基于对比学习与多尺度结合的陶瓷显微图像分类方法

Ceramic microscopic image classification based on the combination of contrastive learning and multi-scale methods

陶瓷文物修复是文物保护研究中一项重要内容,对碎片分类可提高修复效率。针对人工标注分类耗时长、效率低、主观因素大等问题,该文基于对比学习方法对陶瓷显微图像进行分类,然而,传统的SimCLR(a simple framework for contrastive learning of visual representations)对比学习网络不能精准提取陶瓷显微图像细节,因此,该文将SimCLR网络与多尺度方法结合,对陶瓷显微图像进行分类。首先,将采集到的陶瓷显微图像进行增强并提取特征,在特征提取模块使用多尺度卷积操作替换SimCLR中的标准卷积,使得网络具有更大的感受野,提取到更加准确的特征信息;其次,使用多层感知机(MLP)将提取到的特征进行降维处理,提高后续计算效率;最后,使用归一化温度标度的交叉熵损耗对模型进行优化。实验结果表明,改进后的网络在陶瓷显微图像分类中比原始网络准确率提高1.8%,达到98.6%,且网络参数只增加了0.11 m。该方法能以较小的代价有效对陶瓷碎片分类,辅助文物修复。

The restoration of ceramic cultural relics is an important topic in the research of cultural relics protection.The classification of fragment can improve the restoration efficiency.Aiming at the problems of long time-consuming,low efficiency and large subjective factors of manual annotation and classification,this paper classifies ceramic micro images based on comparative learning method.However,the traditional SimCLR(a simple framework for contractual learning of visual representations)comparative learning network can not accurately extract the details of ceramic micro images.Therefore,this paper combines SimCLR network with multi-scale method to classify ceramic micro images.Firstly,the collected ceramic micro image is enhanced and the features are extracted.In the special extraction module,the multi-scale convolution operation is used to replace the standard convolution in SimCLR,so that the network has a larger receptive field and extract more accurate feature information.Secondly,the multi-layer perceptron(MLP)is applied to reduce the dimension of the extracted representation vectors to improve the subsequent calculation efficiency.Finally,the normalized temperature-scaled cross entropy loss is used to optimize the model.The experimental results show that the accuracy of the improved network is 1.8%higher than that of the original network,reaching 98.6%,and the network parameters are only increased by 0.11 m.The numerical results shows that the proposed method is able to classify the ceramic fragments effectively and assist the restoration of cultural relics with small cost.