Low-level Image Generation: Deep Learning for Makeup Transfer and Unified Visual Restoration
- Publication Type:
- Thesis
- Issue Date:
- 2024
Open Access
Copyright Clearance Process
- Recently Added
- In Progress
- Open Access
This item is open access.
With the continuous advancement of artificial intelligence, deep learning-based computer vision technology is making significant progress. As a result, more applications are being integrated into everyday scenarios. In this thesis, we focus on two specific tasks within the field of image generation: makeup transfer and image reconstruction. Both of them have strong practical value, but there are still limitations in the implementation. In makeup transfer, the difficulty of precisely capturing facial contours often leads to generated faces appearing overly smooth and lacking in realism. To address this, we incorporate 3D facial information to accurately preserve geometric features, thereby significantly enhancing the fidelity of the makeup transfer process. In the reconstruct task, models with high accuracy often struggle to maintain real-time inference speed, which limits its application scenarios. To tackle this issue, we select video deraining as a representative task and design a Transformer-based approach. Furthermore, we incorporate a memory bank as auxiliary information, enabling precise video deraining while maintaining high-speed inference and efficient reconstruction without increasing computational overhead. Moreover, most existing reconstruction strategies are designed to address only single degradation conditions, which often results in suboptimal performance when dealing with complex degradation scenarios in the real world. To solve this issue, we design a variety of solutions. First, we introduce diffusion models, which enhance generalization across diverse degradation scenarios through the progressive generation process. Second, we develop meta batch normalization inspired by meta-learning, using precision training for domain-specific parameters to enhance generalization. Additionally, we implement test-time adaptation to improve robustness under unknown weather conditions.
Please use this identifier to cite or link to this item: