Extraction And Restoration For Mural Line Drawing

Line drawing accurately reflects the overall structure and original appearance of murals with generalized lines.It is the starting point for the recording and protection of murals,and it is also an important object and means for historical archaeology,art research,and cultural inheritance.At present,computer-aided software and hardware systems have played a role in the generation of mural line drawing,but the core line drawing process still requires manual interaction,which greatly restricts the progress of the work.Although some scholars have carried out exploratory research on the generation of line drawings of mural images using computer vision and image processing technology,there are still technical difficulties:(1)Most of the existing edge detection technologies perceive the edge of each color area from the perspective of gradient change,unable to recognize and retain artistic styles such as the thickness change of the mural line drawing;(2)The noise in the image is indistinguishable from the line drawing details,causing noise pollution and loss of details during the line drawing extraction process;(3)The general image restoration algorithm focuses on restoring the visual continuity of the image while ignoring the structural and semantic rationality and correctness requirements of line drawing restoration and reconstruction tasks;(4)Due to the lack of massive and high-quality line drawing annotated images as support,general data-driven methods often cannot be carried out smoothly.In view of the above challenges,this thesis deeply combines the task characteristics of mural line drawing generation,take on the extraction and repair of line drawing generation as the core task,and aim to achieve an accurate and efficient line drawing generation algorithm and workflow.The main research contents and contributions of this thesis are as follows:(1)A line drawing generation algorithm based on flexible layering and style preservation is proposed.Based on the unique style requirements of line drawing in the field of cultural relics,in view of the problem of line discontinuity and line information redundancy in the line drawing feature extraction algorithm,it is proposed to convert the line drawing feature extraction into the separation and fusion of the image background layer and the line drawing layer.For background removal,the high-frequency enhancement filtering algorithm based on Gaussian blur is used to remove background information irrelevant to line drawing as much as possible.For line drawing preservation,an edge localization algorithm based on threshold adaptation is used to preserve as many useful lines as possible.By refining the outer outline and inner stroke in the line drawing,the style preservation of the mural line drawing is realized.Based on this algorithm,an interactive computer-aided line drawing system is proposed to achieve fast and accurate line drawing generation.Through a large number of user studies,the effectiveness of the algorithm and interactive system in this thesis is verified,and a batch of high-quality mural line drawing annotation datasets has been accumulated,which provides valuable support for subsequent data-driven method research.(2)A method for extracting mural line drawings based on multivariate supervision and transfer learning is proposed.First,a multi-level and multi-scale network architecture is proposed to solve the problem of missing features and redundant invalid lines caused by the difference of the scale of the line drawing.Second,by designing a multi-dimensional deep supervised feature network,each feature layer is supervised by labels of specific scales,which promotes the effective fusion of multi-scale features,and introduces an attention mechanism and dilated convolution to improve feature perception capabilities.Finally,to solve the problem of an insufficient number of paired training samples for murals,the idea of transfer learning is adopted,the natural image data set is fully used to assist the training process,and the weak label strategy is introduced,successfully solving the multi-label classification problem supported by small samples.Quantitative and qualitative experimental results show that the method in this thesis achieves outstanding extraction results on real mural images and reflects the good generalization ability on data from different sources.In terms of performance indicators,it also achieves leading advantages.(3)A semantic-aware-based restoration algorithm for defective line drawing is proposed.Analyzing the phenomenon of line defects,in essence,is to solve the problem of ambiguity in the completion of line structural defects.In this thesis,the most semantic images,such as the figures in murals,are taken as objects,and the semantic supervision of the line restoration process is completed by constructing a multi-semantic discriminator and a global discriminator to work together to ensure the semantic rationality of the synthesized lines.To solve the problem of lack of texture context information in the completion of structures that miss lines,a progressive image synthesis network structure is proposed.Through the loop feedback mechanism,the restoration process of line structure is gradually completed.It is worth mentioning that,in view of the common problem of missing annotation data for mural figures,this thesis uses a large amount of annotated information in the public real face image data set to improve the local localization module through the domain adaptation method and combines with a small amount of manual work.Annotated human mural data is coordinated to achieve accurate localization of facial local features in cross-domain images.Extensive experiments demonstrate the effectiveness of our method on various types of line defect repair tasks.To sum up,this thesis focuses on the characteristics of murals and proposes three effective and accurate line drawing extraction and restoration algorithms for the line drawing generation task of mural images,which solves the problem of style preservation and data dependence in the line drawing generation task.The method and system based on these principles algorithms have been demonstrated and applied in the actual cultural relics business.The research work in this paper has theoretical significance for related tasks in the field of computer vision and has practical value for the line drawing generation business in the field of cultural relics.