Mural Disease Detection Based on ConvUNeXt with Improved Up-Sampling and Feature Fusion
Xueying Ni, Ying Yu, Huirong Zhao, Yicen Li · 2024
Disease detection in mural paintings provides guidance for digital restoration. To address the existing problems such as insufficient detection in complex areas and erroneous detection when there is a low distinction between the target and the background, this paper proposes a ConvUNeXt network with improved up-sampling and feature fusion for high-precision detection of mural diseases. ConvUNeXt adopts a U-shaped architecture, consisting of an encoder utilizing ConvNeXt and a decoder composed of ConvNeXt Blocks. Firstly, we design a ConvNeXt-Up module to reduce the information loss during the up-sampling process and retain more semantics and details. Secondly, it is combined with the Multi-scale Attention-Guided Fusion (MAGF) module to guide shallower features by deeper features, balancing the global semantics with the local details. On the two established datasets of the Shaanxi Temple mural and the Dunhuang mural with disease masks, our method outperforms other methods and respectively improves 1.4% and 0.78% in IoU metric compared to ConvUNeXt. The model proposed is superior in the task of mural disease detection, in addition, the test results on the Baisha mural show its robust generalization capabilities.