北京信息科技大学学报（自然科学版）2025，Vol.40Issue(6)：69-78,10.DOI:10.16508/j.cnki.11-5866/n.2025.06.008
基于对比学习的机械专利附图相似性计算
Contrastive learning-based similarity calculation for mechanical patent drawings
摘要
Abstract
Semantic similarity comparison of patent drawings is a key step in the automated assessment of mechanical patents.Existing deep learning methods struggle to adapt to the unique data characteristics of mechanical patent drawings.To address these issues,a contrastive learning-based approach called ViT-Mech for similarity calculation of mechanical patent drawings was proposed.PatchMix was incorporated to construct patch sequences,and training samples with soft labels were generated by blending multiple images,to simulate the similarity relationship between complex structures and enhance the model's understanding of semantic information.The Spatial DINO module was introduced for feature extraction by integrating a spatial transformer network(STN)with DINOv2(distillation with no labels version 2).DINOv2's self-distilled vision Transformer(ViT)weights were leveraged to handle the visual features of mechanical drawings that differ from reality and STN was used to improve robustness against varying illustration styles.A small-scale validation set with four levels of similarity annotations was constructed for the evaluation of model performance.Experimental results show that ViT-Mech achieves 70.0％accuracy in similarity assessment task,outperforming DINOv2 by 2.0 percentage points.关键词
计算机视觉/对比学习/机械专利附图/特征提取/相似性计算Key words
computer vision/contrastive learning/mechanical patent drawing/feature extraction/similarity calculation分类
信息技术与安全科学引用本文复制引用
孙潇岳,吕学强,韩晶,游新冬..基于对比学习的机械专利附图相似性计算[J].北京信息科技大学学报（自然科学版）,2025,40(6):69-78,10.基金项目
国家自然科学基金项目(62171043) （62171043）
北京市自然科学基金项目(4254096) （4254096）
北京市教委科研计划科技一般项目(KM202311232003) （KM202311232003）
