Research On Microscopic Image Classification Of Ceramic Relic Fragments Based On Unsupervised Learning

During the excavation process of ceramic cultural relics,a large number of fragments will be produced,which is difficult to be preserved completely.Therefore,the restoration of ceramic cultural relics is an important part of their protection.The ceramic repair process generally includes the steps of fragment classification,fragment splicing,and hole repair.Accurate fragment classification can improve repair efficiency and reduce secondary damage.However,the accuracy and efficiency of the existing ceramic fragment classification methods are low due to two reasons: 1)the large number of fragments makes marking difficult;2)the macroscopic observation of ceramics has less feature information.In order to better classify ceramic fragments,we proposes two unsupervised deep learning methods to train a classification model for microscopic images of ceramic fragments.Specific research progress includes:(1)Establish a microscopic image dataset and image preprocessing of ceramic cultural relics.Firstly,the optical microscope was used to collect the data of ceramic cultural relic fragments.In view of the problems of reflection and uneven brightness in the collected microscopic images caused by the smooth surface of the porcelain,the contrast enhancement operation was used to preprocess the ceramic microscopic images.Secondly,in view of the problem of insufficient data set,data augmentation methods such as rotation,cropping,and grayscale are used to expand the data,so as to provide certain experimental data for the classification network.(2)Since different categories may have the same style(such as color,texture,etc.)in ceramic microscopic images,this paper adopts a deep network based on category style as the backbone network to classify ceramic microscopic images,and adds an attention mechanism module The ability of the network to extract microscopic features from ceramic microscopic images is improved.The network separates the style information of the image from the category information,so that the style information does not affect the classification results.The experimental results show that the network using the combination of category style and attention mechanism can effectively classify ceramic microscopic images of different categories but the same style.(3)Aiming at the lack of public datasets of ceramic fragments and the small scale of existing data sets,this paper adopts a contrastive learning algorithm framework suitable for small datasets,and improves the feature extraction network module,so that the network has a larger receptive field,Extract more accurate microscopic features.The network uses the data itself as a label to classify by training the similarity between pairs of positive samples and the difference between pairs of negative samples.The results show that the classification accuracy of this network framework can achieve 98.6% on the dataset of5000 ceramic microscopic images.