A Method for Labeling Flaking Disease on Ancient Murals Based on Gradient Trainable Gabor and Channel Attention Mechanism

Disease labeling is usually the first step to protect and repair mural relics scientifically and effectively. The most common method for disease labeling at present is manual mapping of disease regions on orthophoto image by humancomputer interaction. However, this method is not only time consuming, but also has different labeling result due to different expert experience. Recently, with the development of artificial intelligence, machine learning and other technologies, automatic labeling can be achieved by deep learning and so on. However, due to the wide variety and varying manifestations of mural diseases, targeted research is needed on each type of disease. For typical flaking in mural relics, this paper presents a gradient trainable Gabor and channel attention convolution network (GTGCAC-Net) to extract texture information of flaking in order to label it efficiently and accurately. Gradient trainable Gabor integrates the characteristics of gradient information and Gabor filter to obtain rich edge information and texture information and other shallow features. The decoder part uses the channel attention module to establish the dependency relationship between each feature graph in the network and enhance the ability of the network to obtain remote context information. Residual skip connections are used to better integrate the texture features extracted by shallow convolution with the deeper abstract features of the encoder. Experiments on the murals of Fengguo Temple in Yi County, Jinzhou, Liaoning Province, China, show that the method proposed in this paper can achieve automatic labeling of flaking of the mural relics.