PatMV: Patent-level Multi-View Retrieval Network

Official implementation of:

"面向外观专利审查的专利级多视图检索方法"

PatMV is a patent-level multi-view retrieval framework designed for U.S. design patent novelty examination research.

Unlike traditional design patent retrieval methods that treat individual patent views independently, PatMV models a complete design patent as a multi-view image set and learns patent-level visual representations by explicitly exploiting cross-view geometric relationships.

---

Overview

A design patent usually contains multiple views (e.g., front, rear, side, top, bottom, and perspective views), which jointly describe the three-dimensional appearance of an object.

Existing image-level retrieval paradigms may lose complementary information among different views. To address this issue, PatMV introduces a patent-level multi-view representation learning framework.

The main components include:

• Patent-level visual encoding

  • Treats all views of one patent as a unified input unit.
  • Learns a global representation for each complete patent.

• View-aware feature modeling

  • Introduces learnable view identifier embeddings together with spatial position embeddings.
  • Preserves both spatial information and view-source information.

• Global Patch-level Cross-View Attention

  • Inserts cross-view attention modules into the Vision Transformer architecture.
  • Enables fine-grained interactions among image patches from different patent views.

• Gated cross-view feature fusion

  • Uses zero-initialized learnable gates to smoothly integrate cross-view information while maintaining pretrained feature stability.

• Multi-objective contrastive learning

  • Jointly optimizes:
    • patent-text alignment;
    • patent-single-view consistency.

---

Dataset

The PatMV dataset is publicly available at:

Hugging Face Dataset

https://huggingface.co/datasets/blueingman/PatMV

The dataset is constructed from official USPTO examination records and is designed for research on U.S. design patent novelty analysis.

It contains:

• examiner-confirmed similar design patent pairs;
• multi-view patent images;
• image-level captions;
• patent-level captions;
• distractor patents for retrieval evaluation.

Please refer to the dataset card for details.

---

Repository Structure

PatMV/
│
├── model/
│   └── Model architectures
│
├── dataset/
│   └── Dataset loading and preprocessing
│
├── loss/
│   └── Contrastive learning objectives
│
├── train/
│   └── Training scripts
│
├── validate/
│   └── Validation and evaluation scripts
│
├── inference/
│   └── Retrieval inference
│
└── data_processing/
└── Dataset conversion and preparation scripts

---

Dataset Preparation

The released dataset provides processed metadata and images.

After downloading the dataset from Hugging Face, organize files as:

data/
├── images/
├── query_metadata.json
└── gallery_metadata.json

The scripts in data_processing/ provide utilities for converting and preparing patent retrieval data.

---

Training

To train PatMV:

python train.py

Training configuration can be modified according to your environment.

The training procedure optimizes a joint contrastive objective consisting of:

1. Patent-level visual-text alignment;
2. Patent-level and single-view visual consistency.

---

Evaluation

To evaluate retrieval performance:

python inference.py

The evaluation pipeline supports patent-level retrieval under gallery distractors.

Common retrieval metrics include:

• Hit@K
• mAP

---

Pretrained Model

The pretrained checkpoint will be released separately.

---

Citation

If you use this project or the PatMV dataset in your research, please cite:

Wang Yi-wei et al.
"面向外观专利审查的专利级多视图检索方法."
Accepted by CCIR 2026.

---

License

This repository is released for academic research purposes only.

Users should ensure compliance with applicable intellectual property regulations and dataset usage policies.

---

Acknowledgement

This work builds upon previous advances in:

• Vision Transformer;
• Contrastive Language-Image Pretraining;
• Patent-oriented multimodal representation learning.

---