计算机科学
人工智能
卷积(计算机科学)
深度学习
学习迁移
点(几何)
班级(哲学)
模式识别(心理学)
特征(语言学)
比例(比率)
机器学习
数据挖掘
人工神经网络
数学
地理
语言学
哲学
几何学
地图学
标识
DOI:10.1109/nmitcon58196.2023.10276018
摘要
In order to solve the problem that there is no publicly available large-scale multi-category image dataset of cultural relics collections, the research on single-label and multi-label classification of heritage images based on deep learning is proposed. In the research, two representative datasets, DPM dataset and MET dataset, are constructed for domestic and foreign collection types respectively through a network approach for single-label classification research, which are useful for the construction of large-scale deep learning datasets in related fields. The experimental results show that for the problem of small samples in DPM dataset, DPM dataset is first classified by means of deep transfer learning for mainstream deep learning models, among which ReSNet50 model Dovo achieves the accuracy of nearly 87%. To address the problem of large intra-class differences and small inter-class differences in heritage images, a multi-feature fusion classification method combining point convolution and integration learning is proposed, in which the locally connected point convolution-based method finally improves the classification accuracy by nearly 5 percentage points on the DPM dataset. It is concluded that the scoring layer fusion method based on the locally connected point convolution+SL algorithm proposed in the research achieves the best results among all fusion methods, which proves the effectiveness of the point convolution+SL method.
科研通智能强力驱动
Strongly Powered by AbleSci AI