基于改进U-Net的壁画颜料层脱落病害区域提取

Extraction of Mural Paint Loss Diseases Based on Improved U-Net

颜料层脱落区域的提取是壁画科学保护和修复的重要环节,由于标准U-Net在传播过程中低维的细节信息逐渐减弱,使得病害区域提取精度受到限制。因此,本文提出一种改进型U-Net网络,用于壁画颜料层脱落病害区域提取。首先,通过影像裁剪和预处理得到青海省瞿昙寺壁画图像数据集。其次,在编码层中引入空间金字塔池化层,构建一种低维特征信息保留结构。同时,在解码层中通过池化索引上采样,减少图像边缘信息在反卷积过程中的损失。最后,在瞿昙寺壁画图像数据集上进行验证。结果表明,改进U-Net方法在IoU和F1-Score两项评价指标上的均值分别达到82.7%和90.5%,优于SegNet及标准U-Net,证明了改进U-Net网络模型对壁画颜料层脱落病害区域提取的有效性,可为壁画保护现状调查提供依据。

Extracting paint loss on mural surface plays an important role in scientific preservation of murals.Because the low-dimensional detail information of the standard network will be weakened during the propagation process,it would limit the accuracy of disease extraction.Therefore,we proposed an improved U-Net model to extract paint loss area in murals.First,we preprocessed and clipped the images to produce the image dataset of murals in Qutan Temple,Qinghai Province,China.Second,the spatial pyramid pooling layer is introduced into the coding layer to construct a low-dimensional feature information retention structure.Third,we used a pooling index in up-sampling in the decoder network to reduce the loss of image edge information during the deconvolution.Finally,experiments were carried out on the imagery dataset covering three different areas of murals in Qutan Temple.The results show:The improved U-Net performed better than SegNet and standard U-Net in terms of average IoU and F1-Score,which are 82.7% and 90.5% by the improved U-Net,respectively.In general,the improved U-Net can extract the paint loss effectively and provide the essential information for the investigation and preservation of murals.