| Affected by preservation conditions,foxing will form on the surface of many paper cultural relics.If effective monitoring and scientific judgment are not carried out,the safety of paper cultural relics will be further affected.For the detection of foxing disease on paper cultural relics,there are problems such as hysteresis and subjectivity.It is difficult to identify the area covered by ink,paint and seals in the painting and calligraphy collection.Therefore,the concept of preventive protection based on cultural relics needs to be developed urgently.Non-destructive testing technology for efficient and accurate identification of foxing.The visible-near-infrared hyperspectral image combines spectrum and image,contains rich spatial information and spectral information,and can achieve lossless batch collection of sample spectral information on the flat.This research proposes a rapid identification method based on hyperspectral imaging technology to detect foxing on paper cultural relics.Obtain hyperspectral images of simulating paper cultural relics at 360nm～970nm,and compare the healthy area with that.In the area of foxing infection,it is found that there is a difference in the spectral curves of the two.Based on this,the main research of this paper is as follows:(1)Foxing image enhancement in hyperspectral imagesA hyperspectral approach to image feature extraction is used to extract the foxing portion of the paper relics.In the extraction process,threshold segmentation,bandmath,threshold segmentation + bandmath and minimum noise fraction rotation are utilised respectively.Threshold segmentation missed the identification of foxing;wave band operation could not completely distinguish the fox spots from the seal part;threshold segmentation + bandmath had the problem that the seal could not be separated from the foxing,which needed further methods to solve;minimum noise fraction rotation eliminated the interference of the seal part compared with the result of wave band operation,which was more favourable for the determination of foxing.By comparing the number of pixels obtained by the three foxing extraction methods,the minimum noise fraction is the most suitable method for foxspot extraction.(2)Pre-processing and dimensionality reduction of foxing spectral informationTo address the problems of noise and data redundancy in the spectral information,the spectral data are smoothed and dimensionality reduced.SG smoothing filtering and principal component analysis are used to reduce the dimensionality of the data while avoiding data catastrophe and noise,improving the accuracy of the system and retaining the feature information in the spectral data.It achieves information extraction of the spectral signal,simplifies the data structure of the spectral information and reduces the complexity of subsequent processing.(3)A foxing recognition method based on spectral information of paper relicsThe recognition and classification of foxing reflectance spectra is accomplished using a convolutional neural network to classify the pre-processed spectral data.Because of the presence of convolutional and pooling layers,the convolutional neural network has the advantage that it can share convolutional kernels and does not require pre-processing of the data by dimensionality reduction.Compared to the dimensionality-reduced KNN classification(PCA-KNN)method and the improved BP neural network,the recognition discrimination rate is not only high(>93.7%),but also eliminates the dimensionality reduction step,making it faster and more suitable for foxing spectral recognition. |
