervision (DSS) module to mitigate background noise using deep semantic information to guide shallow features without adding parameters. Additionally, we achieve model lightweighting through pruning techniques, ensuring competitive performance while reducing the number of parameters to only 4.5 % of the original. The experimental results show an improvement of 10. 4 % AP over existing methods, demonstrating the robustness of DURN for complex and damaged murals. The source code is available at https://github.com/TIVEN-Z/DURN.