| 摘要: |
| 图像去噪是计算机视觉中的基础任务, 其关键在于利用有效先验知识恢复噪声污染下的细节信息. 针对传统卷积神经网络因固定权重与局部感受野限制而存在的性能瓶颈, 以及Transformer在全局建模时面临的高计算复杂度问题, 提出一种基于特征空间上下文的Transformer去噪方法FSCformer. 该方法设计高效感受野模块, 通过动态捕获多尺度上下文信息, 在增强空间感知能力的同时显著降低计算开销; 采用卷积注意力模块, 将局部特征提取与全局依赖建模有机结合, 提高模型在复杂噪声环境下的鲁棒性; 提出跨特征融合机制, 通过多尺度特征的精细化交互增强图像细节保留能力. 大量实验结果表明, 该方法在去噪精度与计算效率之间实现了良好平衡, 并在多个基准数据集上优于现有的多种图像去噪模型. |
| 关键词: 图像去噪 Transformer 大感受野 特征融合 |
| DOI：10.13328/j.cnki.jos.007569 |
| 分类号:TP18 |
| 基金项目:国家自然科学基金重点项目(61836016); 广西多源信息挖掘与安全重点实验室开放基金 (MIMS24-02) |
|
| Feature Spatial Contextual Transformer for Image Denoising |
|
HU Yu-Xuan1, ZHANG Shi-Chao2,3
|
|
1.School of Computer Science and Engineering, Central South University, Changsha 410083, China;2.Key Laboratory of Education Blockchain and Intelligent Technology (Guangxi Normal University), Ministry of Education, Guilin 541004, China;3.Guangxi Key Lab of Multi-source Information Mining and Security (Guangxi Normal University), Guilin 541004, China
|
| Abstract: |
| Image denoising, a fundamental task in computer vision, relies on leveraging effective prior knowledge to restore detailed information corrupted by noise. This study proposes a feature spatial contextual Transformer for image denoising FSCformer to address the performance bottlenecks of traditional convolutional neural network (CNN) caused by fixed weights and the limitation of local receptive fields, and the high computational complexity of Transformer for global modeling. In this method, an efficient receptive field module is introduced to dynamically capture multi-scale contextual information, enhancing spatial perception while significantly reducing computational overhead. Furthermore, a convolutional attention module is adopted to integrate local feature extraction with global dependency modeling, improving the robustness of the model especially in complex scenes. Additionally, a cross-feature fusion mechanism is proposed to preserve image details by promoting the fine-grained interaction of multi-scale features. Extensive experiments demonstrate that the proposed method achieves a favorable trade-off between denoising accuracy and computational efficiency, outperforming several existing image denoising models across multiple benchmark datasets. |
| Key words: image denoising Transformer large receptive field feature fusion |
