Attention-Guided Deep Reinforcement Learning for Realistic Neural Painting
Neural painting aims to produce realistic artworks using stroke sequences, having attracted considerable interest from academia and industry. Previous methods often focus on minimizing the total color distance, neglecting distinctions between foreground and background objects. In contrast, human art...
Saved in:
| Main Authors: | , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
IEEE
2025-01-01
|
| Series: | IEEE Access |
| Subjects: | |
| Online Access: | https://ieeexplore.ieee.org/document/11023249/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Neural painting aims to produce realistic artworks using stroke sequences, having attracted considerable interest from academia and industry. Previous methods often focus on minimizing the total color distance, neglecting distinctions between foreground and background objects. In contrast, human artists typically prioritize important content over background details. Motivated by this observation, we propose an attention-guided deep reinforcement learning framework for stroke planning that emulates human painting processes. Our approach consists of an attention-guided policy network that generates stroke parameters and a stroke rendering network that updates the canvas. The networks interact iteratively: the policy proposes stroke parameters based on attention maps, while the renderer applies them and feeds updated canvas states back to the policy. Specifically, the attention module computes an attention map to guide stroke generation, and a feature-masked reward function prioritizes important regions. Experiments demonstrate that our method achieves high-fidelity rendering of foreground objects with fine details, while approximating backgrounds with coarser strokes. This approach produces visually appealing paintings using significantly fewer strokes compared to existing methods. |
|---|---|
| ISSN: | 2169-3536 |