.png){kind=icon}
.png){kind=icon}
Return
Automatic Locarno Design Classification Using Text, Image, and IPC Information from Patent Documents
.png){kind=icon}
PRE
AI
K
DOI：10.3795/KSME-A.2026.50.4.285.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Prior-art search in patent examination is challenging due to the mismatch between IPC (text-based) and Locarno (visual/functional) taxonomies. This study establishes a corpus by matching Korean patent/ utility-model documents with Locarno codes. We propose a multimodal classifier that integrates IPC, text, and images. The text branch combines KorPatBERT embeddings with embeddings of the IPC hierarchy (section/class/subclass). The image branch fuses AlexNet features with local binary pattern and adaptive hierarchical density histogram descriptors. By comparing text+IPC, image+IPC, and text+image+IPC settings, we demonstrate that the fusion model consistently outperforms unimodal baselines. IPC injection and visual cues improve discrimination for sparse or noisy claims and for visually similar Locarno classes.
Keywords:
International Patent Classification
Multimodal Fusion
Patent Code Embedding
Locarno Classification
Journal
T
IF:
0.2
Papers:
87
Citations:
308
Organization
Cited Papers