Study On Extraction Of Mural Paint Loss Diseases Based On Improved U-Net

As one of the precious cultural heritage,mural is the carrier of the national spirit,historical life and the continuation of traditional culture in China.It can reflect the information of human history,culture,life and art at that time,and has important value.However,a large number of murals are affected by many factors such as natural environment and man-made destruction due to their long history,and they are facing the threat of different degrees of diseases.One of the most common diseases of mural paint loss diseases is the disconnection of the paint layer from the ground layer or the wall layer,which not only weakens the information expression of mural painting,but also has a very negative impact on the preservation and inheritance of historical culture.As the premise of mural protection,disease extraction can provide scientific basis for restoration and monitoring,which has important research significance.At present,most of the disease extraction methods need manual intervention,and the extraction efficiency and effect need to be improved.In view of the shortcomings of current extraction methods and the wide application of deep learning in image segmentation,classification and other fields,this paper puts forward a research idea of using deep learning for image segmentation technology to achieve fast and accurate extraction of mural paint loss diseases.This paper mainly completed the following works:1)Design and production of mural disease data set.Through literature research to understand the research status at home and abroad,a new idea based on deep learning mural disease extraction is proposed.In view of the problem that there is no data set about mural diseases,this paper uses close range photogrammetry technology to obtain the digital orthophoto map(DOM)of mural,and completes the production of experimental data set through related processing to provide data preparation for subsequent experiments.2)A method of mural disease extraction based on encoder-decoder network model is proposed.Firstly,the structure characteristics of U-Net and SegNet network models based on encoder-decoder structure are studied and analyzed,and the two network models are constructed by combining the mural disease extraction task in this paper.Then,the mural disease extraction experiments of model training,model prediction and post-processing are completed by using self-made data set.Finally,from the qualitative and quantitative point of view,the results of the two network models are compared and analyzed.3)An improved U-Net network model is proposed.In order to further improve the accuracy of disease extraction,this paper proposes an improved U-Net based on low-level feature retention and pooled index sampling.First of all,a low-level feature retention structure is constructed by pyramid pooling,which realizes the low-level feature of the coding layer is not lost.In addition,in order to reduce the loss of image edge information in the process of deconvolution,the improved U-Net adopts the pooled index upsampling method.Then the network model is constructed,and the model training and disease extraction experiment are completed.Finally,the disease extraction results of the improved U-Net network model are compared and analyzed.The experimental results show that the average values of IOU and F1-Score in the two test areas are 0.76 and 0.86 for SegNet,0.81 and 0.89 for U-Net,which verifies the feasibility of this method.The U-Net extraction results with low-level feature fusion are better than SegNet in edge detail description and overall accuracy,which shows that the network model is more suitable for the extraction of mural paint loss diseases.The improved U-Net in this paper improves the extraction accuracy in both test areas.Among them,the IOU and F1-Score of the results extracted from test area 1 are 0.85 and 0.92,respectively,which is about 2% higher than the original network,which proves the effectiveness of the improved U-Net network model in this paper.