| In recent years,with the rapid development of deep learning technology,it has been widely used in the field of computer vision and achieved fruitful results.This also brings a new opportunity for the image restoration and restoration of the original appearance of cultural relics.The damage cultural relics image mainly has the characteristics of different missing area size and irregular damage area.At this stage,many image repair algorithms of based on deep learning are mostly for the missing area rules of the image repair,while repairing the texture structure of the object is relatively simple.Most of the cultural relics images are irregularly damaged and the texture structure is relatively complex,so the existing image repair algorithm still has limitations in the cultural relics image restoration scene.In view of the above problems,according to the texture characteristics and damage characteristics of cultural relics images,two kinds of cultural relics image restoration models are designed:the object restoration model based on gated convolution and the cultural relics image restoration model based on the gated convolution and coherent semantic attention mechanism.The main work of this article is as follows:(1)At present,there is no open data set in the field of cultural relics image restoration,for this phenomenon this paper collects three kinds of cultural relics images,such as bronze,Tangka,Dunhuang murals,and then pre-processes them properly,and builds three kinds of cultural relics image restoration data sets for this paper image restoration experiment.(2)An image repair model based on gated convolution is given.Considering that there is a zero-fill problem of the jump connection part of the existing U-net repair method when repairing images with large missing area,the generating network uses an encoder-decoder.Because ordinary convolution treats all input pixels as valid pixels,resulting in visual artifacts and information ambiguity in the results of the repair,this paper optimizes the repair network by gated convolution instead of ordinary convolution and residual blocks,while using the expansion convolution to increase the characteristics of feeling wild and improve the image repair ability.(3)An image repaired model based on the combination of gated convolution and coherent semantic attention mechanism is designed.There are problems of semantic incoherence and texture fracture for the repaired image.In this paper,a coherent semantic attention mechanism is designed,the gated convolution repair model is used as a coarse repair,and the restoration model based on coherent semantic attention mechanism is used as the fine repair,and the two-stage restoration of the cultural relic image is carried out to improve the semantic consistency after the restoration of the cultural relic image.(4)According to the above two-stage image restoration algorithm research and the proposed restoration model,the corresponding cultural relics image restoration application system is designed and developed.Based on the PyTorch framework,the system is developed using python-pyqt5,which realizes the main functions of loading,marking,repairing,preserving and interacting with different types of cultural relics images,and provides a new choice of the information about cultural relics image restoration.The establishment and application system of cultural relic restoration model based on the combination of gated convolution and coherent semantic attention mechanism not only enriches and improves the two-stage image restoration theory technology and model based on deep learning,but also expands and promotes the practical application of image restoration,which can effectively improve the efficiency of cultural relics image restoration. |
