Extracting Sketch Of Hyperspectral Painted Cultural Relics Based On Non-negative Matrix Underapproximation And Deep Residual Network

The sketches of painted cultural relics can clearly reflect the content,pattern structure,painting method and style of the author,so that the protection of painted cultural relics can not only stay in protecting the ontology,but also can carry out a more vivid inheritance of them.In recent years,hyperspectral technology has been applied to extract the sketches of painted cultural relics because of the non-destructive and"image-spectrum merging"characteristics,and has achieved some research results.However,the existing methods have more manual interventions,and the threshold is more difficult to determine;And due to the noise of painted cultural relics,the integrity and accuracy of sketches extraction needs to be further improved.In this paper,a new method for extracting sketches of painted cultural relics is proposed based on the study of non-negative matrix underapproximation and deep learning in image denoising.The main research is as follows:（1）A new method of non-negative matrix underapproximation based on L_{1 2}-norm is proposed.Non-negative matrix underapproximation can represent the spectral curve of each pixel as the additive linear combination of end-member it contains,which is helpful to separate different features.The sparsity constraint can further separate the different end-member of pigments.In this paper,the L_{1 2}-norm is used as the sparse constraint to improve the non-negative matrix underapproximation algorithm,so as to obtain a better feature separation result and obtain a component with rich sketches information,and take it as the optimal components.（2）A new method of sketch extraction based on the optimal components is proposed.The current sketch extraction algorithms only extract the sketches by spectral information.This paper makes full use of the spatial information of the optimal component image to extract the sketch.The experimental results show that the proposed method is better than the existing method,and it can extract the sketches automatically without manual intervention,and there is no limit to the pigment of the sketch.（3）A new method of denoising algorithm of painted cultural relics’sketches image is proposed.In this paper,the DnCNN（Feed-forward Denoising Convolution Neutal Networks）image denoising network is improved.By adding the residual network,it can train networks more deeper so that the noise distribution can be learned better.To solve the problem of insufficient training samples,this paper adopts transfer learning pretraining to ensure the effect of network denoising.Experimental results show that the proposed algorithm can effectively remove the noise in the sketch image,which is helpful to extract cleaner skeches.